.r

.

 

.: 21%..
an... 2E. .

n 3.:

a 23”.. H.

 

57.. 1

3... . ..
.1: Cl. a. I
e. « é:t..a.5§ .. .i

W»! r. .I. .v: 3 .5... .2 73
5. . 2— 7.. .a q

3211.5

VI.) .14”
.w.~h.m:xuﬁ..
t: c 5.. l ..
“My... ff:
‘ .. L: . .

«ﬂurry:
rm..lm€:ﬂ.:.¢h:

a; .2.

.12.»:

a. 2.
{Inn}

. 9: ‘ O ;: u
1115‘s
:.r.3.ll‘

Lr.’

‘ iw‘hddwgzl‘ . , . .
s??§??§ 7 . g «5 . . , Emu

 

This is to certify that the
thesis entitled

Statistical Speech Segmentation: A Neuropsychological
Investigation of Auditory Object Formation

presented by

Daniel Fogerty

has been accepted towards fulfillment
of the requirements for the

degree In Audiolgqy & Speech Sciences

 

 

(Lac; was;

V M5jor Piofessor’s Signature

oq/zsrlﬂacl

Date

MSU is an Affirmative Action/Equal Opportunity Institution

 

 

UBHARY

MiChiQan State
University

 

_l

PLACE IN RETURN BOX to remove this checkout from your record.
To AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 c:/CIRC/DateDue.p65-p.15

 

STATISTICAL SPEECH SEGMENTATION: A NEUROPSYCHOLOGICAL
INVESTIGATION OF AUDITORY OBJECT FORMATION

By

Daniel Fogerty

A THESIS

Submitted to
Michigan State University
in partial fulﬁllment of the requirements
for the degree of

MASTER OF ARTS
Department of Audiology & Speech Sciences

2004

ABSTRACT

STATISTICAL SPEECH SEGMENTATION: A NEUROPSYCHOLOGICAL
INVESTIGATION OF AUDITORY OBJECT FORMATION

By

Daniel Fogerty

During language acquisition, the development of a speech segmentation strategy has
classically been investigated by how language learners discover word boundaries. In
addition to certain acoustic cues (e. g., stress, phonotactics, etc), statistical regularities
have been identiﬁed as an early perceptual cue to segmentation. Sensitivity to statistical
regularities appears to be a domain-general process that is founded in neurophysiological
mechanisms. This study examined the ability to use statistical regularities to identify
word boundaries in an artiﬁcial language. It also expanded upon previous research by
investigating whether statistical regularities are capable of forming auditory objects. In
speech, these objects are the early, prelexical word-forms used to segment speech online.
The creation of word-forms was examined through a pretest-posttest design and
compared to optimal performance on isomorphic English tasks. The manipulability of
these word-forms was also investigated through test protocols that measured a hierarchy
of processing levels and through a hi gher-order cognitive task of phonological synthesis.
Results support the ability of statistical regularities to identify word boundaries and create
perceptual word-forms for speech segmentation. The results also suggest that these
word-forms may be represented at a higher level of processing that is cognitively
accessible. Domain-general theory of statistical learning, second language acquisition,

and therapeutic intervention techniques are discussed.

Cepyright by
DANIEL FOGERTY
2004

To Him who created our ears to hear and our brains to understand. This is my attempt to
learn of your miracle, that we might help those who struggle with this wonderful gift -
that of communication.

and...

A mon amie le plus proches et au compagnon le plus cher. . .Jonni.

iv

ACKNOWLEDGEMENTS

Thank you for the comments and support of my major professor, Jeffrey Marler,
Ph.D., CCC-SLP, and committee members, Ida Stockman, Ph.D., CCC-SLP and Brad
Rakerd, Ph.D. Your comments and technical support encouraged me to precision and
reminded me to refocus on why I’m in this profession. . .for those of us who have
difﬁculty with communication.

Thank you as well to the Michigan State University Statistical Consulting Service,
and in particular Tingting Yi, for their initial power analysis and assistance in data
analysis; Jeffrey Marler, Ph.D. for assistance in participant recruitment, data analysis, and
helpﬁil comments on previous drafts; and Aude Oliva, Ph.D. who guided me through an
experimental investigation of the domain- generality of statistical learning by using an
isomorphic visual paradigm created from portions of this study.

Most importantly, thank you to all of the individuals who participated in this
study, including family and ﬁiends who tolerated me during periods of collecting pilot
data. Your contributions and support have been invaluable to this project.

My parents have been the leading source of strength and guidance in my life.
Your love, support, and faith gave me the freedom to explore, the self conﬁdence to
succeed, and the humor to survive.

J onni, you have refocused me on what is truly important in life and continue to
grant me the courage to live it. Thank you for tolerating me these past two years and for
offering your support, encouragement, faith, slash-counting ability, and brain to test. . .and

retest. Your presence, though distant, has made all things endurable and possible.

TABLE OF CONTENTS

LIST OF TABLES ............................................................................... viii
LIST OF FIGURES .................................................................................. ix
CHAPTER 1
INTRODUCTION ................................................................................. 1
Speech Segmentation Strategies ............................................................. 6
Metrical Stress .............................................................................. 7
Allophonic variation ...................................................................... 9
Phonotactics ................................................................................ 9
Statistical Regularities ................................................................... 1 l
Electrophysiology of Auditory Sequences ................................................ 18
Electrophysiology and Neuropsychology of Speech Segmentation ................... 24
Electrophysiology ........................................................................ 25
Neuropsychology ........................................................................ 27
Proposal for the Current Study .............................................................. 37
CHAPTER 2
METHODS ........................................................................................ 41
Participants .................................................................................... 41
Stimuli ......................................................................................... 42
Speech Synthesis ......................................................................... 42
Language Construction .................................................................. 43
Transitional Probabilities ............................................................... 44
Neuropsychological Test Materials .................................................... 44
Design & Procedure .......................................................................... 45
Neuropsychological Training Tasks ................................................... 47
Behavioral Replication Testing ........................................................ 48
Neuropsychological Testing ............................................................ 49
Slow Online Segmentation Test (SOST) ........................................ 49
Auditory Probabilistic Word Count (APWC) .................................... 50
Data Analysis ................................................................................. 51
CHAPTER 3
RESULTS .......................................................................................... 54
Syllable Identiﬁcation of Synthetic Speech ............................................... 54
Inter-rater Reliability ......................................................................... 54
Standardized Screening Measures .......................................................... 55
Experimental Questions ..................................................................... 56
Question 1: Segmentation of Complex Patterns via Statistical Regularities. . ....56
Question 2: Replication of Discrimination Studies .................................. 56
Question 3: Object Recognition and Identiﬁcation .................................. 57
APWC Recognition Task ........................................................... 57
SOST Identiﬁcation Task ........................................................... 60

vi

Question 4: Changes in Performance due to Enhanced Perception ............... 61

Question 5: Manipulability of Auditory Object Memory Traces .................. 61
CHAPTER 4
DISCUSSION ..................................................................................... 63
Identiﬁcation of Word Boundaries ......................................................... 64
Question 1: Segmentation of Complex Patterns via Statistical Regularities. . 65
Question 2: Replication of Discrimination Studies .................................. 66
Integrating Acoustic Features into Auditory Objects: Creating Word-Forms. . ......69
Question 3: Object Recognition and Identiﬁcation .................................. 70
The Recognition Task ............................................................... 70
The Identiﬁcation Task ............................................................. 71
Question 4: Changes in Performance due to Enhanced Perception ............... 74
Question 5: Manipulability of Auditory Object Memory Traces .................. 75
Building Hi gher-Level Processing: From Discrimination to Identiﬁcation. 76
Evidence of Phonological Synthesis .............................................. 77
Summary of the Research Questions ...................................................... 78
Domain-Generality of Statistical Learning and Perceptual Objects ................... 79
Applications of the Current Study ......................................................... 80
Second Language Instruction .......................................................... 81
Theraputic Intervention ................................................................. 85
Study Limitations and Considerations .................................................... 89
Syllable Identiﬁcation Pilot Data ...................................................... 89
Perceptual Learning of Synthetic Speech .......................................... 89
Altering the Probability of the Artiﬁcial Language ............................. 90
English Validity Testing ................................................................ 92
Additional Considerations ............................................................... 92
Further Questions and Areas of Research ................................................. 93
Conclusions .................................................................................... 96
APPENDICES ................................................................................... 100
Appendix A: Participant Forms and Recruitment ....................................... 101
Appendix B: Pilot Syllable Identiﬁcation Data ......................................... 108
Appendix C: Experimental Procedure and Response Forms ......................... 111
Appendix D: Individual Participant Data ................................................ 120
Appendix B: Data Figures .................................................................. 127
REFERENCES ................................................................................... 130

vii

LIST OF TABLES

Table 1. Standardized test scores ............................................................... 55
Table 2. APWC means and standard deviations .............................................. 59
Table 3. SOST means and standard deviations ................................................ 60
Table 4. Signiﬁcance probabilities and effect sizes for test comparisons ................. 61
Table 5. Effect sizes changing according to task difﬁculty ................................. 62

viii

LIST OF FIGURES

Figure 1. Mean scores for each participant on the ZIP C ..................................... 57

ix

CHAPTER I
Introduction

Conceptualizing the recognition of speech is a daunting task. It may be examined
either through breaking down of a continuous speech stream into meaningful segments,
or joining multiple acoustic signals into auditory objects. It is through an adequate
system of speech perception that meaning emerges and language develops. This paper
investigates a speciﬁc process to acquiring a system of speech perception that is capable
of identifying words. How is a continuous stream of speech, containing endless acoustic
features, initially organized into meaningful perceptual objects? Or conversely, how is
that same speech stream, containing no acoustic breaks to mark the end of words, broken
into meaningful segments? Through a historic line of research, processes of speech
perception have been identiﬁed, as well as rejected.

This paper posits that individual speech stimuli are ﬁrst processed as isolated
acoustic events. Later, the statistical regularities of speech stimuli allow for the
integration of them into larger perceptual objects. An early sensory memory store is
implicated in the ability to compare perceptual features for these objects. It facilitates
online parsing of the continuous speech stream into individual objects that can later be
discriminated in isolation or reformed from temporally segmented object features. These
perceptual objects may, through later processes, eventually become lexicalized as words.

Improvement in identifying perceptual units within the continuous stream would suggest

that perceptual abilities change due to exposure to the statistical probabilities inherent in
the speech stream, perhaps due to the formation of auditory objects.

Kubovy and Van Valkenburg (2001) argued for the concept of an auditory object.
According to them, a perceptual object “is that which is susceptible to ﬁgure-ground
segregation” (p. 102). Historically studied by Gestalt psychologists within the visual
modality, Kubovy and Van Valkenburg describe ﬁ gure— ground segregation in terms of
domain-general objects. Early pre-attentive processes of grouping and feature integration
occur following the principles of Gestalt psychology. Stimulus elements are formed into
perceptual organizations. Later, attention selects one organization (or a set of them) to
become the ﬁgure, while all other information is relegated to the ground. Thus, those
perceptual organizations that form the ﬁgure are perceptual objects, segmented from all
other undifferentiated information of the ground. However, while visual information is
segmented from the spatial scene, auditory information is segmented within a temporal
envelope. Therefore, this paper deﬁnes an auditory object as a set of stimulus elements
that are pre-attentively organized, grouped, and differentially selected (or segmented)
from all other acoustic information. It is the position of this paper that individual speech
stimuli may be organized, grouped, and selected through statistical means to form an
auditory object, which can then be differentiated from the rest of the speech stream.
These auditory objects serve as the initial parsing of the continuous speech stream into
identiﬁable units. However, even if these auditory objects coincide with the boundaries
of syllables or words, such information is prelexical; the meaning of such objects has not
necessarily been accessed (or even acquired) at this time in processing. Therefore, this is

a study of speech object formation and perception, not lexical access.

The concept of a domain in this paper goes beyond the different sensory processes
(or modalities) of vision and audition to include such cognitive processes as morphology
or reasoning. Domain-generality states that the same perceptual/cognitive processes that
underlie one speciﬁc function also underlie other similar functions. This concept is
typically expanded to support distributed network models of neural functioning, rather
than speciﬁc modules within the brain that account for only one ﬁrnction. Therefore,
when domain-generality is discussed in this paper, I am referring to cognitive/perceptual
processes which may apply equally well to visual or auditory domains (and possibly
others).

Infants, as well as other animals (monkeys and birds) have been shown to poses
an innate ability to discriminate differences between speech sounds (Kluender, Diehl, &
Killeen, 1987; Kuhl & Miller, 1975; Kuhl & Padden, 1983; Steinschneider, Schroeder,
Arezzo, & Vaughan, 1995). However, what appears necessary for speech perception is
the acquisition of a system for recognizing and organizing speech patterns. Simos and
Molfese (1997) discuss an innate bases and early reorganization of the
neurophysiological mechanisms involved in categorical speech perception. Infants soon
acquire the ability to ignore certain phonetic contrasts that are not meaningﬁrl to their
language experience and to acquire sharper contrasts in other areas (Cheour, Ceponiene,
Lehtokoski, Luuk, Allik, Alho, & Naatanen, 1998). This ability to organize sound
variations into phonetic categories is essential for speech perception; thereby, allowing
for allophonic and speaker variations to be perceived as belonging to the same phonetic

category.

Phonetic categories allow for the formation of meaningful perceptual units that
can indicate a change in word meaning; these units are recognized as phonemes.
However, listening to a continuous speech stream composed of endless phonemes is no
more worthwhile than when it was composed of endless acoustic features. These
phonemes must again be organized into larger perceptual units. Statistical learning is a
process which creates perceptual units based on the probabilities that one feature (a
phoneme in this case, or a syllable as presented by Saffran, Newport & Aslin, 1996)
occurs adjacent to another one. Weak probabilities will likely signal the end of words.
Stress and other prosodic features likely offer additional cues (see J usczyk, 1999 for an
overview).

The exact process of the perceptual identiﬁcation of words in a novel language is
not fully understood. The historic problem which has obfuscated this identiﬁcation
process is the absence of consistent pauses marking word boundaries (Cole & Jakimik,
1980). To address this problem, acoustic speech cues have been proposed to assist in the
segmentation process. Markers such as metrical stress (Cutler & Norris, 1988; Jusczyk,
Houston, & Newsome, 1999), phonotactics (Mattys & Jusczyk, 2001a; Vitevitch & Luce,
1999), and allophonic variations (Jusczyk, Hohne, & Bauman, 1999) have all been
proposed. Additionally, the method of statistical learning, which calculates the
conditional probability between speech syllables, termed transitional probabilities, has
also been proposed (Christophe, Dupoux, Bertoncini, & Mehler, 1994; Goodsitt, Morgan,
& Kuhl, 1993; Saffran, Aslin, et al., 1996). Transitional probabilities are calculated from
the co-occurring frequency of two stimuli. In speech segmentation research, these stimuli

are typically syllables. The use of transitional probability as a tool to speech

segmentation is predicated on the premise that intra—word syllables have a higher
probability of co-occurring than do syllables that span word boundaries. For example,
Saffran (2003) discussed the phrase pretty baby. The syllables pre and (y occur together
much more frequently than the syllables which span the word boundary, ty-ba. In fact, in
the speech directed to young infants, pre is followed by ty about 80% of the time;
whereas, ty is followed by ba only about 0.03% of the time (Saffran, 2003). Thus, weak
transitional probabilities may be used to indicate the ends of words. Metrical stress and
statistical cues to word boundaries appear to be the earliest segmentation strategies used
by infants (Jusczyk, 1999), emerging around 7.5 months (Jusczyk, et al., 1999) and 8
months (Saffran, Aslin, at al., 1996) respectively. Several computational models of
segmentation learning have attempted to examine these segmentation processes (Brent,
1999) though not one clear method has emerged. In addition, the relative strength of
contribution between different proposed methods of segmentation has received a cursory
examination with varying results (for an investigation of statistics versus stress, see
Johnson & Jusczyk, 2001; Saffran, Newport, & Aslin, 1996).

The use of statistical cues appear to be a domain-general technique for learning,
as its properties have been demonstrated for visual stimuli (Fiser & Aslin, 2002;
Kirkham, Slemmer, & Johnson, 2002), nonlinguistic acoustic stimuli (Saffran, Johnson,
Aslin, & Newport, 1999), artiﬁcial grammar acquisition (Gomez & Gerkin, 1999, 2001;
Saffran & Wilson, 2003), and is present in animal learning (Hauser, Newport, & Aslin,
2001). These studies suggest that sensitivity to statistical structure is neither found solely
in a linguistic domain, nor is it unique to humans. The use of statistical cues may be a

natural technique used to identify perceptual objects within a set context. As such,

statistical cues identify and segment auditory objects within the speech stream. A further
investigation of the perceptual and cognitive processes undergone during this

segmentation strategy is warranted, which this paper proposes to address.

Speech Segmentation Strategies

Infants as young as one month can use the acoustic properties of speech to
differentiate phonetic categories (Eimas, Siqueland, J usczyk, & Vigorito, 1971). By 6
months of age, this ability is altered through linguistic experience to reﬂect phonetic
perception patterns of the native language (Kuhl, Williams, Lacerda, Stevens, &
Lindblom, 1992). The beginning of successﬁrl speech segmentation starts around 7.5
months of age (Jusczyk & Aslin, 1995) and continues to develop as additional strategies
of segmentation are applied (J usczyk, 1999). Infants even apply these segmentation
strategies to other languages that have similar structures (Houston, J usczyk, Kuijupers,
Coolen, & Cutler, 2000). By 8 months of age, infants exhibit some ability to remember
frequently occurring words (J usczyk & Hohne, 1997).

A technique that is frequently used by researchers to examine infant perceptual
abilities is the headtum preference procedure (HPP; see J usczyk, 1998). The HPP, an
effective technique in studying infants between 4.5 and 21 months of age (Jusczyk,
1998), is particularly useful to establish which of two stimuli the infant prefers (Houston,
et al., 2000). Infants are typically familiarized with isolated words and then tested on
passages which either do or do not have the familiarized words. Longer listening times to
the passages containing the familiarized words are interpreted ~as demonstrating

successful segmentation processing by the infant (Mattys & Jusczyk, 2001b).

J usczyk and Aslin (1995) were one of the ﬁrst to adapt this technique to study
speech segmentation processes. They familiarized 7.5 month-old infants to repetitions of
two different words, and then tested whether the infants would listen longer to speech
passages containing the familiarized words. Signiﬁcant evidence for this was found.
Moreover, when infants were ﬁrst exposed to the passages and subsequently tested on
repetitions of words which occurred frequently in the passages, infants again listened
longer to the repetitions of familiarized words as compared to repetitions of words that
did not occur in the passages. In contrast to the performance of these 7.5 month-old
infants, repetition of the same procedure with 6 month-olds yielded no signiﬁcant results.
These results suggest acquisition of speech segmentation abilities between 6 and 7.5
months of age. However, this study did not provide evidence for how 7.5 month-old
infants segment the speech stream, only that they can. Further studies are continuing to

examine the exact mechanism(s) for how speech segmentation is performed.

Metrical Stress

One such mechanism that has been presented and supported as a cue to word
boundaries is metrical stress (Cutler & Norris, 1988; Jusczyk, et al., 1999; Mattys,
J usczyk, Luce, & Morgan, 1999; Morgan, 1996; Norris, McQueen, & Cutler, 1995).
Cutler and Carter (1987) reported that English is a language that is overwhelmingly
characterized by initial syllable stress, with 90% of lexical items beginning with strong
syllables. This allows for the possibility of using initial syllable stress to mark word
boundaries and divide up the speech stream into manageable units that can be analyzed.

Indeed, studies have shown that infants and adults are able to segment speech based upon

stress patterns (Cutler & Norris, 1988; Jusczyk, et al., 1999; McQueen, Norris, & Cutler,
1994), even when other potential cues to segmentation contradict stress (Johnson &

J usczyk, 2001). Based upon evidence such as this, Cutler and Norris (1988) proposed a
stress-based Metrical Segmentation Strategy (MSS). They proposed this model as an
alternative to the leﬁ-to-right sequential processing strategy, such as that suggested by
Marslen—Wilson (1987), which does not allow following context to impact previous
segmentation. Other studies have cast doubt upon sequential models by showing that
lexical units within a speech context are ﬁ'equently recognized only after their acoustic
occurrence (Grosjean, 1985). The MSS allows for following context to affect preceding
segmentation through identifying likely word boundaries based upon stress patterns.
Thus, words such as run and runner might be accurately segmented using MSS, while a
left-to—right processing strategy would incorrectly segment runner into run and ner
(Cutler & Norris, 1988). Considering this evidence, sequential processing strategies are
no longer actively discussed in the speech segmentation literature.

In a study'examining infants’ ability to segment speech based upon a metrical
stress strategy, J usczyk, Houston, and Newsome (1999) exposed 7.5 month-old infants to
strong/weak and weak/strong bisyllables in fluent speech. They discovered that infants
appeared to correctly segment strong/weak bisyllables, but incorrectly segment
weak/strong bisyllables. Furthermore, when the distribution probabilities of words were
manipulated so that these weak/strong syllables were consistently followed by a speciﬁc
weak syllable (e.g., “guitar is”), the infants would misperceive these strong/weak words
(e. g., “taris”). These results suggest the inﬂuence of transitional probabilities on speech

segmentation.

Allophonic variation

A second proposed method of segmentation is the ability to use the information
provided by different phonetic variants of the same phoneme that are restricted to certain
word positions. Thus, the phoneme /p/ in pat is an aspirated allophone, but it occurs as
an unaspirated allophone in the word-ﬁnal position in lap. Therefore, the speech stream
can be segmented by using different allophonic variations as a cue to whether the
phonetic variant is in an initial or ﬁnal position. This information would then signal a
word boundary or a word-medial position. Such a strategy relies on the infant’s ability to
recognize which variants occur frequently or infrequently in certain word positions; thus,
the infant must keep track of certain phonetic probabilities, a concept which will be
addressed later.

Previous discussions of allophonic variations have proposed that such a strategy is
not available immediately, but is only employed after the speech stream has been initially
parsed by other segmentation methods (Jusczyk, 1999; Mattys & J usczyk, 2001b).
Therefore, segmentation by allophonic variations may be a reﬂection of a reﬁning ability
to segment the speech stream, and not a primary means (Jusczyk, 1999). Indeed, while
metrical stress cues appear to be available by 7.5 months (J usczyk, etal., 1999),
allophonic variation appears to ﬁrst become available sometime between 9 to 10.5

months (Hohne & Jusczyk, 1994).

Phonotactics

A third means of segmentation examines what sounds are permitted to occur
adjacent to each other within a syllable. In English, the cluster /st/ occurs relatively
frequently within syllables, while /mg/ is not permitted within a syllable (Jusczyk, 1999);
therefore, /mg/ has a high probability of signaling a syllable boundary. Mattys and
Jusczyk (2001b) used the headtum preference procedure to assess 9 month old infants’
ability to segment the speech stream based upon phonotactic cues. Good phonotactic
cues were estimated by computing the frequency of CC clusters in child-directed speech
(where the dot signals a potential word boundary). Two consonants that had a high
probability of occurring between words were associated with good phonotactic cues.
Results showed that in sentential contexts, infants listened longer to a C VC stimulus with
good phonotactic cues (meaning that the probability that C-C clusters occurred between
words in g V QC was high). Similar results occurred when good phonotactic cues only
occurred at the onset QQVC or offset C VQQ of target words in the utterance, suggesting
that infants calculate the probabilities at both onset and offset boundaries.

Mattys and colleagues (1999) also tested phonotactic probabilities between C-C
clusters. This time, infants were tested on bisyllabic nonwords (i.e., CVC-CVC). Either
the internal C-C cluster had a high probability of occurring at a word juncture or a high
probability of occurring within a word. Infants were sensitive to how the phonotactic
cues aligned with word boundaries.

Similarly, McQueen (1998) also used aligned and misaligned phonotactic cues to
word boundaries in Dutch. His results showed that native Dutch college students are able
to detect word boundaries easier when they are aligned with the word boundary, but also

suggested that these phonotactic cues are secondary to lexical activation. He presented

10

the possibility that competing lexical candidates are eliminated by phonotactic violations.
Vitevitch and Luce (1999) also concluded that increased phonotactic probabilities
facilitate sublexical processing during word competition in adults. However, this method
of lexical activation during segmentation is not available to infants or second language
learners who have a limited lexicon (unlike the ﬂuent adult speakers in these studies).
Therefore, phonotactic cues may also be an advanced, secondary cue to parsing the
speech stream.

As with allophonic cues, for infants to successfully use phonotactic information to
segment speech, they must calculate the transitional probabilities that occur between
different phonemes. The infant can then use these probabilities to determine if the

phonemes occur within the same word, or if they are likely to mark a word boundary.

Statistical Regularities

Allophonic and phonotactic segmentation strategies require the infant to monitor
the probabilities of allophones occurring in speciﬁc word positions and phonemes
occurring in certain clusters within syllables. Therefore, for infants to use these
strategies, it is likely that an established system is required to monitor these statistical
probabilities. Segmentation by allophonic variants and phonotactics are therefore only
speciﬁc applications of a more general segmentation method based on computing the
statistical regularity of information within the speech stream. Also, the segmentation
errors of taris in guitar is (J usczyk, et al., 1999) reﬂect that metrical stress cues to

segmentation may be mediated by transitional probabilities, as tar was always followed

11

by is. Johnson and Jusczyk (2001) also discuss that many of the speech cues that have
been proposed are probabilistic in nature and can be framed as statistical cues.

The use of statistical regularities appears to be a domain-general ability for
identifying perceptual objects (Kirkham, et al., 2002). Several studies have provided
evidence that statistically deﬁned patterns can be used in learning auditory and visual
sequences. In the classic study conducted by Saffran, Aslin, et al. (1996), infants were
exposed to a continuous artiﬁcial speech stream composed of four trisyllabic words. No
auditory cues to word boundaries were provided; therefore, the only cues to segmentation
were provided by the transitional probabilities between syllables. After only two minutes
of exposure to this continuous stream, infants were able to discriminate between words in
the artiﬁcial language and nonwords (i.e., three syllable patterns that never occurred
adjacent to each other during exposure), as well as between words and partwords (i.e.,
three syllable patterns that spanned word boundaries). This study inspired a line of
research supporting sensitivity to transitional probabilities across domains and species
(Aslin, Saffran, & Newport, 1998; Fiser & Aslin, 2002; Hauser, et al., 2001; Johnson &
Jusczyk, 2001; Kirkham, et al., 2002; Saffran, et al., 1999; Saffran, Newport, et al.,
1996). The primitive and ubiquitous nature of statistical sensitivities suggest that a
neurophysiological mechanism may underly them.

An explanation of how statistical patterns across domains are implicitly learned
might be provided through literature in neural networks and neural plasticity. Research
on cross-modal cortical plasticity in ferrets has shown that re-wiring the visual pathway
into the auditory cortex, instead of the visual cortex, causes the auditory cortex to not

only take over the function of the visual cortex, but to also become arranged into

12

orientation maps similar to the normal visual cortex (Sharma, Angelucci, & Sur, 2000;
Sur & Leamey, 2001). This evidence suggests that experience (the input stimulus) to
some degree determines the function of the neural cells which receive the input. In
addition, evidence suggests that the more often a stimulus is presented in the
environment, the stronger the neural connection becomes which receives that stimulus
input (Sur & Leamey, 2001). Thus, higher frequency events are maintained by stronger
pathways. The Hebbian rule suggests that if a certain cell repeatedly or persistently
stimulates another cell, then a growth process or metabolic change will occur in both or
either cell to facilitate future cell ﬁrings (Sur & Leamey, 2001).

This evidence relates to statistical learning in the following ways. First, it is the
frequency of events which creates strong neural connections within the brain. Therefore,
more probable events are more likely to create and travel along well established
pathways, facilitating faster and more efﬁcient processing. Second, neural
representations in the brain are modiﬁed to more closely resemble the incoming stimulus.
Thus, it is likely that sensitivity to statistical properties is not detected by a speciﬁc area
of the brain, but is a generalized property of forming neural connections. In terms of
learning segmentation, this may apply in the following way. Continued exposure to
certain high probability events (certain syllables occurring together) in the environment
cause stronger neural representations to be formed (which over time may result in
speciﬁc connections). These neural representations are constantly being reﬁned by
metabolic or growth processes (i.e. via the Hebbian Rule) to facilitate more efﬁcient
ﬁring when the same stimulus is re-encountered. As neural input to some extent guides

the development of neural processing (Shanna, Angelucci, & Sur, 2000), these highly

13

probably stimulus features (the co-occurring syllables) become neurally represented.
Thus, statistical learning may be a general neural process that allows for the continued
reﬁnement of the brain to interpret environmental stimuli. Kersten and Yuille (2003)
discussed a statistical model that suggests feedforward and feedback neural connections
(found in primates, see Felleman & Van Essen, 1991). These statistically derived
connections allow for higher level processing to reﬁne lower level activity (Kersten &
Yuille, 2003), and play a crucial role in perception (Pascal-Leone & Walsh, 2001) and
image segmentation (Tu & Zhu, 2002).

The very nature of neurally represented co-occurring stimulus features supports
the possibility that these features may become integrated into one neural trace, that of an
auditory object. Murray, Kersten, Olshausen, Schrater, and Woods (2002) discovered
that grouping features into an object reduces brain activity at lower processing levels;
thereby, suggesting easier processing for perceptual objects (Kersten & Yuille, 2003).
Reduced activity to grouped objects may facilitate segmentation from ungrouped, novel
elements (Murray, et al., 2002), which provides a neural basis for the ﬁgure-ground
process that Kubvoy and Van Valkenburg (2001) suggested occurs with auditory objects.
Clearly, it is expected that auditory objects are processed at a level higher than the
processing of their individual features. Indeed, Liebenthal, Binder, Piorkowski, and
Remez (2003) discovered behavioral and physiological differences using fMRI between
processing spectral information as either individual acoustic features or as a grouped
speech object.

Support for the neurophysiology of statistical sensitivity is also derived from

animal research; thereby, demonstrating its primitive nature. The use of statistical

14

information for visual sequences is also an identiﬁed ability of pigeons. In 1991,
Terrance conducted a study which showed that pigeons are able to combine visual
information into one perceptual object when learning serial lists. The ability to combine
adjacent items into one object; thereby, reducing the number of discrete items in the list,
increased recall ability (Terrance, 1991). As items in the list were consistently presented
in the same order, the transitional probabilities between two intemal items was p=l .0.
This statistical regularity provided a mechanism for adjacent items to be perceptually
grouped together. This ability to learn sequential information is important for higher
organisms to adapt and survive within a temporally bounded environment (Conway &
Christainsen, 2001).

Animals not only have the ability to identify sequential visual sequences, but can
also use statistical probabilities to identify speech patterns. Hauser, Newport, and Aslin
(2001) studied cotton-top tamarins using the same artiﬁcial language stimuli Saffran,
Aslin, et a1. (1996) used in their infant study. Like humans, tamarins are able to
discriminate syllable sequences based upon the statistical probability for which they
occurred within the speech stream. Tamarins are also able to extract enough information
from continuous speech to distinguish between Dutch and Japanese above and beyond
speaker variability (Ramus, Hauser, Miller, Morris, & Mehler, 2000).

Statistical sensitivity is also present very early in human development. Kirkham
and collegues (2002) habituated the looking times of two-, ﬁve-, and eight-month old
infants to statistically predictable patterns of visual stimuli. After infant looking times
habituated to a continuous stream of randomly ordered visual pairs, the infants were

tested on displays of familiar sequences (sequences that maintained the statistical pairs)

15

and novel sequences (sequence orders to which the infants had never been exposed).
They found that when tested later, infants at all ages showed greater interest (longer
looking times) to novel visual sequences than familiar sequences. This suggests that
infants as young as 2-months are sensitive to statistical properties, as they recognized the
violation of a learned statistical pattern. Kirkham and colleagues (2002) suggest that this
mechanism may play a role in later cognitive development.

One area where the abilities of humans and animals may begin to diverge is in
analyzing the complexity of statistical information. Hunt and Aslin (2001) showed in a
visuomotor task that adults can utilize simultaneous sources of statistical information.
Infants (Gomez & Gerken, 1999), as well as children and adults (Saffran, 2002), are able
to acquire the legal ordering of words in sentences from probabilistic structures,
independent of the actual vocabulary used. In addition, infants are able to generalize this
statistical structure to novel auditory strings with less than 2 minutes of exposure (Gomez
& Gerken, 1999) and adults can learn the statistical regularities between non-adjacent
sound sequences (Newport & Aslin, 2004). These studies suggest that statistical
information can organize the speech stream beyond simple transitional probabilities and
object formation. Statistics not only deﬁnes the internal structure of words, but also
identiﬁes the grammatical structure of sentences when lexical information is not even
available. Therefore, the suggestion that segmentation cues are secondary to lexical
activation (see McQueen, 1999) is not necessarily true, but may be an independent and
collaborative process. Lexical processing depends upon segmentation to identify
auditory objects for lexical access, and later, to statistical properties in identifying the

overall grammatical structure of the speech stream.

16

It is likely that other processes, in addition to the sensitivity to transitional
probabilities, are involved in speech segmentation (Johnson & Jusczyk, 2001; J usczyk,
1997; Saffran, Newport, et al., 1996). Several of these processes may become more
available after early structure and auditory objects have been identiﬁed using statistical
properties. Such processes may be related to lexical activation of neighboring words, or
further reﬁnements of statistical sensitivity that go beyond simple transitional
probabilities of syllables (such as allophonic variations or phonotactics). Also, this paper
does not propose that initially parsed auditory objects are necessarily ﬁnalized as real
lexical words (though they may be treated as such, see Saffran 2001), but are only an
initial parsing of the speech stream. The purpose of this paper is not to discount other
viable methods of language acquisition, but to describe the signiﬁcance that statistical
structure plays in early speech perception and how it facilitates language acquisition.
Later processes of assigning semantic meaning to auditory objects are required to
complete the process of word learning (see Hollich, Hirsh-Pasek, & Golinkoff, 2000, for
a further overview of later word-leaming processes), and the infant language learner is
also engaged in acquiring the structure of objects and events, as well as conceptual
structure (J usczyk, 1997). It may be possible that cross-modal statistical regularities also
facilitate the acquisition of word-object associations (Roy & Pentland, 2002). Statistical
learning is possibly a very early and primitive pattern detection process which allows for
the building of memory traces of frequent events (auditory objects) and later enables the
recognition and identiﬁcation of the same patterns. New research continues to unfold an
increasing complexity of statistical learning as an elaborate mechanism for perceiving

structure in the environment (Newport & Aslin, 2004).

17

Adults naturally perceive patterns when presented with continuous stimuli, even
when these stimuli are presented at random with no intended pattern (Huettel, Mack, &
McCarthy, 2002), and infants are able to implicitly detect these patterns (Saffran,
Johnson, Aslin, Newport, 1999). Implicit knowledge results from instilling an abstract
representation of the stimulus structure (Reber, 1989), a structure which forms the
perceptual object. They are implicitly learned through an automatic and natural process
of statistical learning, which may arise from brain neurophysiology.

The review of these studies indicates that the use of statistical information is an
ability that spans across species and cognitive/sensory domains, allowing for the
discrimination and segmentation of stimulus patterns, even those as abstract as grammar.
Infants as young as two-months of age are able to make use of statistical patterns to order
their world (Kirkham, et al., 2002). Studies of neural plasticity suggest stronger neural
connections are formed to process more frequent events (Sur & Leamey, 2001).
Neurophysiological changes may facilitate object segmentation (Murray, et al., 2002) and
reduce the processing required (Kersten & Yuille, 2003). As statistical properties are a
generalized mode of learning (and perhaps a general property of neural organization), it is
no surprise that it is also used to identify syllabic patterns in segmenting speech
appropriately into words. Even other suggested methods of speech segmentation, namely
allophonic and phonotactic cues, are founded on the properties of statistical patterns.
Moreover, an early method of segmentation based upon stress is inﬂuenced by

probabilistic patterns.

Electrophysiology of Auditory Sequences

18

This paper suggests that the formation of auditory objects within a continuous
speech stream is necessary to identify individual words. The electrophysiological
literature offers great insight into the processes of how these auditory objects may be
formed, and is therefore worthwhile for our review.

In an effort to identify the perceptual processes necessary in the formation of
auditory objects, the structure of auditory patterns has also been investigated using event
related potentials (ERP). These auditory objects are analogous to those that the pigeons
formed in Terrance’s tasks (1991) or the auditory sequences that are identiﬁed as words
due to high transitional probabilities. The following is a discussion supported in the adult
electrophysiological literature.

The ability to segment speech based upon transitional probabilities is the
discrimination between two units that occur together frequently (in this case, syllables
that occur within words) from two units that occur together relatively infrequently
(syllables spanning word boundaries). A parallel for this process is present in research
examining an ERP component called the mismatch negativity (MMN), an index of
auditory discrimination independent of attention (Naatanen, 1992, 1995). The oddball
paradigm is typically used to elicit the MMN by presenting a highly probable stimulus
(called the standard) and a highly improbable stimulus (the deviant). The response
elicited by the standards is subtracted from the response to the deviant, creating a
difference wave that peaks at 100 to 200 ms after stimulus onset.

In this paradigm, the duration of time between the offset of one stimulus and the
onset of the next stimuli is referred to as the Stimulus Onset Asynchrony (SOA), or

sometimes as an Inter-Stimulus Interval (181). The logic behind using SOAs is that by

19

delaying presentation, decay occurs in the neural representation of the standard
(Naatanen, 1992). Also, when SOA durations are reduced, stimuli elicit larger MMN
amplitudes at the faster presentation rate (Naatanen, 1992). It is also possible that longer
SOA times assist in providing a temporal gap for dissociating stimuli. During continuous
speech, there are relatively few SOAs (acoustic breaks/silences) between syllables.

Saffran, et al. (1999) reported that adults can learn auditory sequences, even when
uninformed that the stimulus stream contained units and were instructed to avoid
analyzing it. Therefore, infrequent probabilities within the speech stream may provide
the cues to segmentation independent of attention. The pertinence of the MMN in the
study of speech segmentation through statistical means is made stronger by recent
evidence that suggests the MMN may be more closely related to auditory regularities
than to deviance detection (Winkler & Czigler, 1998). The MMN is also sensitive to the
degree of probability, with higher probable standards eliciting larger amplitudes
(Naatanen, 1992; Naatanen, Sams, Jarvilehto, & Soininen; 1983; Sato, Yabe, Hiruma, et
a1,2000)

Several recent studies using the MMN have begun to examine detection of
deviancy within a repeating auditory pattern or patterns. The investigation of pattern
violation detection is important to the understanding of auditory object formation and
subsequent segmentation from the acoustic environment. It is these auditory patterns that
are formed into gestalts, or auditory objects (see Ritter, Deacon, Gomes, Javitt, &
Vaughan, 1995 for a discussion of auditory stimuli stored as gestalts in a distinct
physiologic form). The detection of a deviant event within an acoustically varying

environment (such as that presented by many of these auditory patterns) requires that

20

these patterns be identiﬁed; an auditory object, a ﬁgure, needs to emerge from the
ground. Violation to this ﬁgure can then be readily perceived. The elicitation of the
MMN requires detection of the deviant event; and therefore, perception of auditory
objects within the varying acoustic environment.

Vaz Pato, Jones, Perez, & Sprague (2002) presented adults with synthesized
musical instrument tones in a continuous sequence of four and of ﬁve tones in a repeated
rising or pseudorandom pattern at a rate of 16 tones/second. Deviant tones occurred after
every 20 tones and were of higher pitch. A MMN was elicited to the deviant
immediately following the standard pattern. These results suggest the segmentation of
sequential information from repetitive and pseudorandom patterns. In this study, the
standard was not individual tones, as was the case in many early studies examining the
MMN, but was a repeated sequence of tones. The elicitation of the MMN occurred with
a violation of this pattern (due to presentation of the deviant, a higher pitched stimulus),
suggesting two conclusions. First, the repetitive standard pattern was recognized as one
auditory object due to the high probability of those tones occurring together. Second, it
suggests that these auditory objects can be formed in the environment of continuous
auditory streams, independent of other acoustic cues such as pauses indicating the end of
the sequence.

The concept of auditory object formation has further support in the literature.
Atienza, Cantero, Grau, Gomez, Dominguez-Marin, and Escera (2003) showed that
sequential grouping of auditory stimuli does occur. Also, memory traces of the MN are
not only formed for individual tones, but are also present for larger acoustic events that

contain several sounds and span larger temporal scales. This supports the results of Vaz

21

Pato, et al. (2002) which suggested that patterns of auditory stimuli presented as
standards can be grouped perceptually as auditory objects, and Winkler, Schroger, and
Cowan (2001) also support the concept of preperceptual organization for auditory events.

Winkler, Cowan, Csepe, Czigler, and Naatanen (1996) demonstrated that these
auditory objects are preserved over the course of different presentations of the auditory
object, and do not need to be recalculated each time the auditory sequence is encountered.
Trains of 6 tones (ﬁve standard tones and one deviant) were presented to subjects with an
intertrain interval of 9.5 seconds. If the pattern needed to be reestablished each time, then
the initial presentation of at least two or three standards would have been required to
detect a deviant. However, Winkler and colleagues found that a single reminder of the
standard tone reactivates the representation of the standard stimulus; a deviant in the
second position elicited the MMN. These results indicate that a sensory memory trace,
indexed by the MMN, is able to store complex auditory events for extended periods of
time; thereby, allowing for the establishment and maintenance of auditory objects. This
paper proposes that these auditory objects are not transient sensory traces subject to
almost immediate decay, but remain established in a memory store for later comparison
and organizational processes mediated through statistical means.

Auditory objects can also include other information, in addition to the spectral
patterns discussed thus far. Alain, Cortese, and Picton (1999) presented a continuously
repeating sequence of four tones where deviants varied in either frequency or timing.
Both types of deviants generated the MMN with identical scalp topography, suggesting
that spectral and temporal relations of auditory patterns are encoded in a uniﬁed memory

trace (Alain, et al., 1999).

22

Presented thus far is electrophysiological evidence that individual acoustic events
can be grouped into larger auditory objects on the basis of the statistical regularity of
their pattern, that this object formation occurs within continuous presentation
environments requiring the identiﬁcation and segmentation of individual auditory
patterns, and that auditory objects can simultaneously include spectral and temporal
information. Furthermore, the MMN, a preattentional index of auditory memory, is
sensitive to probabilities, detects these auditory patterns, and is preserved over time for
use on later pattern detection and comparison processes.

Also, multiple acoustic patterns can be maintained at the same time (Brattico,
Winkler, Naatanen, Paavilainen, & Tervaniemi, 2002). Brattico, et al. presented two
different standard sequences of four tones. Unlike many previous studies that used a
different auditory stimulus for the deviant, they constructed their deviant from the ﬁrst
two tones of one standard and the last two tones from the second standard. Therefore, the
deviant constructed was not a violation of the auditory event, but was a violation in
pattern (similar to the partwords used to detect segmentation by Safﬁan Aslin &
Newport, 1996). Only a system capable of recognizing and organizing these patterns into
perceptual objects would be capable of detecting this deviance. The MMN was elicited
by this pattern deviance in the absence of a change in the acoustic environment. These
results also indicate that multiple sound patterns can be maintained during an automatic
stage of auditory processing, as two different standard patterns were maintained. This
clearly demonstrates the formation and maintenance of multiple auditory objects that

were segmented from the presentation of a continuous auditory stream. This study

23

warrants ﬁrrther investigation of pattern deviancy detection and multiple sequence
maintenance with more complex auditory signals, such as speech.

While studies using pure-tone stimuli have provided an excellent discussion of
auditory objects, Rauschecker (1998) discussed their limitation. These studies are unable
to examine how complex sounds are re-assembled, lead to perceptual or cognitive
performance, or form memory traces for the recognition of complex sounds
(Rauschecker, 1998). The current study investigates these issues within a

neuropsychological framework.

Electrophysiology and Neuropsychology of Speech Segmentation

Speech segmentation studies have been limited to an investigation of
segmentation ability after the speech stream has already been parsed using the proposed
speech cues. These studies are therefore unable to establish the time course of the
perceptual processes performed (Sanders, Newport, & Neville, 2002). Few studies have
attempted to investigate the online perception of these cues, or the process of identifying
the auditory objects (word-forms) within the stream. (This paper uses the term word-
fonn to refer to auditory objects that are the prelexical auditory structure of words.)
These issues have not been addressed because relatively few behavioral methods have
been used to investigate speech segmentation, and these have all been limited to studying
ofﬂine detection of isolated words. Ofﬂine tasks use stimuli isolated from the natural
context of a continuous speech stream to assess speech segmentation. This project will
investigate segmentation through online protocols that embed the stimuli within

continuous presentations.

24

The concept of the word as an auditory object is not new. As early as 1970,
Hayes and Clark discussed the word as a perceptual object of language. From their
experiments, they concluded that a primitive mechanism for clustering continuous speech
into words exists for segmentation. They state that “this mechanism appears to operate
only on the information available in the intercorrelations between successive sounds in
the speech stream; it identiﬁes as words the clusters of sounds that consistently recur in
an unbroken sequence” (Hayes & Clark, 1970, p. 233). Therefore, the concept of using
statistical information to unite individual acoustic events into perceptual word-forms is
long standing. Even as early as 1899, Bryan and Hatter discussed how repetitious
exposure could train people to perceive the individual tones in morse code as whole
words and phrases.

Electrophysiological methods have the advantage of directly indexing the neural
correlates of perceptual abilities. The results of these studies provide great insight into
how the perceptual abilities of the brain work. Therefore, evidence ﬁom these studies
offer a neurophysiological framework with which to integrate behavioral evidence into a
working theory of the mind. This study attempts to take electrophysiological evidence of
auditory object formation and explain how it applies to a behavioral investigation of
speech segmentation. It is from an accurate understanding of the neurophysiological
correlates involved in perception that accurate neuropsychological test protocols may be
developed to behaviorally examine these processes in higher-order tasks such as speech

segmentation, and from these results infer neuropsychological mechanisms.

Electrophysiology

25

Perhaps the electrophysiological study most relevant to speech segmentation was
conducted by Sanders, et al. (2002) in an adaptation of a behavioral investigation done by
Saffran, Aslin, et al. in 1996. The purpose of this study was to conﬁrm that larger N100$
are evoked by word onsets. They recorded ERPs before and after participants had
learned to segment an artiﬁcial speech stream identical to that created by Saffran, Aslin,
et a1. (1996). Sanders, et a1. (2002) discovered a larger N100 for high learners
(participants with the greatest behavioral change in segmentation performance) after as
compared to before training participants to segment. This study by Sanders, et a1. (2002)
suggests that early ERPs can index perceptual changes developed through learning and
provides an electrophysiological correlate to measure segmentation ability. However,
this study did not address the ability of incidental language learning shown in Saffran,
Aslin, et al.’s (1996) study, as Sanders, et a1. (2002) actively trained participants on the
correct segmentation. Also, a larger N100 in this study may be more related to
perceptual features at word onsets rather than an actual formulation and detection of
auditory objects necessary for word identiﬁcation. In order to examine this, the actual
formation of words as entire auditory objects needs to be assessed; an investigation that
can be achieved using the mismatch negativity. This claim is based upon the precedent
of using MMN in the study of other similar auditory patterns.

An overlooked area in behavioral speech segmentation research is how perception
might change after successful parsing of the speech stream is learned. As the study by
Sanders, et a1. (2002) indicates through an electrophysiological paradigm, the perception
of the speech stream can be recorded before word-forms are identiﬁed and then compared

to perception after segmentation properties are learned. Such a study would be able to

26

examine how learning word boundaries might inﬂuence perceptual processing.
Experience has already been shown to inﬂuence the MMN. In a study by Naatanen,
Schroger, Karakas, Tervaniem and Paavilainen (1993) a standard sequence of 8
consecutive segments of different frequencies was presented. In the deviant condition,
the sixth segment had a different frequency. During the beginning of the study, the
MN was not elicited to the deviant. However, after continued exposure, participants
who learned to discriminate the sequences revealed a MMN to the deviant at the end of
exposure. This demonstrates that the MMN can index changes in the ability to
discriminate auditory sequences. Evidence such as this can be built into

neuropsychological paradigms to measure changes in performance ability.

Neuropsychology

Electrophysiological investigations add to our understanding of the perceptual
processes involved in speech segmentation by providing a framework for the time-course
of segmentation properties and the examination of these properties online during auditory
object acquisition. Evidence from electrophysiological studies provides a larger picture
of the perceptual processes that underlie speech segmentation. The unique abilities of
electrophysiological methods can be adapted to ﬁt a neuropsychological paradigm.
Speech segmentation abilities have been measured using high amplitude sucking,
headtum preference procedures, and two-interval forced choice paradigms. However,
these methods are only able to show that acoustic or statistical cues were identiﬁed
during exposure and that these properties are detected later when presented in isolation.

They lack the ability to examine the perceptual processes that support the acquisition of

27

novel words within the speech stream. Though they have shown speech segmentation
through various mechanisms, they also lack an adequate theory of how the segmented
syllables are clustered into single, unitary objects that may later become lexicalized. This
is despite the evidence that these perceptual objects, segmented and formed from the
speech stream, are treated as English words (Saffran, 2001).

The distinction between processes of discrimination, recognition, and
identiﬁcation must also be discussed brieﬂy. Grifﬁths (2003) organized complex sound
perception into a processing hierarchy. The levels are: l) the simple processing of
acoustic features, 2) complex processing of temporal, spectral, and spatial patterns (which
is the level of the auditory object), and 3) semantic processing (or the symbolic use of
auditory objects). It is at this last level that meaning is associated with prelexical objects.
Craik and Lockhart (1972) proposed that memory is a bi-product of the level of
processing. Thus, different processing may recruit different memory stores.
Discrimination, recognition, and identiﬁcation are interactive and related, but separate
levels of processing that are involved in speech segmentation and auditory object
formation. Bumham, Earnshaw, and Quinn (1987) state that “. . .different neural and
cognitive processes may be involved in discrimination and identiﬁcation” (p. 256).

DiscrimirLtion is the detection of any difference between two tokens (Bumham,
Earnshaw, & Quinn, 1987). Thus, discrimination is a very early process that enables later
perception, but does not necessarily lead to it. The method of a two-interval forced-
choice task is a discrimination paradigm in which the participant detects the difference

between two stimuli. Other methods, such as the head-tum preference procedure or high

28

amplitude sucking, also indicate this preperceptual process of detection. Discrimination
of stimuli uses an early sensory memory store (Winkler, et al., 1996).

Recognition on the other hand, is the implicit knowledge that a stimulus has been
encountered before (Galotti, 1999). It is the early implicit detection of a recurring
structure of stimulus features. Recognition is at higher level than discrimination as it is
no longer preperceptual detection, but requires the integration of features into an early
structure. According to Craik and Lockhart (1972), this higher processing level would
recruit memory resources different than the early sensory memory store implicated in
discrimination.

Identiﬁcation “. . .involves ﬁrst the abstraction of common bases for grouping

 

sounds together, and then the comparison of each token with this abstraction (Bumham,
Earnshaw, & Quinn, 1987, p. 256). This is no longer a lower, preperceptual process like
discrimination, or an implicit detection of pattern structure like recognition.

Identiﬁcation requires an abstraction representation of the object. It examines the holistic
characteristics of the stimulus and requires an explicit knowledge structure; or in other
words, identiﬁcation labels the stimulus as an object. This is a higher-order process that
recruits memory resources for maintaining an abstract representation of the object, which
is penetrable to cognitive tasks.

This study proposes to measure detection, by replicating previous studies, and
build upon them by also measuring recognition and identiﬁcation in two different online
tasks where the participant must sort through distraction stimuli to perceive the learned
words online. These later two conditions implicate different requirements of memory

that are associated with their level of processing.

29

The neuropsychological methods designed for this project measure recognition
and identiﬁcation of word-forms before and after learning segmentation, examine
segmentation abilities online, and investigate the formation of words as auditory objects
in a recognition task. This paper makes the distinction between online and ofﬂine tasks
as such: ofﬂine tasks extract the stimulus from the continuous speech stream and present
it in isolation; whereas, online tasks present the stimulus within the speech stream. Thus,
online tasks more closely model perception during natural speech contexts and actively
require speech segmentation processes to occur during testing. Ofﬂine tasks already have
the stimulus segmented during the test presentation. This study also utilizes a cognitively
more complex identiﬁcation task that requires phonological synthesis (the conceptual
reformation of a word-fonn from its individual elements) of these stored auditory objects.
Future studies may wish to examine the electrophysiological substrates that are involved
in these processes.

This paper proposes that the online recognition and identiﬁcation of words
suggests successful speech segmentation. This furthers earlier reports using isolated
stimuli in discrimination tasks; such tasks may index the detection of familiar patterns or
the presence of a less probable feature (a word-boundary), rather than the identiﬁcation of
a full auditory object — the word (Fiser & Aslin, 2002). This is the difference between
feature detection and gestalt processing abilities. The current project examined this
difference by using tasks structured on a hierarchy of processing difﬁculty.

The ﬁrst level of this processing hierarchy examined is discrimination. A two-
interval forced-choice (ZIFC) paradigm was used as an ofﬂine discrimination task like

that used by Saffran, Newport, et al. (1996) to conﬁrm the results of this study in

30

comparison to previous studies. Successful performance on this test would conﬁrm the
occurrence of segmentation. As stated earlier, this paradigm provides evidence for a
feature detection of statistical segmentation cues that were learned during exposure.
Unlike the next two testing paradigms, this task does not require gestalt auditory
processing (as in the Auditory Probabilistic Word Count) or higher cognitive tasks such
as phonological synthesis (as in the Slow Online Segmentation Test), but is a relatively
early process of discrimination and can only infer word boundary detection, not word-
fonn learning. Both of these tasks involve higher-order, online processing of learned
auditory objects. The combinations of these testing methods thus allow for the
dissociation of segmentation, object perception, and the hierarchy of perceptual/
cognitive processes.

The Auditory Probabilistic Word Count (APWC) is an online measure of
recognition, the second level of the hierarchy. This task was designed from the
probabilistic paradigms of ERP studies to examine detection of low occurrence auditory
objects (from the artiﬁcial language) within highly frequent randomized syllables. Thus,
this task is very similar to odd-ball paradigms used for auditory deviance detection within
continuous complex auditory sequences. The participant’s task is to count the number of
artiﬁcial words (learned from the exposure period) that they hear within a continuous
speech stream presented at a normal rate and composed mostly of randomized isolated
syllables. This method conjectures that if memory traces for the exposed words were
formed during exposure using the statistical distribution of syllables (which required
word segmentation), then this same probabilistic information should be accessible to

successfully segment the learned words in a novel language exposure with distracters

31

(which are randomized isolated syllables from the exposed language). To accurately
detect the learned words, the participant must have already integrated the statistical
distributions of the syllables into a implicit, early structural formation of a perceptual
object (word-form). As such, this is a measure of auditory gestalt processing. An
accurate count across subjects is dependent upon successful segmentation performance.
The third level of processing was assessed using the Slow Online Segmentation
Test (SOST). This protocol (described in further detail in the methods section) requires
the participant to identify word boundaries online during the continuous presentation of
isolated syllables that follow the same probabilistic structure of the exposure speech
stream. In order to perform successfully at this task, participants are required to combine
individual speech syllables into one of the auditory objects learned during exposure.
They then must identify the object boundary within a several sentence-length
presentations of other syllables belonging to other learned objects or used as distracters.
Support for this method comes from the phonological processing literature that is
typically used to investigate early reading abilities. However, phonological processing
abilities (speciﬁcally phonological awareness and memory) are arguably essential for the
timely acquisition of a novel language. Phonological awareness is the explicit knowledge
of the language’s phonological features (Bradley & Bryant, 1985). It is the awareness of
individual phonemes that comprise words and requires the same level of mental
representation that is required for identiﬁcation processing. The auditory objects, in
contrast to discrimination and recognition processing, need to have an explicit symbolic
structure for identiﬁcation, and may imply the recruitment of prelexical semantic memory

l‘CSOlll’ C CS.

32

Phonological synthesis, typically studied using Sound Blending and Nonword
Blending subtests in neuropsychological test protocols, is the ability to combine an
isolated set of phonemes together into a recognizable word (Torgesen, Morgan, & Davis,
1992; Wagner, Torgesen, Rashotte, Hecht, Barker, Burgess, et al., 1997). For an example
of a phonological synthesis task, the participant might be asked, “What word is this: /k/,
/a=:/, /t/?” The correct response would be “cat”. Evidence suggests that the syllable is the
unit that children ﬁrst become phonologically aware of (McClure, F erreira, & Bisanz,
1996; Wagner, et al., 1997), which complements the current study’s use of syllables as
the individual features of the novel words.

Tests of phonological awareness and auditory processing have traditionally used
phonological synthesis as an index of how we integrate and say whole words (CTOPP;
Wagner, Torgesen, & Rashotte, 1999; WJ-R; Woodcock & Johnson, 1989, 1990).
Wagner, et a1. (1997) also used Sound Blending and Nonword Blending to measure
phonological synthesis. Though typically measured as correlating with reading ability,
this synthetic process enables learning and is not reciprocally facilitated by later
acquisition of reading. It is described as an aspect of phonological sensitivity which is a
very basic recognition of the phonological elements in oral language (Wagner, Torgesen,
& Rashotte, 1994). Thus, it is appropriate to use in an oral language task to study
integrating isolated syllables into an auditory object.

The development of the SOST was based upon this concept. In this task, the
syllables, presented at a slow rate, must be phonologically synthesized into words (as in
the Sound Blending and Nonword Blending tasks). Torgesen and Morgan (1990)

discussed a model of phonological synthesis outlined in a presentation by Perfetti, Beck,

33

and Hughes in 1981. The model suggests several processes that are involved in
phonological synthesis. The most basic component is awareness that individual sounds
can be combined in various ways to form words. But, these sounds must also be
represented and stored in memory and then combined to form a word-like representation.
However, during the SOST, participants must be able to do these processes online during
the continuous presentation of sentence-length speech streams. Thus, in addition to
phonologically synthesizing these syllables into words corresponding to the auditory
object memory traces, the participants must also be able to segment these words from the
speech stream. Success at this task would support statistical speech segmentation, as well
as, the formation of auditory objects. (A standardized screening of phonological memory
and phonological awareness were conducted with participants to compare each
participant’s phonological abilities to a normative sample.)

Auditory perception is inﬂuenced by presentation rates, as reﬂected in
electrophysiological studies which vary SOA times (Naatanen, 1992). By testing in a
temporal window that differs from exposure rates, the stimuli are no longer perceived in
the same relation to each other. Therefore, the successful identiﬁcation and segmentation
of the learned words is dependent upon ﬁrst synthesizing the independent syllables. This
gestalt processing of auditory sequences has been shown to inﬂuence the perception of
auditory spaces by perceptively shrinking the amount of time between stimuli (Sasaki,
Suetomi, Nakajima, & Hoopen, 2002). If only the statistical features are learned during
exposure, and no memory traces for perceptual objects are formed, then the isolated
syllables will not be able to be phonologically synthesized appropriately within the

continuous speech stream.

34

While the 2IF C was an ofﬂine discrimination task, the APWC and SOST assessed
participants’ ability to segment speech online, during continuous presentations. These
online tasks were measured through a pretest-posttest design. Pre- and post-exposure
tests are relatively straightforward. As discussed, electrophysiological studies have
provided evidence for how perceptual processes change as a result of exposure to
stimulus patterns (N aatanen, et al., 1993). In the current study, any perceptual changes
that occur are a result of the exposure to the probabilistic structure of the artiﬁcial
language. Improved performance is therefore related to changes in perceptual ability.
This paper predicts that a wider distribution of scores and poorer accuracy will be
measured during the pretest as opposed to measurements taken after exposure. It also
predicts that processing for the artiﬁcial language will become more like native language
processing. This will be assessed through comparisons with performance on isomorphic
tests using English words.

While these English tests measure native processing abilities, their primary
purpose is to serve as a measure of external validity for the APWC and SOST
neuropsychological measures. The participants performed these two tasks with low
frequency English words. Low frequency words reduce the confound of lexical access
facilitating performance (which is not available in the experimental tasks). If the
participants were unable to successfully perform these tasks with English words, then
they would almost certainly perform poorly with the experimental stimuli. However, if
they were able to perform accurately in English, then they should also have the potential

to perform accurately on the experimental stimuli, given that they have been able to

35

segment the artiﬁcial language using the statistical distributions. These English tests
were considered a measure of optimal segmentation ability.

Research on phonological processing has also revealed differences in the
processing and even localization of function between males and females. Data from
functional imaging studies has revealed more bilateral activation during phonetic
processing tasks in females; whereas, males primarily activated the left hemisphere
(Pugh, Shaywitz, Shaywitz, Constable, Skudlarski, Fulbright, et a1, 1996; Shaywitz,
Shaywitz, Pugh, Constable, & Skudlarski, 1995). Structural MRI studies have also
revealed that the left planum temporale is signiﬁcantly larger in males than females
(Kulynych, Vladar, Jones, & Weinberger, 1994), which is believed to be involved with
language components that require rapid temporal processing (Lambe, 1999).

Neuropsychological studies have also discovered differences in processing. In
dichotic listening tasks, a signiﬁcantly stronger right ear advantage occurs for right-
handed males than for right—handed females (Lake & Bryden, 1976). The right ear
advantage is also produced for stop consonants which have rapidly changing acoustic
cues, rather than for steady state vowels (Fitch, Miller, & Tallal, 1997). Coney (2002)
also observed that females are faster to respond in a nonword rhyming task in which
successive rhyming words were presented to different visual ﬁelds, which is consistent
with the view that females may possess “a greater facility for dual hemispheric
processing in phonological operations” (p. 363). These results of anatomical, functional,
and behavioral differences in the ways males and females processes phonological
information led Lambe (1999) to recommend the separation of males and females in

research studies to minimize variance; thus, enabling subtle differences that are

36

functionally relevant to be detected. In following this suggestion, this study only used
females as participants, with the hope that future studies will be able to examine a more

diversiﬁed population.

Proposal for the Current Study

This study is designed to address several questions within the speech
segmentation literature. It examines online auditory object perception related to the
statistical distribution of syllables within and between artiﬁcial words. The
neuropsychological protocols developed measured changes in perceptual ability during
the acquisition of speech segmentation abilities through pre- and post-leanring
measurements. Such evidence would suggest that perceptual processes change due to
exposure to the statistical probabilities inherent in the speech stream.

Current trends in the literature support the rationale for investigating these issues
using more advanced behavioral methods. While continued support for statistical
learning from the behavioral literature is established, an expansion of the
neuropsychological methods is required to more fully investigate this phenomenon.

The current proposal investigates these issues through the use of a
neuropsychological paradigm. Participants will listen to a continuous auditory stream of
an artiﬁcial language analogous to the speech stream created by Saffran, Aslin, et a1.

( 1 996).
The current study was designed to examine the following areas:

1) Segmentation of complex patterns via statistical regularities

37

This study will test the use of statistical regularities in analysis of complex

stimulus patterns, namely speech syllable distributions, and to segment those

patterns within a continuous auditory stream that offers no acoustic breaks.
Participants will be exposed to 3-syllable artiﬁcial words that are
concatenated together in a pseudo-random order. The only cue to word
boundaries will be the statistical regularities between syllables. Syllables
within words will always occur together, p=1.0. Syllables between words
will only occasionally occur, p=0.33. Participants must therefore learn the
statistical regularity to segment the artiﬁcial words.

2) Replication of discrimination studies

This study will conﬁrm the ability to discriminate segmented words from foils by

expanding upon previous studies.
The Two-Interval Forced Choice paradigm has been used in the literature
to indicate that speech segmentation did occur during exposure and that
participants are able to identify exposed words in an ofﬂine discrimination
task (Saffran, Johnson, et al., 1999; Saffran, Newport, et al., 1996). This
paradigm will be replicated to conﬁrm results with previous research and
identify the particular words that have been learned. It also tests the
ability of statistical probabilities to inﬂuence a lower-level feature-
detection process of discrimination.

3) Object recognition and identification
This study will investigate the online ability to recognize and identify auditory

objects.

38

Participants’ ability to identify the artiﬁcial words online during
continuous presentation of the speech stream will be measured using the
Auditory Probabilistic Word Count and Slow Online Segmentation Test.
These tests examine the ability of statistical probabilities to inﬂuence
hi gher-level gestalt processes of recognition and identiﬁcation.
4) Changes in performance due to enhanced perception

This study will identify changes in performance related to an enhanced ability to

perceive perceptual units.
Participants’ ability to identify artiﬁcial words will be tested pre- and post-
exposure to the continuous stream using the Auditory Probabilistic Word
Count and the Slow Online Segmentation Test. Low accuracy scores
pretest and signiﬁcantly improved scores posttest would indicate leanring
the perceptual structure of the artiﬁcial words.

5) Manipulability of auditory object memory traces

This study will examine the manipulability of memory traces for the learned

auditory objects in hi gher-order processing.
Temporally separated syllables will be presented to the participants at a
continuous rate. As temporal information affects our perception of
auditory events, for the participants to identify word boundaries they must
ﬁrst phonologically synthesize the individual syllables into the learned
auditory objects maintained as memory traces from exposure. Also, the
perceptual hierarchy of discrimination, recognition, and identiﬁcation will

examine what representation levels auditory objects can inﬂuence.

39

The present study investigates these perceptual changes within an artiﬁcial speech
stream. Neuropsychological protocols will measure perceptual processes pre- and post-
segmentation learning. This study hypothesizes that the statistical properties of speech
will provide for the formation of auditory objects; thereby, inﬂuencing perceptual states
and enabling successful behavioral performance in identifying artiﬁcial words.
Therefore, a quantiﬁable difference in behavioral performance is predicted to occur

before and after speech segmentation is learned.

4O

CHAPTER II

Methods

Participants

Twenty-two college students participated in this study; however, two subjects
were eliminated from analyses due to demonstrated and reported fatigue by one subject,
and failure to complete the testing protocols by the second. Implications for these
exclusions are addressed later in the discussion of this paper. Prior to selection,
participants completed a short survey via phone or in person on previous
language/auditory training and a medical history. Criteria for exclusion included: ﬂuency
in a language other than English; previous speech, language, hearing, or reading
difﬁculties; or presence of a developmental disorder. All participants were monolingual
English females by self-report and were either paid for their participation or received
course credit. Females were selected due to the demonstrated anatomical, functional, and
behavioral differences between males and females that have been demonstrated in the
literature (see the introduction for a discussion). All participants passed a bilateral pure-
tone hearing screening at octave frequencies 500 to 8000 Hz presented at 20 dB HL
(ANSI, 1996) and had otoscopy performed to verify the ear canal was free from
occlusions.

To be included in the study, all participants were required to achieve standard
scores of an 8 or above (representing average or above phonological processing abilities)

on Memory for Digits and Blending Nonwords subtests of the Comprehensive Test of

41

Phonological Processing (CTOPP, Wagner, Torgesen, & Roshotte, 1999). Participants
also completed the Nonword Repetition subtest to achieve a Phonological Memory
Composite score. Blending Nonwords is an acknowledged test of phonological
synthesis, while Memory for Digits and Nonword Repetition are recognized measures of
Phonological Memory (Wagner, Torgesen, & Roshotte, 1999). They also completed the
Matrices subtest of the Kaufman Brief Intelligence Test which is a nonverbal measure of

ﬂuid thinking (K-BIT, Kaufman & Kauﬁnan, 1990).

Stimuli
Speech Synthesis

Speech syllables were created using the built-in speech synthesizer on a
Macintosh OSX system. The speech output was synthesized using a MacInTalk 3 female
voice (Kathy) and converted to a digital audio ﬁle using Voice Box 1.4. Audio ﬁles were
then digitally edited to control for amplitude, fundamental frequency, and duration of
each syllable. Edited syllables were strung together to form the three-syllable words
using a waveform editor and saved as .wav ﬁles. Distracter syllables used in the
neuropsychological protocols were individual syllables from the medial position of the
words. All syllables were 333 ms in duration, had a fundamental frequency of 186 Hz,
and an amplitude with an average root mean squared (RMS)=.2200 during the steady
state of the vowel and a RMS=.17 74 across a concatenation of all 12 syllables. English
stimuli consisted of low frequency 3-syllable English words (as provided in the Word
Frequency Book, Carroll, Davies, & Richman, 1971) to more closely model the artiﬁcial

language and were also synthesized using this method. Low frequency words were

42

determined to be both a better control and a measure of optimal performance as they
reduced the confound of lexical activation assisting in segmentation.

Pilot data on these syllables was collected using an identiﬁcation task where
participants transcribed each syllable that they heard. The experimental syllables were
repeated ﬁve times with a 0.167 second ISI before the participant was required to identify
the syllable. Each experimental syllable was identiﬁed three times for a total of 36 items.

Results of this pilot study are discussed in the following chapter.

Language Construction

Speech syllables were assembled to form four tri-syllabic words, as in Saffran,
Aslin, et a1. (1996). Two counterbalanced conditions were created in order to control for
any perceptual biases unrelated to the experimental investigation (Language A: daropi,
pabiku, golatu, tibudo; Language B: bikuti, tudaro, pigola, budopa). Both languages
were constructed identically according to the methods described here. An equal number
of participants heard each language. Participants listened to the same language during
each exposure period and during testing.

Tri-syllabic words were concatenated together in pseudo-random order to form
three different blocks of 7-minutes each with the stipulation that the same word never
occurred twice in a row. This created a total of 21 minutes of exposure to the
probabilistic structure of the artiﬁcial language. No pauses or other acoustic cues to word
boundaries occurred between syllables. The only cues to word boundaries were the
statistical probabilities between words. An orthographic representation of the resulting

speech stream is as follows: pabikutibudogolatudaropi. . ..

43

Transitional Probabilities

In Saffran, Aslin, et al. (1996) the internal probability between syllables within a
word (e.g., pabi and biku in pabiku) was p=1 .0 because pa was always followed by bi
(and bi always by ku) in the speech stream. The external transitional probabilities used
by Saffran, Aslin, et a1. (1996) for syllables spanning a word boundary (e. g., ku#ti in
pabiku#tibudo) were p=0.33. This was because the ﬁnal syllable (in this case ku) could
be followed by any of three different initial-position syllables (ti, go, or da). This
external probability of p=0.33 was maintained in the present study, as the same four ﬁnal
syllables were each followed by one of three initial syllables, and as a word could not

follow itself.

Neuropsychological Test Materials

The Auditory Probabilistic Word Count (APWC) consisted of a speech stream of
approximately 3 minutes in duration and was presented at the same presentation rate as
the exposure speech stream (i.e., 3 syllables per second). Syllables that occurred in the
medial position of the exposed words were randomly presented with the 3-syllable words
from exposure (either from Language A or Language B). The four exposed, 3-syllable
words had a 20% chance of being presented, as lower probabilities yield a stronger
deviant, and therefore allowed for easier detection of the attended stimulus (the word).
Five different presentations were generated: one English version, and a pre- and posttest

each for Language A and B.

44

The Slow Online Segmentation Test (SOST) consisted of 25 isolated speech
streams that were approximately sentence length. Each stream consisted of 15 syllables
which formed 4 words and 3 distracter syllables (to prevent pattern responses) for a total
of 100 words to identify. Syllables were presented at a rate of 2 syllables per second (one
every 500 ms), which is the rate that has been used in phonological synthesis tasks
(Wagner, et al., 1997), and were in a pseudo-random order (with the stipulation that no
stimuli occurred twice in a row). The distracter syllables were from the medial position
of the exposed words; thus, bi became a distracter from the word pabiku. Each speech
stream started with a tone (a sine wave of 1 second in duration) followed by one second
of silence, and ended with a 500 ms pause immediately followed by a series of 5 beeps
(ﬁve triangle waves at 150 Hz with a duration of 200 ms each). An interval of 4 seconds
passed between streams. These prevented participants from losing their place during the
continuous presentation and to control the amount of time available for stream analysis;
as this test was designed to measure online processing, not ofﬂine reanalysis of the
stream. Again, ﬁve test streams were created: one English version, and pre- and posttests
for the two artiﬁcial languages. A response sheet containing 25 rows of boxes (15 boxes
per row) was created for participants, each box represented one syllable and each row

corresponded to each speech stream.

Design & Procedure
Participants attended one session that lasted approximately 120 minutes, including
scheduled breaks. The hearing screening and all experimental tasks were completed in a

sound attenuated room. All protocol directions were audio recorded and presented prior

45

to each testing section, in addition to the written transcript being provided at the
beginning of each protocol response form. All stimulus presentations and test protocols
were transferred onto a single CD-Rom, played on a KLH 5-disc CD changer player
(DA1602), and presented via ER-3A 500 insert earphones with E-A-R LINK foam
eartips positioned by the experimenter. The system was calibrated weekly to present the
syllables within a 2 dB difference of 60 dB SPL using a Larson Davis Laboratories
precision integrating sound level meter (Model 800B). The calibration was done using a
sine wave generated at the syllable fundamental frequency of 186 Hz and RMS of 0.22.
This calibration setting was then checked during a continuous presentation of the
syllables. A biologic check was also performed weekly.

Participants ﬁrst completed the standardized test screenings, followed by three
training tasks designed to familiarize the participants with the synthesized English words
and teach them how to complete the SOST and APWC tests. These training tasks were
performed before any experimental testing began.

All participants completed pretests of the APWC and SOST. Then, they listened
to 3 blocks of a 7 -minute artiﬁcial language exposure (described above) which was
interspersed with English versions of the APWC and SOST. Next, participants received
posttests of the APWC and SOST and completed testing with the two-interval forced-
choice task which replicated Saffran, Newport, et al.’s (1996) method. Pretest and
posttest versions of the APWC and SOST were counterbalanced with each other. The
APWC always preceded presentation of the SOST as this was viewed as a possible initial
exposure period. (The APWC was selected for ﬁrst presentation as it did not require the

explicit identiﬁcation of word-forms that the SOST requires). Two 5-minute breaks were

46

suggested and made available to participants, though they were not mandatory: one
occurred following the pretests and the other occurred right before the third exposure to
the 7-minute speech stream. The entire testing session lasted approximately 2 hours.

Further descriptions of these training tasks and experimental tests follow.

Neuropsychological Training Tasks

The English Word F amiliarization was a task that familiarized participants with
the synthesized nature of the speech they would hear. Subjects were read off a word list
in natural speech and its synthetic counterpart of eight words (four English control words
and four additional example words). The four English control words were included, as
pilot tasks indicated that some participants had difﬁculty deciphering the synthetic nature
of the speech. As the English control tasks were a measure of optimal ability,
familiarization of the synthetic quality (which subjects had never heard) was necessary to
measure optimal task ability without the confound of perceptual difﬁculty. However,
participants were not explicitly told which of the eight words would be heard during
testing. After presentation of the word list, participants completed an eight-item
identiﬁcation task where they numbered each of the eight words in the order that they
heard them. This task was repeated if the participant was not successful the ﬁrst time. All
participants reached the 100% accuracy criterion on this task, with the exception of four
participants who required two presentations. Participants also heard a reading of the
word list prior to each English protocol with natural and synthesized speech.

The APWC Training Task consisted of three examples using non-experimental

English words. Before each example, the participant was told the words which she would

47

hear in a continuous speech stream and given the synthesized stimulus. The participant
counted the number of times she heard those words within the sample speech stream.
Following the stream, the correct number of words was announced, and the participant
listened to the same speech stream a second time. This was performed for all three
examples, which built in complexity from listening from just two words in the ﬁrst
example to four words in the third.

The SOST Training Task consisted of four trials and three practice examples. The
training task followed the procedure of the SOST, except that the word(s) to segment
were provided aurally and in writing before each presentation. Trial tasks increased in
complexity and practice examples were directly identical to the SOST testing items with

the exception that the English training stimuli were used.

Behavioral Replication Testing

To assess behavioral performance of leaming segmentation for an artiﬁcial
language, a 32-item two-altemative-forced-choice (21FC) test was constructed, analogous
to the test used by Saffran, Newport, and Aslin (1996). For each test item, participants
heard two tri-syllabic strings, separated by an inter stimulus interval (ISI) of 500 ms.
One of the two strings was a word from the exposed language, while the other was a
partword. Partwords were formed from the syllables which spanned word boundaries.
The probability of these words occurring in the exposed speech stream was equal to the
transitional probability between words, p=0.33. The partwords that were tested were the
words from the unexposed language (as used in Aslin, Saffran, & Newport, 1998). For

example, the word tudaro from Language B occurred as a partword in Language A from

48

adjoining the words golatu#daropi. Thus, in testing a participant exposed to Language
A, golatu (a word) was compared against tudaro (a partword). Therefore, all
participants, regardless of language exposure, received the same behavioral test, although
correct responses were exactly the opposite for participants who received different
language exposures. The nature of the language allowed for the testing of two different
partword types. Initial partwords contained the initial two syllables of a word (e.g., the
initial partword tudggg contained the initial syllables ﬁom Mi) and ﬁnal partwords
contained the ﬁnal two syllables from a word (e.g., the ﬁnal partword _b_lk_l_lti contained the
ﬁnal syllables from the word paw).

During behavioral testing, participants were provided with a response sheet
containing three English training examples and 32 numbered experimental items. All
stimuli were presented to the participants via headphones. Participants were asked to
indicate which of the two strings was more familiar by circling a “1” or a “2” on the
response sheet corresponding to the indicated string. Three practice items were given to
the participants on English versus nonword stimuli to familiarize them with the task. As
with the SOST, a tone followed by one second of silence was played prior to the stimuli
presentation and a series of ﬁve beeps occurred at the end of item columns. Four seconds

‘61,,

elapsed between item presentations. For half of the items, was the correct answer,
and for the other half, “2” was correct. Each word and partword was paired exhaustively

with each other to yield a total of 16 items. Participants were tested on each item twice,

rendering 32 test items. The entire test lasted approximately six minutes.

Neuropsychological Testing

49

Auditory Probabilistic Word Count (APWC)

The APWC was given pre- and post-exposure and validity for the APWC was
measured using English words (E-APWC). Participants listened to the pre- and posttest
continuous syllable streams that were constructed respective of their exposed language,
and the English version. A total of 36 words were presented.

The participants’ task was to count the number of artiﬁcial words (learned from
the exposure period) that they heard within the continuous speech stream composed
mostly of randomized syllables from the medial position of the words. Participants drew
tick marks each time they identiﬁed a familiar word. If a memory trace for the exposed
words was formed (which required segmenting words) using the statistical distribution of
syllables, then this same probabilistic information should have been accessible to
successfully segment the learned words in a novel language exposure with distracters
(which were randomized isolated medial syllables from the exposed language). To
accurately detect the learned words, participants must have integrated the statistical
distributions of the syllables into a perceptual object (word). As such, this was a measure
of auditory gestalt processing. Participants were provided with a brief trial speech stream
of 10 seconds, which contained non-experimental English words embedded in it to
familiarize them with the task. The three minute speech stream was divided into six
segments of 30 seconds each. Participants recorded the number counted after each 30

second segment. The entire APWC test lasted approximately 5 minutes.

Slow Online Segmentation Test (SOST)

50

The SOST was given pre- and post-exposure to measure the amount of
improvement that could be attributed to exposure sessions. Validity for the SOST was
measured using English stimuli (E-SOST), instead of the artiﬁcial words to determine
optimal performance. This test allowed for testing segmentation ability online, rather
than discriminating isolated words, and investigated auditory object formation for
artiﬁcial words.

Participants, using a colored marker, followed along with the auditory speech
stream by placing a dot inside each box on the response sheet for every syllable heard
(syllables presented at a slowed rate of two per second), and a slash after a box upon
perceiving the end of a word. Thus, this test examined online segmentation of the
speech stream by requiring participants to synthesize temporally segmented syllables into
the learned auditory objects, and then identify the boundaries of these objects online.
Four practice items using non-experimental English training stimuli were provided before
beginning each test presentation (pre-, post-, and English), the ﬁrst two of which had the
correct answer provided to the participants so that they had a clear model of how they
were expected to complete the task. Responses to the practice items were monitored to
ensure appropriate responses (i.e. drawing a slash after word-ﬁnal syllables). The entire

test lasted approximately 9 minutes.

Data Analysis
The alpha level for all test comparisons was set at p=.05. In order to measure the
percent nonoverlap of the two samples, effect sizes were calculated for signiﬁcant results.

Cohen (1988) deﬁned levels of effect sizes using the d statistic: for small effect sizes

51

d=.2, for medium effect sizes d=.5, and for large effect sizes d=.8. The pooled sample
variance was used in calculating effect sizes. Bonferonni corrections were made for all
post-hoe comparisons using the English testing. Inter-rater reliability was also measured
by having 20% of the data rescored by an independent rater.

Scores obtained from the 21FC were the total correct (which was converted to
percent correct) and the number of false positives to ﬁnal and initial partwords. Means
and standard deviations were calculated on each of these scores. Analysis consisted of a
series of t-tests. The ﬁrst test compared percent correct between Language A and
Language B to determine if any factor other than the statistical nature of the language
(e. g., phonetic structure, acoustic characteristics, etc.) differentially impacted
performance between groups. Equal variances were tested between the languages and
then analyzed using the pooled sample variance. All other analyses were conducted by
pooling the two languages together as one group. A single-sample t-test was used to
determine if percent correct was signiﬁcantly above chance (50% correct) and a matched-
pairs t-test compared selection errors of initial and ﬁnal partwords.

The APWC had raw scores for the number of words counted at each of the six
speech streams. Absolute differences were calculated between each of the six raw scores
and the correct number of words in each stream. These six difference scores were then
summed to create the test score used in analyses. An absolute difference score of 0
indicated an approximation of perfect performance (as the exact words counted could not
be determined); thus, lower scores indicated better performance. A series of matched-

pairs t-tests were conducted on this data: pretest versus posttest, pretest versus English,

52

and posttest versus English. Means and standard deviations were also calculated for each
test presentation.

Raw scores for the SOST included the total number of words identiﬁed and the
total number of words correct. Percent correct was calculated by dividing the number of
words identiﬁed by the number correct. Matched-pairs t-tests were performed using the
number correct and the percent correct for the following comparisons: pretest versus
posttest, pretest versus English, and posttest versus English. Analyses using the English
condition as a control for performance were also conducted by dividing the test score by
the English score. This data for the pretest versus the posttest were analyzed using t-
tests. Means and standard deviations were also calculated for each test presentation on

both number correct and percent correct.

53

CHAPTER III

Results

Syllable Identification of Synthetic Speech

Results of the syllable identiﬁcation pilot study indicated that participants
identiﬁed syllables correctly 58% of the time across all syllables. Identiﬁcation accuracy
scores ranged from 0% for /ro/ to 100% for /do/ and /la/. Accuracy scores across all
participants and syllables increased to 81% when the most frequent response was
considered correct and response accuracy ranged from 47% (/go/) to 100% (/do, la, rol).
Within subject consistency scores (the percent of syllables an individual subject
identiﬁed the same across all three trials) was an average of 92% across all syllables;
ranging from 67% (/pa/) to 100% (/bi, bu, do, la, ro, ti/). Only two syllables were
consistently identiﬁed as another syllable within the artiﬁcial language (/go/ and /pi/ were
consistently identiﬁed as /do/ and /ti/ respectively). Three other syllables were identiﬁed
as non-language syllables (/bi/, /ku/, and /r0/ were identiﬁed as /di/, /pu/, and /10/
respectively). Identiﬁcation results with a smaller sample who heard the syllables within
the artiﬁcial words had similar results. Implications of these results are addressed in the

discussion of this paper.

Inter-rater Reliability

Inter-rater reliability was calculated for the individual test protocols. An

independent rater with no knowledge of the theoretical foundations of the tests scored a

54

sample of 20% of the data as a measure of inter-rater reliability. Raters were in 100%
agreement on the APWC and 21FC protocols, which was likely the result of very easy
scoring on the APWC and score rechecking procedures with the 2IFC. Total agreement
on the SOST was also high at 99%, with 99% agreement for the total number of words
counted and 98% agreement for the number of correct words counted. Due to the high
inter-rater agreement, in the event of a disagreement the ﬁrst score calculated was chosen

for use in the analysis.

Standardized Screening Measures

All twenty participants scored above 90 on the Matrices subtest of the Kaufman
Brief Intelligence Test (Kaufman & Kaufman, 1990), indicating at least average
nonverbal abilities (M=104, SD=8.0). All participants also scored above an 8 on the
Memory for Digits (M=12, SD=1.8) and Blending Nonwords (M=12, SD=2.3) subtests of
the Comprehensive Test of Phonological Processing (Wagner, Torgesen, & Roshotte,
1999), again indicating at least average abilities. The Phonological Memory Composite
score had a mean that fell within the average range (M=108, SD=9.9), where a score
between 90 and 110 indicates normal performance. The Nonword Repetition subtest
(M=1l, SD=1.8) was also administered as a non-criterion measure to obtain the

composite score.

Table 1. Standardized test scores

 

 

 

 

 

Memory N onword Memory Blending

Matrices for Digits Repetition Composite Nonwords
Mean 1 04 12 l l 1 08 12
Standard Deviation 8.0 1.8 1.8 9.9 2.3

 

 

 

 

 

55

 

Experimental Questions
An alpha level of p=.05 was used as the signiﬁcance level for all statistical tests.

Signiﬁcance probabilities are reported here.

Question 1: Segmentation of Complex Patterns via Statistical Regularities

This is a global question as to the amount of learning which occurred on the three
experimental measures; as such, speciﬁc results will be discussed in the sections that
follow. Language A and Language B were ﬁrst examined to determine if any factor other
than the statistical nature of the language (e. g., phonetic structure, acoustic
characteristics, etc.) differentially affected performance between groups. This was
examined using a t-test conducted between the Language A and Language B 21FC scores,
as used in previous research (Saffran, Newport, & Aslin, 1996). Results indicated no
signiﬁcant difference between the languages as detected by the 21FC test, t(l 8)=.56,
p=.58. These results suggest that there was no perceptual difference between the two
languages that impacted participants’ performances. As these languages were perceived
in the same way as predicted, the scores for the two languages were collapsed in the

analyses that follow.

Question 2: Replication of Discrimination Studies

For the 21F C Discrimination Task, the mean score was 20.3 of a possible 32 items
(63.4%), where chance performance equals 16. A single-sample t-test (two-tailed)
revealed an overall performance signiﬁcantly above chance with a large effect size:

t(l9)=5.15,p<.0001, d=l .08.

56

Figure 1. Mean scores for each participant on the 21FC

 

 

 

 

 

 

30_
25g :’
no
A .C.
Q o
u— 20_ coo
0 co.
‘5 o
3 0
Q8) 15H . =chance If
g- .. performance ‘
U
311:
10_<
i
l
5_ I
i
l.
0 ..

 

Previous research has demonstrated differential performance between
discrimination pairs when a partword contained the ﬁnal two syllables of a word versus
when a partword contained the initial two syllables of a word (Saffran, Newport, & Aslin,
1996). A t-test comparing the errors of these two types of partwords was not signiﬁcant:
t(l9)=1.86,p=.07. The mean number of errors for initial partwords was 6.7, versus 5.1

for ﬁnal partwords.

Question 3: Object Recognition and Identification

APWC Recognition Task

Results of the optimal English testing on the APWC revealed a mean total word
count of M=28, SD=10.8. To assess for participants’ accuracy in counting the correct

words, the absolute difference (where a difference of 0 indicates a perfect score) was

57

calculated between the participant’s counted score and the correct number of words
presented during each of the six different speech streams. These scores were then
summed. The mean absolute difference from the correct count was M=12, SD=7.6. The
English test was also compared against both pre- and posttests. These comparisons used
a Bonferonni-adjusted signiﬁcance level of p=.025. Results demonstrated signiﬁcantly
better performance on the English test as compared to both pretest: t(l 9)=5.03, p<.0001,
d=1.89, and posttest: t(l9)=3.64, p=.002, d=1.10. These results indicate that while
participants performed signiﬁcantly better on the English testing, the greatest effect size
was between the English and pretest with a percent nonoverlap of 79.4%. The percent
nonoverlap of the posttest and English test was 58.9%. The percent nonoverlap is the
area that is not covered by both distributions combined.

Experimental results from the APWC Recognition Task demonstrated improved
accuracy and decreased variability between subjects when comparing the pretest to the
posttest. For the pretest, participants’ mean number of total counted words was 51 (total
number of words actually presented during the test was 36) with a standard deviation of
53.8. This is compared to their more accurate and less variable posttest score (M=39,
5%306). The mean absolute difference on the pretest was higher (M=44, SD=37.6)
than the posttest (M=24, SD=19.7), indicating increased accuracy after exposure for the
posttest. Ranges of the absolute difference also decreased with exposure (pretest=l37;
posttest=83; English=26).

The group data for the APWC, particularly on the pretest scores, had several
outliers. While the number of words presented during the test was 36, six participants

counted more than 36 words, three of whom counted over 100 words. This lack of

58

normality did not allow pre- and posttest absolute difference scores to be compared using
a t-test. Therefore, the data was transformed through logarithms (Irnan, 1994). This
allowed the data to ﬁrlﬁll the normality assumption. The log of the pretest mean absolute
difference (from the correct number of words) was compared to the log of the posttest
mean absolute difference, demonstrating a signiﬁcant difference between tests with a
large effect size: t(l9)=-3.00, p=.007, d=.88. This indicates more accurate performance

on the posttest than on the pretest.

Table 2. APWC means and standard deviations (in parentheses)

 

 

 

 

 

APWC Pretest Posttest English
# Counted 51 (53.8) 39 (30.6) 28 (10.8)
Absolute Difference from # Correct 44 (37.6) 24 (19.7) 12 (7.6)
Log of Absolute Difference 1.5 (0.3) 1.3 (0.3) 1.0 (0.3)

 

SOST Identiﬁcation Task

The optimal English scores were higher than both experimental measures with a
total number counted of M=81, SD: 16.6; correct counted of M=74, SD=23.2; and
percentage correct of M=88.3%, SD=16.3%. The differences for the pre- and posttests
with the English test on the total number of correct words identiﬁed score were also
compared. These comparisons used a Bonferonni-adjusted signiﬁcance level of p=.025.
For the pretest, a signiﬁcant difference was found when compared to participants’
English performance: t(l9)=-10.18, p<.0001, d=2.96. This was also true for the posttest
comparison, t(l9)=-9.42, p<.0001, d=2.56. Comparison of the percent correct was also
signiﬁcantly different for the pretest, t(l9)=-9.60, p<.0001, d=2.85, and posttest, t(l9)=-

7.45, p<.0001, d=2.36.

59

Analysis of the experimental SOST Identiﬁcation Task data demonstrated
increased performance on the posttest as compared to the pretest. The mean total number
of words identiﬁed increased from 46 in the pretest to 59 in the posttest. Of those words
identiﬁed, a mean of only 19 words were correct in the pretest versus 26 in the posttest.
This measure was viewed as an indicator of participant accuracy in completing the task.
A matched-pairs t-test between these pre and post measures of the total number of
correctly identiﬁed words revealed a signiﬁcant difference with a medium effect size:
t(l9)=-2.39, p=.028, d=.52. The percent correct was also calculated for pre- and posttest
scores by taking the number correct over the number counted. This signiﬁcant difference
was still upheld when English test scores were used as a control factor to account for
individual subject differences in general SOST test performance, t(l9)=-2. l 8, p=.04,
d=.16.

The mean percent correct for the pretest was 37.3%; whereas, it was 45.6% in the
posttest. This measure was viewed as a possible indicator of participant precision in
completing the task. A greater dissociation (and lower percentage) between the number
identiﬁed and the correct number identiﬁed could be a measure of intra-subj ect
reliability. A matched-pairs t-test of these scores was also conducted, but demonstrated
no signiﬁcant difference between the scores: t(l9)=-1 .43, p=.17. However, there was a
trend toward increased scores for the posttest. No signiﬁcance on this measure was

found when using the English tests scores as a control factor, t(l9)=-1.59, p=.13.

Table 3. SOST means and standard deviations (in parentheses)

 

 

 

 

 

SOST Pretest Posttest English
Total Identiﬁed 46 (23.4) 59 (20.5) 81 (16.6)
Correct Identiﬁed 19 (12.8) 26 (13.6) 74 (23.2)

 

60

 

l%CorrectofTotal(%) T374209) 46 (22.0) 88 (16.3) |

Question 4: Changes in Performance due to Enhanced Perception

Perceptual changes were indicated in the above sections by the improved
performance on posttests. Also as outlined above, performance scores on the English
tests indicated higher task accuracy, as was expected from an optimal performance
measure. The percent of nonoverlap between the experimental and English samples
decreased between pre- and posttests: 79.4% pre- to 58.9% posttest on the APWC; 92.8%
pre- to 89.3% posttest on the SOST. The difference was most reduced for the APWC.
As predicted, variability indexed by standard deviations and ranges, also decreased for
the APWC from pretest (SD=37.6, range=l37) to posttest(S%19.7, range=83). These

results are examined in more detail in the discussion of this paper.

Table 4. Signiﬁcance probabilities and effect sizes for test comparisons

 

 

 

 

 

Test Measure Pretest vs. Posttest Pretest vs. English Posttest vs. English
APWC log of
absolute difference p=.007, d=.88 p<.0001, d=1.89 p=.002, d=1.10
SOST # correct p=.028, d=.52 p<.0001, d=2.96 p<.0001, d=2.56

 

Question 5: Manipulability of Auditory Object Memory Traces

The examination of task hierarchical difﬁculty from the predicted easier 21FC
discrimination task, to more difﬁcult the APWC recognition task, to the most difﬁcult
SOST identiﬁcation task was not directly testable through a statistical model. However,
examination of effect sizes provides some indication for this hierarchy of processing. For
the discrimination task, the large effect size was d=1.08 with a percent nonoverlap of

58.9% for the difference between participant performance and chance. This size was

61

reduced for the recognition task, which had an effect size of d=.88 and percent
nonoverlap of 51.6% for the difference between pre- and posttests. The medium effect
size for the SOST was the least at d=.52 and percent nonoverlap of 33.0% for the pre-

and posttest difference.

Table 5. Effect sizes changing according to task difﬁculty
21FC APWC SOST
d value 1.08 .88 .52

 

 

 

 

 

 

 

 

62

CHAPTER IV

Discussion

The segmentation problem, within the context of novel language acquisition,
arises from the quasi-continuous acoustic structure of speech. All listeners must learn to
solve this acoustic problem in order to learn and comprehend spoken language. The
acoustic problem to speech segmentation is twofold. First, how are word boundaries
identiﬁed from an often unbroken acoustic stream? Second, how are the endless acoustic
features of the speech stream combined into word-forms? The work of researchers such
as Saffran, Jusczyk, Aslin, Newport, Johnson, among others, has been focused on the ﬁrst
question of how infant and adult listeners learn to identify word boundaries. This project
replicated their methodology of using an ofﬂine discrimination task to conﬁrm the
learning of word boundaries, and expanded upon it by instituting more rigorous acoustic
stimulus controls and using online tasks requiring hi gher-level processing. Research
cited in the electrophysiological literature has focused on this second question:
integrating multiple features into one object. These studies have looked at primarily early
stages of auditory processing with tone patterns. This project attempted to expand upon
these ﬁndings by using this theory of auditory objects (Hayes & Clark, 1970; Kubvoy &
Van Valkenburg, 2001) to explain a mechanism by which statistical structure could
integrate individual aspects of speech (syllables) into a preliminary auditory object
(prelexical word-fonn). It is the position of this paper that an early sensory memory store

built up by statistical regularities allows for the detection of word boundaries; thus,

63

segmenting speech. Novel individual acoustic features are later integrated into perceptual
objects (at least in part) through these statistical regularities, and allow for the recognition
of uniﬁed objects within the continuous speech steam. These objects may later be
entered into hi gher-order memory to establish interaction with cognitive-linguistic
processing and explicit object identiﬁcation. Unarguably, cues other than simple
transitional probabilities also facilitate object formation and segmentation. However, it is
the early nature, saliency, and ability of these cues to work in isolation from other cues,
which make transitional probabilities particularly interesting and signiﬁcant. Through a
series of neuropsychological protocols investigating the recognition and identiﬁcation of
word-forms, this study investigated statistical auditory object formation.

It was these two questions of how word boundaries are identiﬁed and how
acoustic features are combined into word-forms that guided this project. The task
complexity hierarchy of discrimination (21FC), recognition (APWC), and identiﬁcation
(SOST), allowed for the unique investigation of these two questions in both online and
ofﬂine tasks. The sections that follow will discuss the results and implications this study
has for answering these two questions; as well as, address other relevant issues such as

domain-generality and implications for therapeutic intervention.

Identiﬁcation of Word Boundaries

Several techniques for identifying word boundaries were reviewed in the
introduction to this paper and have consistently been supported in the literature. Most
segmentation techniques that have been proposed are acoustic cues inherent in the speech

stream. Such characteristics are metrical stress, allophonic variation, and phonotactics.

64

However, most of these cues are either mediated by (in the case of metrical stress), or are
more sophisticated levels of statistical structure (such as phonotactics and allophonic
variation). It is statistical regularities that seem to be an essential aspect of perceiving
complex scenes (Kersten & Yuille, 2003) and may be an innate process (Canﬁeld &
Haith, 1991; Kirkham, et al., 2001). Saffran and colleagues have presented transitional
probabilities as one early statistical cue to identifying word boundaries (Aslin, et al.,
1998; Saffran, Aslin, et al., 1996; Saffran, Newport, et al., 1996). To investigate this
question, this paper replicated the discrimination task of Saffran, Newport, et al. (1996)

with more highly controlled stimuli.

Question I: Segmentation of Complex Patterns via Statistical Regularities

The overarching purpose of this study was to test the use of statistical regularities
in the analysis of complex stimulus patterns, namely speech syllable distributions, and to
segment those patterns within a continuous auditory stream that offered no acoustic
breaks.

Humans seem to be able to identify word boundaries on the bases of several
different statistical parameters. Sensitivity to these conditions allows individuals to
detect what stimulus occurrences are probable, and to discriminate those events from
ones that occur rather infrequently. The detection of a word boundary reﬂects sensitivity
to an infrequent event (two syllables which do not occur frequently together or predict
each other).

By reviewing the present study comprehensively, the results indicate successful

learning of the two artiﬁcial languages, with no difference in performance between the

65

languages (as tested by the ZIP C). This study’s results indicate that adults are sensitive to
statistical information. They can use this information to identify word boundaries (21FC
condition) and auditory objects (APWC and SOST conditions).

The investigation of Question 1 draws two primary conclusions: 1) Online speech
segmentation can be measured by the neuropsychological protocols developed in this
study, and 2) statistical regularities are sufﬁcient for initial parsing of the speech stream,

identiﬁcation of word boundaries, and formation of auditory object word-forms.

Question 2: Replication of Discrimination Studies

This second question was designed to conﬁrm participant ability to discriminate
segmented words from foils so that conclusions from the neuropsychological protocols
could be drawn. Conﬁrmation of participant abilities was addressed through the
replication of previous stimulus discrimination methodology. These studies are based
upon the underlying assumption that segmented words from the continuous speech
stream will be identiﬁed at above-chance levels as compared to nonsegrnented words.
This was tested in an ofﬂine presentation of words versus nonwords/partwords. This test
examines whether words from the artiﬁcial language were discriminatively selected from
syllable strings that spanned word boundaries.

Discrimination is the detection of a difference between stimuli (Bumham, et al.,
1987). This was the process that was tested during the 2IFC task. An ability to select the
appropriate stimuli above chance reﬂects the participant’s learning of the artiﬁcial
language on a very early, detection level. This task does not require the participant to be

able to explicitly identify the words that were a part of the language’s lexicon, or

66

recognize these words within a continuous speech sample. It is instead, the detection of
an infrequent, deviant event. Therefore, discrimination success demonstrates sensitivity
to the statistical structure of the language and detection of word boundaries. Analyses of
the ZIP C indicated a signiﬁcant selection of the word over the partword stimuli. These
propitious results suggest that participants were able to identify word boundaries above
chance following statistical exposure.

However, detection of a foil’s word boundary does not guarantee recognition of
the word as one perceptual object. Fiser and Aslin (2002) suggested that the nature of the
ZIP C is such that “it is unclear whether participants implicitly extracted triplets [objects]
during the familiarization phase, or whether they simply became sensitive to the pairwise
statistical relations present in the stream...” (p. 464). They continued by stating that
“. . .pairwise shape information present during the familiarization phase is completely
sufﬁcient to discriminate between familiar and novel triplets” (p. 465). Thus, successful
completion of the 21F C indicates sensitivity to statistical structure and to the
identiﬁcation of word boundaries. However, it is not able to determine if the three
syllables forming a word are perceived as one uniﬁed object, or if participants are merely
recognizing the presence of an infrequent pairwise relationship (a word boundary) in the
partword. It is unclear whether participants were selecting the word or selecting against
the presence of a word boundary in the partword. To investigate this issue, an analysis of
initial and ﬁnal partwords was conducted.

Previous research in visual (Fiser & Aslin, 2002) and auditory (Saffran, Newport,
& Aslin, 1996) domains has indicated a discrimination asymmetry between initial and

ﬁnal partwords. These partwords had either an X-1-2 or a 2-3 -X structure (where X is an

67

“out-of—order” syllable or visual symbol). The partword X-1-2 has a joint probability of
the ﬁrst shape pair which differs from the exposed word; whereas, 2-3-X maintains this
same joint probability of the initial pair. Subjects consistently and incorrectly chose
partwords that contained the ﬁnal syllables of words more often than partwords that
contained initial syllables. Thus, both studies found above-chance performance in
discriminating X-1-2 sequences, but not 2-3-X sequences. Fiser and Aslin (2002) and
Safﬁ'an, Newport, and Aslin (1996) interpreted these results to suggest that subjects focus
on the ends of sequences.

The results of the present study did not support previous results. The analysis of
initial versus ﬁnal partword selection errors indicated no signiﬁcant difference. This
suggests the possibility that participants completed the 2IFC task in this study differently
than in previous studies. To perform successﬁrlly on the 2IFC, a participant only needs
to be able to recognize pairwise relations to detect the word boundary (Fiser & Aslin,
2002). The previous studies showed that participants failed to detect the word boundary
in the 2-3-X condition, possibly because of their reliance on pairwise relations. However,
subjects in the present study accurately discriminated words from both 2-3-X and X-1-2
partwords in the 21F C condition (as analyzed by the initial and ﬁnal partword t-test).

This suggests that subjects were not simply using pairwise relations to discriminate words
from foils, but were possibly recognizing the words as uniﬁed objects. This conclusion
was tested using the APWC and SOST protocols which examined object recognition and
identiﬁcation (see Question 3). What made participants in this study perform differently
from earlier studies? It is possible that consciously recognizing and identifying words in

previous conditions (namely, the APWC and SOST) trained participants to also complete

68

the ZIP C in an equivalent manner. Thus, a change in actual perception may not have
occurred, but rather a change in discrimination strategy occurred from looking for word
boundaries to word-forms. Performance ability measured in percent correct does not
seem to vary much from previous studies: 65% in Saffran, Aslin, et a1. (1996) and 63% in
the present study. This hypothesis is supported in this study by no signiﬁcant difference
between 2-3-X and X-1-2 partwords. A simple discrimination strategy of selecting
against weak probabilities would support differential performance on these two types of
part words (as found by F iser & Aslin, 2002; Saffran, Newport et al., 1996). A strategy
of selecting word-forms predicts no difference between partword types, and was
supported in the current study. This segmentation strategy was investigated in this study
by testing higher-order abilities of recognition and identiﬁcation (see next section).

In conclusion, the examination of Question 2 suggests that word boundaries can
be detected ﬁom continuous statistical presentations and discriminated ofﬂine. Also,
investigation of partword errors suggests that discrimination success may occur due to
word boundary detection (suggested in previous studies) or auditory object detection (as

indicated in this study).

Integrating Acoustic Features into Auditory Objects: Creating Word-Forms

The second aspect of the acoustic problem to speech segmentation presents a
slightly different question. It is one that has not received much attention with speech
stimuli in the segmentation literature. How are the seemingly endless acoustic features of
the speech stream combined into word-forms? It is from an adequate understanding of

this second question, that the ﬁrst question can also be more fully understood. Once

69

word-forms are created and lexicalized, additional information is available to aid in
speech segmentation, such as lexical activation (McQueen, 1999).

The grouping of auditory features is important in forming coherent speech
percepts (Barker & Cooke, 1999). Hayes and Clark (1970) proposed that the word was
one type of auditory object. It is a short cut for reducing the amount of stimuli that need
to be processed and considered by the brain; thereby, improving perceptual performance.
The perception and formation of auditory objects is essential for the development of
adequate speech perception and language acquisition. This study examined how the early
formation of these auditory objects within a novel speech environment are recognized

and identiﬁed online during perception.

Question 3: Object Recognition and Identiﬁcation

This third question investigated participants’ online ability to recognize and
identify patterns. The last question addressed used a discrimination task that only
examined word boundary detection. This question expanded upon that task by examining
exactly how individual syllables might be combined into one object. The protocols used

to investigate question three measured recognition and identiﬁcation of these objects.

The Recognition Task

Recognition is the implicit knowledge that some present stimulus has been
encountered before (Galotti, 1999). It requires discriminating the stimulus from other
information and implicitly knowing that it is a reoccurring event. Access to the actual

symbolic content of the stimulus, or any higher-order structure that might be present is

70

not required. For example, recognition of the word “dog” requires the perceiver to
segment “dog” from any other information and to access the prelinguistic knowledge that
the sound structure of “dog” has been heard before. However, recognition does not
require explicit access to higher-order structure of the word. Recognition is the implicit
knowledge of a recurring sound structure.

Results from the APWC recognition task indicated increased accuracy in counting
words following the statistical exposure. The APWC examines the beginning stage of the
structural building of perceptual objects. During this task, the participant did not
deﬁnitively state what syllables formed the word they were counting. However, in order
to obtain an accurate count, they needed an implicit recognition of the auditory object
structure. This is a prelexical mental representation of the object form. Clearly,
participants had no such representation in the pretest (as indicated by posttest and English
comparisons). However, the improved accuracy of the posttest APWC, which more
closely approximated optimal performance, indicates that early formation of object
structures in memory. In addition, the predicted trend for decreased variability between
subjects was noted from the pretest to the posttest. This indicates a greater precision for

between-subj ect performance.

The Identiﬁcation Task

Identiﬁcation entails recognizing the abstract representation of a perceptual object
(Bumham, et al., 1987). This level of processing goes beyond recognition. It requires
the perceiver to access the mental representation of the stimulus, either by accessing

semantic meaning, or by explicitly accessing the higher-order stimulus structure. The

71

word “dog” needs to be identiﬁed as a word-form, not just as a recurrent sound structure.
Identiﬁcation also implies an ability to explicitly label the stimulus event.

Results of the SOST identiﬁcation task again indicate improved posttest
performance. The SOST expands upon the APWC by requiring the participant to
explicitly identify the word-form. While this word-form has no lexical meaning, it must
have a clearly formed, holistic, mental representation of the auditory object such that it
can be consciously retrieved from memory. Therefore, participants likely had an early
cognitive representation. Such a representation was even more important considering
that it needed to be phonologically synthesized in memory.

The smaller difference between pre— and posttests on the SOST may also have
been due to the APWC serving as an initial exposure period (as the pretest APWC was
presented before the SOST pretest). This seems possible as two subjects achieved an
average of 82% correct on the SOST pretest; which was not likely due to chance alone.
This initial exposure may have reduced the difference between the pre- and posttests.
However, despite this possible reduction, participants on average still performed
signiﬁcantly better on the posttest than on the pretest when measuring the total number of
correct words counted. Change in perceptual performance for infants has been
demonstrated to occur in periods of time that were shorter than the 3-minute APWC test
stream used in this experiment (Saffran, Aslin, et al., 1996). As statistical learning is
proposed to be a very early, initial parser, it is possible that some learning and perceptual
improvement occurred before the presentation of the SOST pretest.

Also worthy of discussion is the difference in signiﬁcant ﬁndings between the

analysis of the total number of correct and the percent correct. The total number correct

72

was signiﬁcantly higher posttest, while the percent correct showed a clear trend toward
improvement. However, it also appears that these two calculations may be measuring
two different, though related, phenomena.

The signiﬁcant increase in number correct suggests that participants entered
mental representations of the auditory objects into a memory form that is cognitively
accessible. The ﬁndings of percent correct indicate that while participants increased in
number correct, they also increased more in their total number of responses. This is
related to the increase in between-subj ect variability in the posttest, which is a measure of
how precise the participants were as a whole. Thus, while accuracy improved for the
posttest, precision did not. This was not a factor on the English test, which had
signiﬁcantly better accuracy and precision. This ﬁnding may indicate the very early,
tenuous position the mental representations of these auditory objects held in memory.
Perhaps information other than a uniﬁed acoustic structure needs to be present to ﬁrrther
improve performance and implant auditory objects in memory to improve both accuracy
and precision. This information could be lexical meaning, as this was present (although
to a controlled minimum degree) in the testing of low-frequency English words.

The results of Question 3 suggest that recognition is the early structure building of
auditory objects and occurs implicitly during online presentations. Also, identiﬁcation
implicates forming a higher-order mental representation of the auditory object that is
accessible to cognitive processing. Statistical probabilities are sufﬁcient in forming these
representations, though more exposure or information may be required for native-like
processing. Also, performance improved in accuracy on the SOST, while both accuracy

and precision increased for the APWC.

73

Question 4: Changes in Performance due to Enhanced Perception

An enhanced ability to perceive perceptual units was believed to be implicated in
performance changes that occurred over the time course of the experimental session.
This change in abilities was assessed through the use of a pretest-posttest experimental
design. While the English conditions were designed as a measure of validity to
determine optimal performance on the novel neuropsychological protocols used in this
study, they were also available to serve as a comparison to native language performance.
Therefore, this study also compared perceptual changes to the optimal (English)
performance level, which would be expected of an adult native language speaker. The
results showed that participants performed signiﬁcantly better on both recognition and
identiﬁcation tasks on the posttest measure; as well as, signiﬁcantly better on the optimal
performance tests. On the recognition task (APWC), participants performed four times
less accurately on the pretest measure as compared to their optimal performance. They
were only two times less accurate on the posttest. Thus, the posttest had a smaller
percent nonoverlap with the optimal measure than the pretest. (This would be expected
from more native-like processing.) While posttest identiﬁcation (SOST) did signiﬁcantly
improve, participants maintained a similar percent nonoverlap with the optimal test from
pretest to posttest. This was perhaps due to the increased difﬁculty of the test and lack of
lexical association.

Overall, these results suggest greater proﬁciency for recognition and identiﬁcation
tasks following statistical exposure to the language. However, participants did not reach

a level of native language processing. Greater acquisition of recognition than

74

identiﬁcation abilities was demonstrated by a closer approximation of optimal
performance during the APWC.

In summary, the results of Question 4 demonstrate perceptual changes can occur
in the way humans perceive auditory information. Initial perception of auditory features
may, during reﬁnement by statistical regularities, become holistically perceived as
auditory objects. The results also suggest that with statistical exposure, recognition
ability may more closely approximate optimal performance, while identiﬁcation ability

may require additional information.

Question 5: Manipulability of Auditory Object Memory Traces

Statistical learning of auditory objects may facilitate later cognitive processing.
Fritch, Large, and Pisoni (2000) presented adults with nonwords that were created
according to the probabilities of the onset and rime constituents of syllables contained in
an English dictionary. Subjects rated the wordlikeness of the nonwords containing high
probability constituents higher than low probability nonwords. Subjects had better
recognition memory performance for the high probability constituent nonwords. These
results suggest that adults maintain the probability of linguistic segments within memory
and are able to recognize novel combinations of these segments based upon their
probabilistic nature. Therefore, statistical information is not accessed only during early
sensory processing, but is stored to facilitate later processing of novel sequences.

In the present study, a memory trace to the probabilistic nature of the artiﬁcial
language was created during exposure. An attempt was made to determine whether these

memory traces could be used to facilitate ﬁrrther processing, as Fritch, Large, and Pisoni

75

(2000) demonstrated could be done in a recognition memory task. This study examined a
hierarchy of processing difﬁculty designed to determine the ﬂexibility of these auditory
object memory traces. It also investigated whether these memory traces could integrate
temporally separated acoustic information into the learned auditory object through a

phonological synthesis task.

Building Higher-Level Processing: From Discrimination to Identiﬁcation

The manipulability of memory traces for the learned auditory objects was
examined through the use of higher-order processing tasks rather than the widely used
discrimination task. The hierarchy of perceptual difﬁculty between tasks can be inferred
from participant accuracy. (Direct comparison of the conditions was not possible due to
the highly different measures and different theoretical assumptions.) Comparison of
signiﬁcance probabilities between tasks indicates less likelihood that participants’
behavior was the result of chance for the discrimination task. The possibility of chance
behavior increases for the recognition task and most greatly for the identiﬁcation task
(although participants still increased performance signiﬁcantly). Additionally, effect
sizes were largest for discrimination and smallest for identiﬁcation, supporting this
hierarchy.

Levels of processing were proposed by Craik and Lockhart (1972) and applied to
complex auditory processing by Grifﬁths (2003). The hierarchy of processing examined
through the protocols of the current study occurs in the following way. In the
discrimination task, participants did not need to have any preliminary form of a

perceptual object. They must have only been able to detect statistically infrequent

76

syllable pairs at word boundaries. However, this simple sensitivity to weak transitional
probabilities was not enough to accurately complete the recognition task. Participants
must have formed a higher level of representation based upon the statistical structure of
the language. This representation is at the preliminary structure building of the auditory
object where the explicit word-fonn does not have to be stated, but where an implicit
detection of its holistic structure occurs. It is not until identiﬁcation occurs that an
explicit word-fonn (that is stored as a cognitively penetrable memory form) needs to be
accessed. An implicit detection of the preliminary structure of the auditory object would
have been insufﬁcient to complete the identiﬁcation task. Participants were required to
explicitly label the exact structure of the word. This was different from implicitly
detecting its occurrence (as in the recognition task). A ﬁnal level of representation, not

examined by this study, would be the retrieval of the word-fonn stored in the lexicon.

Evidence of Phonological Synthesis

The ability to phonologically synthesize auditory objects provides another
argument for a higher level of representation than that offered by early sensory
discrimination processes. To regenerate a word-fonn from temporally separated
segments, an explicit mental representation of the word-fonn is required. As this is a
phonological task, it might be argued that this word-form is now stored as a prelexical
form in memory. Participants demonstrated their ability to complete the phonological
synthesis task of the SOST after exposure to the artiﬁcial language. This suggests that
statistical regularities are sufficient for forming and entering auditory objects into

memory as a hi gher-order representation.

77

In conclusion, Question 5 suggests that processing of the speech stream may
occur on a hierarchy of difﬁculty building from discrimination to recognition to
identiﬁcation. Each of these levels may entail its own level of representation, may
require different involvements of memory, and may to various degrees be accessible to
explicit cognitive processing. Auditory objects are susceptible to higher levels of
processing, such as phonological synthesis, suggesting the possibility of maintaining a

prelexical representation in memory, accessible by cognitive processing.

Summary of the Research Questions

The results derived from the research questions of this project conﬁrmed the
presence of a statistical learning mechanism. This project replicated previous studies by
verifying the ability to discriminate word-forms from partwords based solely upon
statistical exposure. It also expanded upon these previous studies by examining the
strength of statistical learning in higher-level processing during recognition and
identiﬁcation tasks. These higher levels of processing required more than a simple
detection of low probability word boundaries. They required the formation of auditory
objects. The results of the recognition (APWC) and identiﬁcation (SOST) tasks suggest
that statistical learning mechanisms are able to combine multiple features (syllables) into
one perceptual object, a word-form, which is available to cognitive processing. These
results support a processing hierarchy and provide a mechanism which is capable of
creating abstract representations of word-fonn structure to which meaning can later be

applied. The sections that follow will further discuss this statistical learning mechanism,

78

apply these results to language learning and therapeutic settings, address possible

limitations of this study, and suggest further avenues of research.

Domain-Generality of Statistical Learning and Perceptual Objects

An investigation of the domain-generality of statistical learning can further help
to explain the process by which perceptual objects are formed. Domain- general
processes may work to form and/or support domain-speciﬁc modules. Elrnan and Bates
(1997) discuss statistical learning as an innate default assumption to evaluating all
stimulus input (regardless of domain), which evolved for other reasons, but is now
recruited for language learning. While Cosmides (1989) supports a more specialized
process approach, she states that an organism’s behavior will be random unless it has a
reliable and efﬁcient means of extracting information from the environment and a well-
deﬁned system of rules to use the information. She continues by saying that specialized
leaming mechanisms organize experience into meaningful units which focus attention,
organize perception and initiate procedural knowledge that lead to domain-speciﬁc
processing (Cosmides, 1989). The current paper suggests that statistical leaming may be
such a leanring mechanism that works to shape and implement future domain-speciﬁc
processing modules. Probabilities may be particularly important for initial parsing and
ordering of stimulus events; thereby, enabling the best mode of processing the
information to be determined and assigned.

Statistical sensitivities may work like a person ﬁling a stack of assorted forms in a
law ofﬁce. The person recognizes the pattern of the letterhead, the shape of the text

body, or the form title, but does not process the actual content of the form. Processing

79

occurs when the form reaches the ofﬁce of the speciﬁc professional who was trained to
work with that form. That specialized person is doing the processing work of a speciﬁc
domain module. The person received speciﬁc information that was identiﬁed and
assigned on the bases of statistical pattern sorting. Therefore, this paper does not claim
that it statistical learning is the only information available to segment speech for the
acquisition of language. Other domain-speciﬁc processes are also likely to facilitate and
carry on this work, especially when a higher-order system is established to speciﬁcally
recognize and identify perceptual objects after statistical learning has shaped those
modules.

This study is unable to directly test domain-generality, as it only studied one
domain of speech perception. However, it is relevant to this discussion within the context
of previous research, as statistical learning appears to be a ubiquitous process.
Investigating the domain-generality of statistical learning requires the examination of
processes isomorphic to those examined in the speech domain by this paper. Thus,
investigating musical and visual domains can help explain the process by which the brain
creates representations of word-forms via statistical learning.

Music is a different form of the same ability that underlies speech perception; it is
the ability to organize complex stimuli into temporally ordered sequences (Jourdain,
1997). Electrophysiological results suggest that music and language appear to have
similar electrophysiological correlates for auditory and temporal structure (the level of
the auditory object), while processing semantic meaning in context appears to be
language-speciﬁc (Besson & Schon, 2003). Tervaniemi (2003) discussed ERP and MEG

evidence that may implicate Broca’s Area in processing chord cadences (an auditory

8O

object), but differential processing for individual musical sounds. This evidence suggests
that there may be a common way the brain processes auditory objects. Saffran (2003b)
continued this discussion by suggesting that the learning and memory for music and
language occur without instruction or reinforcement. Humans are able to implicitly learn
and remember structured environmental information.

These studies implicate common processing of auditory objects in music and
speech domains and in certain cases, speciﬁc brain locations. Auditory objects are not
speciﬁc to speech, but are a domain-generalized process of perception that can be derived
implicitly through the natural probabilistic structure the environment. Indeed, tonal
patterns have been perceived as whole units based upon the statistical presentation of
their structure (Brattico, et al., 2002; Saffran, Johnson, Aslin, & Newport, 1999; Vaz
Pato, et al., 2002).

Vision is also processed according to statistics. Kersten and Yuille (2003) wrote
that “statistical regularities in natural images and scene properties are essential for taming
the complexity and ambiguity of image interpretation” (p. 150). They cite studies of
homogeneous textures and object boundaries, as well as, scene and object recognition
studies where statistical patterns help to defrne, constrain, and/or explain the perception
of visual objects. Fiser and Aslin (2002) furthered the support for the domain-generality
of statistical learning by adapting Saffran, Aslin, et al.’s (1996) study to a visual shape
paradigm. Sequences of three visual shapes were presented during a 6-minute movie.
Their results replicate past ﬁndings in that the three shape sequences were successﬁrlly
discriminated from novel sequences and part sequences in an ofﬂine two-interval forced-

choice task.

81

Evidence from visual studies has provided neurophysiological explanations for
how perceptual objects facilitate processing. It appears that the formation of a uniﬁed
object creates a better perceptual ﬁt (Kersten & Yuille, 2003) and reduces brain activity;
thereby, making detection of novel stimuli easier (Murray, et al., 2002). This may be
similar to the neural mechanism behind the MMN, which indexes increased activity

following novel stimuli that break an auditory pattern.

Applications of the Current Study

In addition to investigating the theoretical aims of this research, this study also
has applications to other areas. The theoretical foundations proposed in this study
support the importance of statistical and probabilistic structure in learning theory. This
theoretical base adds to the understanding of very early, basic language acquisition
mechanisms. It provides a rationale for methods of teaching second languages and for
various therapeutic interventions for language-delay and neurogenic clinical populations.

At the most basic level, this theory suggests that in order to learn language, we
need to hear language—repetitively. We must be exposed to enough recurrences of a
particular token to form a mental representation of it; thus, perceiving the probabilistic
structure and deriving a perceptual object. Therefore, experience guides perception,
which is supported in neural plasticity research (Sharma, Angelucci, and Sur, 2000).

This paper stresses that statistical learning is only one mechanism that the child
employs during language development. As discussed by Hollich, Hirsh-Pasek, and
Golinkoff (2000) and J usczyk (1997), many other cues certainly play a role in language

development, including higher-level learning structures such as that employed during

82

social interaction. However, the theoretical foundation of this study suggests that
statistics is a highly important cue that infants and adults can tap into. It is available
across domains (Kirkham, et a1, 2002) and may even be important for integrating cross-
modal information, such as pairing word-forms to real world visual objects for lexical

development (see Roy & Pentland, 2002).

Second Language Instruction

The results of this study, along with the previous research on statistical
regularities suggest that humans are naturally sensitive to patterns; we perceive patterns
even when none exist (Huettel, et al., 2002). Deriving those patterns from the
environment allows us to understand the structure of language and how those components
interact. Our sensitivities also allow us to identify patterns in context, which is essential
to ﬂuent spoken language comprehension. While learning a second language through a
memorization of isolated vocabulary words is common, an exposure to words in context
forces us to derive the patterns of the language (syllabic and grammatical). This concept
may be similar to a theory proposed in the problem solving literature (the problem here
being the attempt to learn a foreign language). Vollrneyer, Burns, and Holyoak (1996)
reported that when participants are presented with speciﬁc goals, they will learn how to
achieve the goal, but will be unable to generalize to other related tasks. However, when
participants were provided a nonspeciﬁc goal, they worked to discover the operators by
which the problem could be solved. This strategy enabled the participants to generalize
their strategies to solve other similar problems. Translating this literature to learning

language: a speciﬁc goal for the language learner is to memorize the correct translation of

83

vocabulary words; whereas, a nonspeciﬁc goal would be to listen to a speech sample to
derive the gist, or overall meaning. It may be that when learning language, a more
efﬁcacious approach would be to use a nonspeciﬁc goal. This approach is supported by
the present study. Our human capacities enable us to identify statistical patterns online
within a continuous speech stream. Therefore, we have the ability to derive the patterns
of a language, including grammar (Gomez & Gerken, 2001), without explicit instruction
on individual words. This is how infants must learn language, and this study suggests
that the young adult still has the capacity to do this. Statistical learning suggests why
people typically do not ﬂuently leanr languages until they are immersed. Goal-speciﬁc
vocabulary teaching may encourage “translation thinking,” while a nonspeciﬁc goal
immersion approach forces the learner to identify language patterns; thereby, facilitating
understanding of the language structure and allowing for “language thinking.”

In applying this research to the student leanring a foreign language, it is apparent
that enough auditory information is required to form an object for word learning. This is
evident from numerous antidotal reports that people learn a new language better when
completely immersed in the foreign language. Many mechanisms are at play in such a
setting, just as they are at play for the infant. Social and situational cues are likely
important. But also important is the rich and constant exposure to the structural
probabilities of the language. Constant exposure to the statistical structure of the
language allows for the perceptual formation of many different auditory objects, which
can later be identiﬁed and entered into the lexicon when the objects are assigned semantic

meaning through social—pragmatic learning processes.

84

However, the conclusion that adults learn language as efﬁciently as a child cannot
be derived from the results of this study where adults learned to identify prelexical
structures. Research suggests that as language learners mature, their ability to learn
language broadly deteriorates, thus supporting a critical period (Johnson & Newport,
1991). A critical period for musical ability has also been discussed (Pantev, Oostenveld,
Engelien, Ross, Roberts, & Hoke, 1998), possibly suggesting the development of an early
neural organization for complex sound perception based upon early environmental

exposure to speciﬁc sound pattems.

Theraputic Intervention

In addition to second language acquisition, statistical learning has implications
within the therapeutic environment. Breitenstein & Knecht (2003) suggest the use of
massive, repeated interactive exposures; high concurrence of language and corresponding
sensory processes; and intense training frequencies. Creating a rich, stimulating
environment is the essence of most treatment techniques. The Power-Law of Practice
states that the time required to complete a task decreases over the number of trials; thus,
practice is an important component for acquiring a new skill (Lovett, 2002) or
rehabilitating an old skill.

The use of highly frequent, statistical presentations is found in a number of
different treatment techniques. One general category is behavior modiﬁcation, based
upon operant conditioning. Lovaas (197 7) proposed using operant conditioning to
facilitate language development in nonverbal children. Behavior modiﬁcation has also

been used in aphasia treatment (Davis, 2000). Another aphasia treatment technique is

85

programmed stimulation, which relies on stimulation to elicit repeated responses (Davis,
2000). Articulation therapy has a treatment technique called Bombardment, which is
designed to increase the frequency of a target stimulus in the child’s environment
(Nemoy & Davis, 1954). Computer-assisted therapy, such as Fast ForWord (R)
(Scientiﬁc Learning Corporation, 1998), also provides statistical environments for
learning.

While many of these (and other) therapy techniques may be effective in changing
or modifying behavior, possibly in part due to statistical presentations, it is important to
consider other factors when selecting a treatment technique. Behavior modiﬁcation
techniques have been criticized as teaching speciﬁc behaviors that do not generalize
outside of the therapy environment, and ignoring the actual underlying source of the
impairment (Davis, 2000). Treatments other than these cited may be more effective at
producing functional behavior that transfers and generalizes to other situations outside of
the treatment setting. The Cognitive Stimulation approach to aphasia treatment takes
some of this into account by suggesting that therapy tasks should be structured to ensure
60-80% success so that successful cognitive processing is practiced (Davis, 2000).
Constant failure only practices the inappropriate processes that lead to the failure.

Statistical leaming may also have a role to play in assigning meaning to word-
forms during the development of a lexicon. Cross-channel Early Lexical Leanring
(CELL) is a computer model that was designed to ﬁnd and model consistent structure
from multi-sensory information (Roy & Pentland, 2002). CELL creates a lexicon
through segmenting speech and assigning semantic meaning through statistical parings of

speech and visual objects. Therefore, word learning may be a function of identifying

86

 

statistical regularities cross-modally. This suggests that to facilitate development of a
lexicon, words should be consistently paired with real world visual/tactile objects to
which they refer. Intensive statistically cross-modal associations may be essential. This
strategy of word-obj ect pairings is integrated into naturalistic early intervention methods
(Bailey & Wolery, 1992).

Many possibilities exist for the delayed development of language. However, this
study proposes one such possibility: a decreased or impaired ability to process the
statistical structure of language. Therefore, language therapy for such a child should
focus on increasing the salience of other cues to word leanring, or to increasing statistical
input. By exaggerating other cues, the child may be able to tap into unimpaired leaming
mechanisms which may thereby assist in the acquisition of language. By increasing
statistical cues, the child may be able to more easily detect the probabilities inherent in
the language.

It is possible to envision several scenarios of language-delay where statistical
regularities may play a role. The child may be in an impoverished language environment
where he or she is not receiving enough language exposure to identify statistical
structure. This intuitively would suggest that increased language exposure is needed,
which is supported now by the theory of statistical learning. Another situation might be
that the complexity of the language exposure has decreased probabilities to such an
extent that the child is not able to identify structure (which could result ﬁom a highly
complex and variable semantic and syntactic structure in the environment). Reducing the
complexity of the language exposure and increasing perceptual cues, such as provided by

melodic intonations characteristic of motherese, may be appropriate.

87

While these types of therapies are rather intuitive and are already being employed
by many speech-language pathologists, this study provides a theory to justify such
therapy techniques and explains a possible contributor to the underlying causes of the
language delay. While the child’s language delay might be the result of many impaired
learning processes, such as a lack of appropriate social language models, an impoverished
statistical environment may also confound such problems.

This study also examined the transfer of speech sounds stored as a sensor memory
trace to encoded auditory objects. Early on, the phonological aspect of word learning
may be more important than semantics (Hu, 2003). A breakdown in the temporary store
of word-forms may impair construction of phonological representations (Hu, 2003).
Underspeciﬁcation of phonological representations may underlie speciﬁc language
impairment (Maillart, Schelstraete, & Hupet, 2004).

This study suggests, along with Breitenstein & Knecht (2003), that massive,
repeated exposure is important in therapy environments, particularly when the treatment
must be performed at a very low, sensory level of processing. However, this statistical,
therapeutic environment needs to be structured in such a way as to facilitate behavioral
generalization and transfer. It is essential that trained professionals use their clinical
expertise to structure appropriate learning environments. Statistical learning is a general
tool that is likely to always be available for use in therapy (as it may represent a general
brain process). As such, it is a powerful instrument for structuring the environment for
early acquisition of abilities and for cognitive retraining after brain damage. This study
develops an avenue for future research which can apply the theory of statistical learning

to clinical populations and therapies.

88

Study Limitations and Considerations
Syllable Identiﬁcation Pilot Data

A preliminary study of the speech syllables used in this study was conducted to
investigate how participants would perceive and identify the synthesized speech used in
this study. Fifteen participants were presented all 12 syllables across three trials, to

which they were required to transcribe the syllable as they perceived it.

Perceptual Learning of Sygthetic Speech

The synthesized speech syllables had reduced intelligibility due to the elimination
of acoustic cues by controlling for frequency, amplitude, and duration across all syllables.
Previous research has indicated that listeners have an increased difﬁculty for
understanding non-native synthesized speech such as the artiﬁcial language used in this
study (Reynolds, Bond, & F ucci, 1996). Studies suggest that training in perceiving the
synthesized speech is required for decreasing error rates (Conroy, 2003; Francis, 1998).
Also, listeners will change their judgments of synthesized consonants after training,
which may indicate a change in the features where attention is focused (Francis, 1998).
Thus, it may be that participants required a training exposure ﬁrst in order to adequately
perceive the synthesized, syllabic stimuli used in this study. This may be why
participants ﬁrst required the initial English familiarization task to perform adequately on
the English testing portions of this study. However, it may be that the statistical exposure
periods were enough to improve perceptual judgments of the experimental stimuli.

Future studies may wish to train participants ﬁrst on the individual synthesized syllables

89

that will be used by presenting the syllable in natural speech followed by its synthetic
counterpart (as this study did with the English words). Such a procedure should help to
improve identiﬁcation scores of the experimental stimuli and decrease variability

between subjects.

Altering the Probability of the Artiﬁcial Language

This identiﬁcation data suggests the possibility that participants perceived a
language that was different from the one intended. Based upon the most frequent
identiﬁcation results for each syllable, the resulting languages were the following:
Language A: padipu, tibudo, dolatu, daloti; Language B: tudalo, tidola, diputi, budopa.
As there were differences between subjects in the actual transcription they used to
identify certain syllables, subject perception of the languages could have varied from
these modiﬁed languages. As the syllables /do/ and /ti/ now occur twice within the
languages in two different words, participants would have had to make use of an altered
statistical environment. This language was modeled after Saffran, Aslin, and Newport’s
(1996) stimuli and synthesized on a newer version of the text—to-speech synthesizer that
they used. They reported that transitional probabilities within words were p=1 .0 and
between words was p=0.33. In light of this identiﬁcation data, these transitional
probabilities change depending upon the word examined. For example, padipu
(originally pabiku) still maintains the p=1 .0 / p=0.33 probabilities as no syllables within
the word were repeated in another word. However, daloti (originally daropi) repeats the
syllable /ti/. This weakens the internal transitional probability of the ﬁnal syllable pair to

p=0.50 as it occurs 50% of the time in a different word. However, the external

90

transitional probability between words following daloti is also weakened to p=0.167 as
/ti/ is followed by /bu/ (in tibudo) 50% of the time and the initial syllables /pa/, /ti/, and
/do/ the other 50%.

Although the identiﬁcation data altered the perceived languages and the
transitional probabilities, participants still had clear probabilities to learn the languages
based on statistics. Saffran, Newport, and Aslin (1996) demonstrated that participants
can learn a language with weakened transitional probabilities when they used the same
syllables within multiple words, creating internal word probabilities ranging from 0.31 to
1.0. Therefore, the weakened probabilities of the languages used in the present study
were not a likely cause for not leaming the language to a higher degree of proﬁciency.
As an older computer synthesizer created Saffran, Aslin, et al.’s (1996) stimuli, they also
likely had similar identiﬁcation characteristics that were not identiﬁed in a separate
identiﬁcation paradigm as in this pilot study, yet infants were still able to learn the
language during a 2-minute exposure presentation.

Although this pilot data does raise some interesting considerations for the use of
synthetic speech in perceptual speech research, the impact on this study’s outcome and
interpretation seems minimal, if not non-existent. Any effect of these identiﬁed factors
(e.g., variability between subjects on syllabic identiﬁcation, altered transitional
probabilities, etc.) could only have increased variability between subjects and made the
language’s statistical structure harder to perceive. The signiﬁcant differences obtained in
this study suggest that these characteristics of the synthetic stimuli had little effect. This
strengthens support for the incredible ability humans have to recognize statistical

structure, even with possibly ambiguous stimuli such as those used in this study.

91

English Validity Testing

The results on the English control tasks within this study demonstrated that
participants can successfully complete the neuropsychological protocols that were
developed speciﬁcally for this investigation. This suggests that these measures can be
used in future studies to further investigate online segmentation in speech or possibly
other domains. The discrepancy between scores on the posttest and English measures
suggest that a strong competency in perceiving the language is needed. Perhaps
additional exposure, statistical exposure in additional domains, or other information
(statistical, lexical, or otherwise) is required to achieve the same level of proﬁciency as

with English stimuli.

Additional Considerations

The tasks of this study required motivation and attention to complete. Participant
abilities in these areas may have decreased by the end of the session inﬂuencing
performance on the SOST and 21FC posttests.

Possible effects of dropping out several subjects from the analyses of this study
must also be considered. Nittrouer (2001) stated that “children may become
uncooperative precisely because they cannot discriminate the stimuli presented. ...on a
different day they usually become uncooperative with the same or similar stimuli”
(p.1603). Nittrouer continues by saying that “it would be inappropriate to assume that the
dismissed infants would have performed as the infants who were not dismissed” (p. 1603).

It is possible that a similar phenomenon may have occurred with the adult participants

92

who were not able to complete the experimental tasks. Indeed, 3 participants counted no
words from the artiﬁcial language in the pretest of the APWC; as Nittrouer’s infants, they
offered no response during this impossible task. It was impossible because participants
had never been exposed to the words in this language before. How could they be
expected to count them from within a continuous speech stream? One participant was
eliminated ﬁom the analysis because she failed to identify any artiﬁcial words in either
pretest or posttest of the SOST and APWC. She also did not attempt to discriminate
between the words and foils during the 21FC. Was this lack of response due to a
behavioral problem? As she did successfully complete the English testing (identifying 95
out of 100 words correctly on the English SOST, which would have tied for the fourth
highest score had she been included), it is not likely that the task structures themselves
required too difﬁcult of a response. Therefore, it seems possible that she was unable to
discriminate (and therefore also recognize and identify) the words in the artiﬁcial
language; or, simply confused by the task requirements, even after task training.

While the possibility of this study’s results being affected by eliminating subjects
exists, it is unlikely that the results of this study are compromised. Of the 22 subjects
tested who met inclusion criteria, only two were eliminated ﬁ'om analyses. The ﬁrst
subject was excluded due to demonstrated and reported fatigue during testing, the second

subject already discussed failed to complete the testing.

Further Questions and Areas of Research

The neuropsychological protocols developed for this study were an attempt to

behaviorally measure segmentation performance online before and after learning a novel

93

 

language. The development of an electrophysiological paradigm to test the
neurophysiological correlates of statistical learning and the identiﬁcation of auditory
objects online may now be warranted. The MMN may be able to index the change from
perceiving the novel continuous speech stream as random syllables to perceiving it as a
series of auditory objects that are prelinguistic word-forms. A pretest-posttest design to
measure how learning the language inﬂuences the electrophysiology of the brain, as well
as subsequent discrimination paradigms, could be developed to identify
electrophysiological changes which occur when leanring a new language.

Further isomorphic studies of the domain-generality of statistical learning and the
similarity between auditory object formation and perception with visual object formation
and perception may also be warranted. Similarities between speech and music perception
may also help to delineate how statistical regularities help to identify perceptual objects
as a general process that the brain uses in a variety of domains, as well as investigate
where speech, music, and other domains begin to differ in their representation and
processing. Do different domains tap into statistical information differently? Or does it
provide the same information and means to an end for several different brain processes?
Also, neuroimaging studies may wish to examine whether statistical regularities tap into
the same anatomical regions, or if they are a more generalized mechanism of brain
neurophysiology.

Evidence from artiﬁcial grammar studies suggests that the same process may be
involved in higher order domains of grammar and vocabulary (Gomez & Gerken, 1999,

2001). Studying the time course of acquisition of these grammatical processes may help

94

to understand the order of how we learn grammar and provide intervention techniques for
individuals who are not developing grammatical skills within a normal timeframe.

Examinations of the degree of complexity that can be extracted from statistical
structures is also warranted. Newport and Aslin (2004) have recently discovered that
adults are sensitive to non-adj acent patterns. This study has suggested that statistical
regularities can form objects and enter them into a level of representation where they are
susceptible to cognitive-linguistic processing. What exactly is the upper bound of
statistical leaming?

Along with this are investigations of higher order grouping. This study examined
how perceptual objects are formed through combining individual syllabic features into
one unit. Past research on morse code has demonstrated an ability to go from perceiving
only individual letters to perceiving whole words and eventually higher language habits
(Bryan & Harter, 1899). Studies in perceptual grouping for learning artiﬁcial grammars
are now ongoing (Gomez & Gerken, 1999, 2001), and grouping structures have also been
identiﬁed for highly skilled activities, such as chess (Chase & Simon, 1973; Gobet, Lane,
Crocker, Cheng, Jones, Oliver & Pine, 2001; Simon & Gobet, 2000). It is possible that
clauses, sentences, and even discourse scripts become recognized as units based upon the
frequency of the co-occurrence of their individual elements. Therefore, individuals who
do not perceive these scripts, or have difﬁculty forming higher-order objects may beneﬁt
from intensive treatment based upon developing these high probabilistic memory traces.
It may be warranted to investigate possible intervention techniques once the formation of

these higher-order groupings is more clearly deﬁned.

95

This paper presented that statistical learning is a domain-general process that we
do innately and implicitly. It is possible that intense and repetitive exposure to an
interactive therapeutic environment would increase client success. To that end, studies
may be designed to investigate how to tap into these statistical processes within the
intervention environment. As sensitivity to statistical regularities occurs very early on in
life, it is likely that this structure could be used successfully with developmentally
disabled and neurologically impaired individuals. Studies should examine more diverse ’5
populations of race, age, and gender, and their sensitivity to statistical structure. :‘

This study also examined perceptual changes resulting from creating higher-order

representations of auditory objects. Some studies have connected this process to l

 

language learning (Hu, 2003) and implicated resulting structures to language impairment
(Maillart, et al., 2004). Future studies may wish to continue deﬁning how this perceptual
process impacts clinical populations and whether statistical learning can be used to

facilitate the development of more appropriate representations.

Conclusions

The changes in performance and above-chance word discrimination measured in
this study clearly demonstrate the ability to use statistical regularities to segment speech.
Adults are able to use this information to detect word boundaries and form prelexical
auditory word-forms. Participants in this study were able to solve three different
perceptual/cognitive tasks which required three different levels of processing. The
successful completion of these tasks suggests that statistical regularities can be used to:

I) perform preperceptual discrimination tasks of infrequent stimulus detection (weak

96

transitional probabilities), 2) structure early perceptual forms for the implicit recognition
of familiar pattern structures, and 3) enter into memory prelexical, mental
representations of auditory object word-forms which are explicitly available for language
processing in phonological synthesis and identiﬁcation tasks. In short, statistical
structure forms auditory objects, which are the acoustic structure to which meaning is
symbolically assigned. Speech segmentation is not just a simple detection of word
boundaries. Instead, it is an active process where prelexical word-forms are created. In a
naturalistic environment, it is possible to envision how statistical pairings of auditory
objects to visual (or tactile) objects associates these word-forms to meaning and enters
them into the lexicon. I go so far as to predict that statistical leaming is a
neurophysiological phenomenon of neural plasticity: it is how our brains form structural
representations of the random environment in which we ﬁnd ourselves. As such,
statistical regularities inﬂuence domain-general processes, and are an important, innate,
and environmentally adaptive process of perception. They form structural representations
for later cognitive processing. Statistical learning opens an avenue of research to
examine the connection and interaction between perception and cognition, and offers
potential insight into new ways of therapeutically adapting the environment to facilitate
language learning and cognitive retraining in multiple domains.
This study supports conclusions in the following areas:

1) Segmentation of complex patterns via statistical regularities

a. Online speech segmentation can be measured by the neuropsychological protocols

developed in this study.

97

b. Statistical regularities are sufﬁcient for initial parsing of the speech stream,

identiﬁcation of word boundaries, and formation of auditory object word-forms.
2) Replication of discrimination studies

a. Word boundaries can be detected from continuous statistical presentations and
discriminated ofﬂine.

b. Discrimination success may occur due to word boundary detection or auditory
object detection, as indicated by different results on partword selection errors.

3) Object recognition and identiﬁcation

a. Recognition is the early structure building of auditory objects and occurs
implicitly during online presentations.

b. Identiﬁcation implicates forming a higher-order mental representation of the
auditory object that is accessible to cognitive processing. Statistical probabilities
are capable of forming these representations.

4) Changes in performance due to enhanced perception

a. Perceptual changes occur in the way humans perceive auditory information.
Initial perception of auditory features may, during reﬁnement by statistical
regularities, become holistically perceived as auditory objects.

b. With statistical exposure, recognition ability may more closely approximate
optimal performance, while identiﬁcation ability may require additional
information.

5) Manipulability of auditory object memory traces
a. Processing of the speech stream may occur on a hierarchy of difﬁculty building

from discrimination to recognition to identiﬁcation. Each of these levels may

98

entail its own level of representation, may require different involvements of
memory, and may to various degrees be accessible to explicit cognitive
processing.

Auditory objects are susceptible to higher levels of processing, such as
phonological synthesis, suggesting the possibility of maintaining a prelexical

representation in memory, which may be accessed by linguistic processing.

99

 

APPENDICES

100

. V‘- I- I

 

APPENDIX A

Participant Forms and Recruitment

101

 

1"” ,

PARTICIPANT CONSENT FORM
Statistical Speech Segmentation:
A Neuropsychological Investigation of Auditory Object Formation

You are invited to participate in a study that will examine how second language learners
identify words within long speech passages. My name is Dan F ogerty and I am using this
study as the foundation of my master’s thesis in Speech-Language Pathology at Michigan
State University. This study is being advised by Jeff Marler, Ph.D., who is an Assistant
Professor in Audiology & Speech Sciences. 1 hope to leanr more about how people
identify words while listening to someone speak, particularly if this person is speaking a
foreign language. You are eligible to participate in this study if you have normal
language and hearing, and if you are ﬂuent in only one language. You will be one of 20
college students to participate in this study.

If you decide to participate in this study, you will:

1)

2)

3)

4)

5)

You will be asked questions concerning your language and musical experiences,
as well as your medical history. These will require yes/no or open-ended
responses.

Screen your hearing. This will include three measures: a) you will respond every
time you hear a beep-like tone, b) your ear canal will be viewed through otoscopy,
and c) your middle ear functioning will be assessed with tympanometry.

You will be asked to serve in the current study for one 120-minute session (with
short breaks when necessary). During this session, you will receive a hearing
evaluation and be asked to complete a standardized screening of your ability to
manipulate speech sounds. You will also be asked to sit in a sound-treated room
while you wear earphones and listen to continuous artiﬁcial syllables. None of
the sounds you will hear poses any risk of hearing damage. I will ask you to
listen and respond to different speech-like words. I will ask you to identify which
artiﬁcial word sounds the most familiar or to identify those word sounds while
listening to another artiﬁcial speech passage. You will also be asked to complete
these tasks with English speech passages.

This study does not involve risks or harm any greater than those ordinarily
encountered in daily life. You may choose to withdraw from further participation
at any time during the session (and will be paid for your time up to that point).
You should inform the investigator immediately if there is discomfort of any kind
during the course of the experiment. The investigator may periodically request
such information from you during the session.

Potential beneﬁts from participating in this study are to provide a stronger

theoretical understanding of speech perception and language acquisition. If
requested, you may receive a personal written report at the end of the study.

102

 

6) Your privacy will be protected to the maximum extent allowable by law. Any
information that is obtained in connection with this study and that can be
identiﬁed with you will remain conﬁdential and will be disclosed only with your
permission. All forms containing your information will be stored in a locked
cabinet in the Cognitive Auditory Research Lab. Only Dan F ogerty and Jeff
Marler, Ph.D, will have access to this cabinet. Any publication resulting from this
study will identify you only according to code.

You are being asked to participate in a study investigating the learning of speech
segmentation. Your decision to participate will not inﬂuence your future relations with
Michigan State University. You will receive either $15.00 or course credit for the
completion of the experimental session. You may refuse to answer any question during
the experimental session without penalty.

Please initial here if you wish to receive course credit:

 

If you have any questions about this study at any time, please contact the investigator
Dan Fogerty by phone: 394-7412, email: fogertyd@msu.edu, or regular mail: 221 E.
Edgewood Blvd., Apt. F, Lansing, MI 48911; or his advisor Jeff Marler, Ph.D. by phone:
355-7628, fax: (517) 432-2370, email: marler@msu.edu, or regular mail: 378 Comm.
Arts & Sci. Bldg, East Lansing, MI 48824. If you have questions or concerns regarding
your rights as a study participant, or are dissatisﬁed at any time with any aspect of this
study, you may contact — anonymously if you wish — Peter Vasilenko, Ph.D., Chair,
University Committee on Research Involving Human Subjects (UCRIHS) by phone: 355-
2180, fax: (517) 432-4503, email: uchrihs@msu.edu, or regular mail: 202 Olds Hall, East
Lansing, MI 48824.

Should you choose to discontinue participation in this study, you may withdraw at any
time after signing this form.

Thank you for your time in visiting about the project and your consideration to
participate. You will receive a copy of this form to keep for your ﬁles.

I voluntarily agree to participate in this study.

 

 

 

Participant Signature Date Phone #

 

 

Signature of Investigator Date

103

 

 

Statistical Speech Segmentation:
A Neuropsychological Investigation of Auditory Object Formation

RESEARCH PARTICIPANTS NEEDED

For a study in
Audiology & Speech Sciences

Research will be conducted in a session of about 120- minutes.

Listen to artiﬁcial speech...

And get $15.00!

Project Description

 

 

Have you ever listened to someone speak a foreign language and have difﬁculty telling
where the words are? We might say that the person is speaking too fast — all the words
blend together. This study attempts to learn more about how people identify words while
listening to someone speak a foreign language. Participation requires some perceptual
testing, a hearing screening, and listening to artiﬁcial speech. You are eligible to
participate in this study if you have no history of a language, reading, or hearing
impairment and if you are ﬂuent in only English. Also, due to demonstrated differences
in speech processing, you need to be female. Sorry guys!

 

104

 

PHONE SCREEN SUBJECT ID:
Called Date: Screen Results: IN OUT

Interviewer’s Initials:

Hello, this is Dan F ogerty ﬁom the Michigan State University Speech Segmentation study. Is
@erson ’s name) available? I am calling in response to your expression of interest in the study. I
have a few questions that help us to determine if you are eligible for the study. Would you mind
spending 5—1 0 minutes on the phone with me at this time to answer these or is there a better time
I could call back? Your conﬁdentiality concerning this interview will be protected to the
maximum extent allowable by law. If you have any questions along the way, please let me know,
and I will do my best to answer them. You indicate your voluntary agreement to let me ask you
these questions by beginning this phone interview. If you qualify for the study and you are still
interested in being a participant, you will be asked to read and sign a consent form when you
come for the study.

 

 

 

 

 

Contact Information Demogranhics
Name: DOB:

First Last
Phone: (Home) (Work) Age:
Email: Gender: C] M D F
Maj or:
Educational Histogy

What is the highest educational level you have attained?
CI High school
C] Some college

Cl Graduated from college
Cl Graduate school

Language Experience

Are you a native English speaker? CI Yes CI No
Have you ever received formal instruction in a foreign language? D Yes D No
Do you speak any language other than English? Cl Yes C] No

If yes, please continue with the following questions:

What language(s)?

 

Please rate your proﬁciency (l=limited experience, 3=ﬂuent) 1 2 3

105

 

Medical Histog

Do you have any major health, medical, neurological, or psychiatric conditions? Cl Yes CI No

If yes, please specify:

 

Have you ever received treatment for speech, language, hearing, or reading? CI Yes D No

 

If yes, please specify:
Any physical handicaps? I D Yes D No Describe:
Normal vision and hearing? D Yes CI No Describe:
Any speech or language difﬁculties? D Yes CI No Describe:
Is your whole family native English speakers? D Yes Cl No Describe:
Any learning disabilities? CI Yes CI No Describe:
Previous or existing reading difﬁculty? D Yes Cl No Describe:
Training in either phonetics or phonics? CI Yes Cl No Describe:

Musical Experience

Have you ever received any musical training? Cl Yes D No

If yes, please continue with the following questions:

Please describe your experience (e.g., choir, private lessons, etc.):

 

 

How many years did you receive this training?
C] 1-2
CI 3-4
C) 5-6 ‘
C) 6+

Thank you for your participation in this initial screening. (If they are eligible for the study,
continue): Your responses show that you are eligible to participate in this study. If you decide to
continue with the study you will be asked to make one visit to campus that will last about 2 hours.
You will receive either $15. 00 or course credit for the completion of the experimental session.

Would you like to schedule your visit with us now?

Date: Time:

 

106

‘-..1& J1!

 

1' IR" 7&' #‘IL' .& 1'. 01
I

PARTICIPANT WORKSHEET

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

ID: LANGUAGE: A B
DATE: COUNTERBALANCED: A B
ENGLISH
VEXPERIMENTAL Scores PRETEST f Scores VERSION Scores
APWC APWC E- APWC
Form:A/B Form:A/B
DEIDEJEIEI 36 DEIDEIEIEI‘ 36121012113012] 36
SOST SOST E- SOST
Form:A/B Form:A/B.
100 ’ 100 100

21FC

Initial:

Final: 32

SUBTEST Raw %tile Standard

Score Score
Matrices
Memory for Digits
Nonword Repetition
Blending Nonwords
COMPOSITES 3“” %tile C°mp°s"e
of SS Score

Phonological Memory
COMMENTS:

107

 

 

APPENDIX B

Pilot Syllable Identification Data

108

 

SYLLABLE IDENTIFICATION IN ISOLATION

S les Presented for Identiﬁcation
bi bu da do ku la
3

36

Syllables within the Artiﬁcial
Language

 

ll...‘ -
1

Additional Syllables Identiﬁed in Participant Responses

  
 
 
 

9 100 36 100 49 20 0 93 89

Between

C % 93 80 93 100 47 53 100 53 87 100 93 87
Within

C 100 100 93 100 80 87 100 67 80 100 100 93
Bold highlights the most frequent response N=15

* Accuracy reﬂects percentage of correctly identiﬁed syllables over all responses between subjects.

1' Between Consistency is the percent of subjects who identiﬁed the syllable the same over at least 2 trials.
e.g., 14 of 15 subjects identiﬁed bi and di at least twice; therefore, bi has a 93% consistency value.

1 Within Consistency is the percent of subjects who identiﬁed the syllable as the same over all individual
trials. e.g., 15 of 15 subjects identiﬁed bi the same over all trials (14 as di and 1 as bi); therefore, bi has a
100% within subject consistency value. (All subjects were 100% consistent for all syllables over 2 trials.)

109

SYLLABLE IDENTIFICATION IN WORDS

S Presented for Identiﬁcation
bi bu da do ku la
1
7

ro
ti
tu
di
ha
hi
i
ki
10

Syllables within the Artiﬁcial
Language

AddItIonal Syllables
Identiﬁed in Participant
Responses

ta 4

% 0 58 100 100 33 0 100 33

 

Between
C % 75 S8 100 100 67 50 100 33

Within

% 67 67 100 100 100 67 100 100
Bold highlights the most frequent response =3

Word List: (subjects heard each syllable four times, twice in each of two words)

pigola tibudo
tudaro pabiku
bikuti golatu
budopa daropi

"' Accuracy reﬂects percentage of correctly identiﬁed syllables over all responses between subjects.

T Between Consistency reﬂects the percent of which the syllable was identiﬁed most frequently across
subjects. e.g., 9 of 12 times bi was identiﬁed as di; therefore, bi has a 75% between word consistency value.

1 Within Consistency reﬂects the percent in which a syllable was identiﬁed as the same over aﬂ trials of
an individual word. e.g., The three subjects identiﬁed bi as the same for a total of 4 out of a possible 6
times (as each of 3 subjects heard bi in two different words); therefore, bi has a 67% within word
consistency value.

110

APPENDIX C

Experimental Procedure and Response Forms

111

Testing Procedure Sequence

Screening Tests
0 Language and medical history interview

0 K-BIT
I Matrices
I CTOPP

I Memory for Digits
I Nonword Repetition
I Nonword Blending
o Otoscopy and Hearing Screening

Test Training
0 English Word Familiarization Task
0 APWC Test Training
0 SOST Test Training

Experimental Testing
- APWC Pretest
0 SOST Pretest
<5-minute Break>
0 Exposure 1
0 English APWC
0 Exposure 2
0 English SOST
<5-minute Break>
0 Exposure 3
APWC Posttest
SOST Posttest
21FC

Note: As two different tests were created for the APWC and SOST, these tests were
counterbalanced between test presentations.

112

ID: DATE: FORM: SCORE:

ENGLISH WORD FAMILIARIZATION

Possible English Word List:
Tuxedo
Textural
Juxtapose
Republic
F ixative
Explosion
Deception
Sketchily

Please number the words in the order that you hear them.

_Deception
_Juxtapose
_Republic
___Textural
_Fixative
_Tuxedo
_____Explosion

Sketchily

113

1D: DATE: FORM: SCORE:

APWC Training Form

Instructions: You will be listening to a continuous speech stream. Your job is to count
the number of words that you hear. When you hear a familiar word, press the counter.
You will hear the example once, and then it will be repeated. Please write the number
counted after each presentation.

 

EXAMPLE 1:
Please write the number of words counted (displayed on your counter):
A)

13)

 

 

EXAMPLE 2:

Clear your counter

Please write the number of words counted (displayed on your counter):
A)

B)

 

EXAMPLE 3:

Clear your counter

Please write the number of words counted (displayed on your counter):
A)

B)

114

ID: DATE: FORM: SCORE:

SOST Training Form

Instructions: You will hear a series of syllables through your headphones. Each box on
this sheet represents one syllable. Some syllables will be unrelated to each other, while
other syllables will combine to form a word in English. Your task is to place a dot in
each box as you hear a syllable and to draw a slash after the box which represents the last
syllable in the word. Each line begins with a tone and is followed by 5 beeps to signal
you to move to the next line. Follow along with the answer provided for you in the ﬁrst
example of each trial. Then, complete the second example on your own.

TRIAL 1: deception
A) I: El E1 I:I/I:I 13m
B) El El C] ['1 Cl BEH’

TRIAL 2: deception, republic
C)ElElEl/DEIEID/D 3m
D) [I El El [1 El D El 13 am

TRIAL 3: deception, republic, tuxedo
E)EIEII:I/EIDCIEI/CII:II:/I:I
F)EIEICIDEIEICICIEIDD BEEP

TRIAL 4: deception, republic, tuxedo, explosion
c)I:II:II:II:I/I:II:II:I/I:ICICII:IEI/I:II:II:I/-
H)UDEIDEIDEIEICIDDDDCIEI am

PRACTICE EXAMPLES:
1)E|CIDEIDDDEICIDDCIDDCI BEH’
2)DDUC]EIDEIDEIEIEICICICID BEEP

3)

UDDDDDEIEIEIEIDDEIEIEI BEEP

115

1D: DATE: FORM: SCORE:

APWC Response Form

Instructions: You will be listening to a continuous speech stream. Your job is to count
the number of words that you hear. When you hear a familiar word, press the counter.
When you hear a series of 5 beeps, write the number displayed on your counter in the
space provided on your AP WC response form. Please practice with the following
example.

 

EXAMPLE:

After you hear 5 beeps, please write the number of words counted (displayed on your
counter):

C)

D)

 

 

EXPERIMENT:
PLEASE CLEAR YOUR COUNTER

You will now begin. Remember; press the counter when you hear a familiar word

Please write the number of words counted (displayed on your counter):
1) __
2)
3)
4)
5)

6)

116

ID: DATE: FORM: SCORE:

SOST Response Form

Instructions: You will hear a series of syllables through your headphones. Each box on
this sheet represents one syllable. Some syllables will be unrelated to each other, while
other syllables will combine to form a word from this study’s artiﬁcial language. Your
task is to place a dot in each box as you hear a syllable and to draw a slash after the box
which represents the last syllable in the word. Each line begins with a tone and is
followed by 5 beeps to signal you to move to the next line. Please listen and follow along
with the ﬁrst example, then complete the remaining three examples.

EXAMPLES:
A)IIlElEl/EIDE1DEJ/DDEIEI/EJEID/BEEP

B)I3I:II:II:I/I:II:II:JI:I/UI:ID/I:II:II:1I:/Bm
C)DDDDDDDDDDDDDDDBEEP

D)DCIE]CICIC]CJEIEICICICICIEIEIBEFP

Now you will begin the test. Remember: Draw a slash after the last syllable in the
word.

1)UDEIEIEIEIDDEICIEIDCIDCIBEEP
2)DDDDDDDDDDDDDDDBEEP
3)DDDEIDEIDEICICIDCIDCICIBEEP
4)DCIDEIDEIDDDDDDEIDEIBEFP
5)DDDDDDDUDDUDDDDBEEP
6)UCIDEIDCIDEIDDUDDEIDBEEP

7)DUDDDDDDDUDDDDDBEIIT

 

8)DC]DE|C1CIDDE]CIEICIEIEID BEFP

 

117

 

9)DDUDDDDDDDDDDDDBEEP

10)

11)

12)

13)

14)

15)

16)

17)

18)

19)

20)

21)

22)

23)

24)

El

E1

DDDDDDDUDD

E]

E]

1:]

C1

El

El

E1

[:1

Cl

C]

118

[31:11:11]
Ell] Cl
CID 1:]
C113 1:]
CID Cl
CID

C11] [:1
DE] [:1
[IE] Cl
[11:] [:1
C113 [:1
DD El
CID E]
DC] [I
CID Cl
CID El

BEEP

BEEP

BEEP

BEEP

BEEP

BEEP

BEEP

 

 

21FC Response Form

Instructions: You will hear two words for each item. Circle the number corresponding
to the ‘familiar ” artificial word. Circle 1 if the word was the ﬁrst one presented, or 2 if
it was the second. Each item will begin with a tone and each column will end with 5
beeps. Please complete the ﬁrst 3 examples of English words.

Example 1: l 2
Example 2: 1 2
Example 3: 1 2

Please continue with the remaining items. Circle the number that corresponds to the
familiar word.

1) 1 2 12) 1 2 23) 1 2
2) 1 2 13) 1 2 24) 1 2
3) 1 2 14) 1 2 25) 1 2
4) 1 2 15) 1 2 26) 1 2
5) '1 2 16) 1 2 27) 1 2
6) 1 2 17) 1 2 28) 1 2
7) 1 2 18) 1 2 29) 1 2
8) 1 2 19) 1 2 . 30) 1 2
9) 1 2 20) 1 2 31) 1 2
10) 1 2 21) 1 2 32) 1 2
11) 1 2 22) 1 2

BEEPS BEEPS BEEPS

ID: __ DATE: __ FORM: __ SCORE: __

119

 

 

APPENDIX D

Individual Participant Data

120

Standardized Criterion Testing

Standard Scores
Memory Nonword Blending Memory
Subject Matrices for Digits Repetition“ Nonwords Composite

 

 

 

 

03Al 120 14 13 13 121
05A1 109 12 13 17 115
06A1 101 13 12 10 115
07A1 98 13 12 11 115
08A2 99 9 8 14 91
09A2 103 12 12 10 112
10A2 96 10 ll 11 103
12A2 109 13 ll 15 112
1581 115 12 10 13 106
1681 101 10 9 14 97
1782 103 8 7 9 85
18B2 94 ll 10 13 103
19B2 116 13 12 16 115
20B] 99 13 12 9 115
2181 101 11 9 11 100
23A1 94 15 10 11 115
25B2 115 12 9 10 103
2682 92 13 13 12 118
27A2 104 10 9 15 97
28B] 105 13 12 14 115
MEAN 104 12 11 12 108
SD 8.0 1.8 1.8 2.3 9.9

*Nonword Repetition was a non-criterion subtest administered to achieve the composite score.

121

 

Two-Interval Forced-Choice (21FC)

 

 

 

 

 

 

# % Initial Final
Subject Correct Correct Errors Errors
03A] 22 69 8 2
05A1 26 81 5 1
06A] 26 81 l 5
07A] 25 78 6 1
08A2 13 41 1 1 8
09A2 19 59 7 6
10A2 16 50 9 7
12A2 20 63 5 7
15B] 13 41 13 6
16B1 21 66 8 3
17B2 19 59 8 5
18B2 18 56 5 9
19B2 24 75 4 3
20B 1 1 7 53 6 9
21 B1 20 63 5 7
23A] 22 69 9 1
2532 24 75 3 5
26B2 20 63 6 6
27A2 19 59 10 4
28B1 22 69 4 6
MEAN 20.3 63.4 6.7 5.1

SD 4.0 12.0 2.9 2.5

122

Auditory Probabilistic Word Count (APWC): Total Scores

 

 

 

 

 

 

 

 

 

Pretest Posttest English

Subject Raw Abs,r Log i Raw Abs Log Raw Abs Log
03A1 0 36 1.6 5 31 1.5 23 13 1.1
05A] 51 17 1.2 25 15 1.2 17 19 1.3
06A1 54 30 1.5 15 21 1.3 28 8 0.9
07A] 43 15 1.2 60 24 1.4 31 5 0.7
08A2 90 58 1.8 20 16 1.2 31 5 0.7
09A2 13 23 1.4 23 13 1.1 27 9 1.0
10A2 179 151 2.2 126 90 2.0 32 4 0.6
12A2 25 15 1.2 58 22 1.3 65 29 1.5
15Bl 0 36 1.6 20 16 1.2 24 12 1.1
16B1 20 22 1.3 23 15 1.2 21 15 1.2
17B2 36 14 1.1 39 15 1.2 28 8 0.9
18B2 83 47 1.7 82 46 1.7 19 17 1.2
19B2 1 35 1.5 34 14 1.1 11 25 1.4
20B] 11 25 1.4 46 16 1.2 15 21 1.3
21B] 148 112 2.0 30 8 0.9 33 3 0.5
23A1 15 21 1.3 0 36 1.6 33 3 0.5
2582 0 36 1.6 35 7 0.8 33 3 0.5
26B2 67 45 1.7 87 53 1.7 28 14 1.1
27A2 43 35 1.5 24 12 1.1 27 11 1.0
28B1 150 114 2.1 24 12 1.1 30 6 0.8
MEAN 51 44 1.6 39 24 1.4 27.8 12 1.1
SD 54.0 38.0 1.6 31.0 20.0 1.3 10.8 7.6 0.9

TAbs is the Absolute Difference of the Raw Count from the Correct Count
ILog is the logarithmic transformation values of the Absolute Difference scores. This
was calculated to fulﬁll the normality assumption which was required for the analyses.

123

.885 :oLLSEmoE some E $83 .3 SALE? Looboo ELL oLmoLLEL 8:6:

wéL N.L a; N m.N N.N 94 Sam L.m VN m L.m mi m6 «.mm m6 NL L: Md LL LL Gm

 

 

 

 

 

 

 

 

 

 

wNN N.m N6 m.m 9m N.m Wm ”in Wm wN m6 a6 9 v6 v.3 ed ad ad .3 LL N.N Z<M=>L
9.. v N o N m m N m m m m m m omL mm N N N 0L NOL LmeN
NN m c c m w v N v v m m L6 o m... o v LL 2 ML o N<NN
N N o N v o m Na 3 LN LN NL a w No 0 m LN mL 3 NL NmoN
mm L6 w o N L. v mm L. o N N o m o o o o o o o NmmN
mm v w c w v m o o o o o o o m L m N N m m N L<MN
mm v w o w m V cm L. m o L. L. N m3 om N N NN N 2 LmLLN
2 N m m m N N 3. L L: N LL o w LL L L N m L m LmLON
LL N N L N L m Lam N w c o N o L L o o o o o NmmL
aL m m L. m N N Na LL NL o NL mL wL mm LL .2 LL LL cL oL waL
N v o m o v m an N m m L: w L6 om v m a L6 w c NmLNL
LN N m m L6 N m N m o m v m N 8 N m N m N m LmLoL
LN m N m m L m ON 4 v m w m N o o o o o o o meL
mo L: 2 LL LL 3 a mm a 0L 3 a w NL N o m m N o N N<NL
Nm v N m w L. L. eNL oL Lm NL NL NL 0N aNL a 9. NM LN mm 3 N<oL
NN v N o m m N N m v o m N m mL o m m v m o N<ao
Lm v w c N v N ON m v v m N N cm x NN 0L 0 Nm c N<wo
Lm L. N c o m m ow 0 2 N L: o L: M... m m 3 m a N L<No
MN m o m o m m m L L L. m N N m Vm w a LL NL E o L<oo
NL m L. m m N N N N N v N o v Lm o N N LL L: L: L<mo
N v v m m m L6 m o o N m o o o o o o o o o L<mo
cm L. No a m s S cm L. w n my L. e on L. m c m V c Suﬁsm
83m 88% :08 BL @8580 Lu Sam 83?. :28 L8 @8550 LL 85m 883 LL68 L8 3:250 u
.LuLLuem «acumen 7.80.5

LOSE—OD SALE—.2 3.3L 3550 v.53 org—Lasso; ESL—Saw

124

 

EN 94 NA 9L a; L.N ad LLN v ﬁm Nam «N m6 ed. mNm 9N Nd LN 5.9 a N.» Gm
NL L N 94 w.N 94 «N N Wm vim «in .3. ed ﬁe mév Nam N.w wN ed ad N.N 7L<HL>L

 

 

 

 

 

 

 

 

o o L o L L m NL L m m m L L LVLL ON LN NN NL NL 2 meN
LL L N o m L N NL o v m m o o mm v v m N a o N<NN
LVL N N L v N m mm mL 2 2 L6 m L Q L. m mL N L: c NmoN
m o o o L c N N o N L L N L cm L8 w o w v c NmmN
m o o o o o m cm v w o w v o LN L o v m L L6 L<mN
m o o o o L N w o m o v o L NL L N NL ON oL LN a LmLLN
LN N m m m N v cL m N L m m N N m N v m m m LmON
N N o m e m m .1 m o o w m o mm m w c w L. o NmLoL
NL L m N m N v we N v m o LL NL Ne N N m m NL mL NmmL
w o N L N o m m L m m L N v N «L o m m L6 e o NmLNL
m L N m L e N m m L L N m w L v NN N m w m m m LmoL
NL L L L m m m 0L o v .6. v L L. em v w o w L. o me L
N e N m m 3 m NN m N w L L6 c m L w m L L N Lu N<NL
v o L L o o N co NL mN LL LL 2 N LmL .6 mm oN mL Lam om N<oL
o o L o m L 6 mL L w o m N m N v m m w L o N<oo
m o o o L o v oL L v N m N v mm v S L: N N o N<wo
m o L o N L L LVN m N L N m L6 m L L o w L m L. L<No
w L N L N L L LN m L. m o N m Om w L m L. L: o L<eo
0L L L6 m m N .6 2 N c N L N N NL N L L m c v L<mo
m L o v m m L N Lm v w v w v 0 mm v w o w v c L<mo
85m 885 some 8L LOBSSU u 85m 835 some 8m @8550 La Sam 88?. some 8L 93550 n Suﬁsm
.LmLLwemL .8325 3895

EEO 39.30 :8...— ooeouoba 828.2 3550 v.83 oLLmLLLLLanLekL CELLO—E

125

 

m.OL N.MN 0.0L O.NN O.mL m.ON 0.0N m.NL O.mN Gm

 

 

 

 

 

 

 

 

 

 

MOO NON OAO O.mO O.mN 0.0m m.Nm NdL OOO 72H:
0.00L O OO O.Lm wN Om O.Lm mm OO meN
L .NO wO ON N.OL O Om m .Nm m L OO N<NN
m.OO Ow LO O.NO mN Om me LN mO NmLON
0.00L O OO L.mm mm OO 0.0 O O NmLmN
0.00L OO OO m.Om mm OO L.LO Nm OO L<mN
me NO OO N.wO Nm ON m.NN LN NO LmLLN
m.NO OO mO m.Om OL mm mmm ON NO LmLON
N.mO ON mN N.Ow Nm mm O.mO NL Om NmLOL
0.0m mm mm m.Om ON OO O.mN N N NmmL
w. LO NO wO O.mO ON mm .OO mN Om NmNL
NmO mN ON N.Om NO LN 0.0w mO mO LmLOL
O.mO Om OO O.mm mN Ow 0.0N N mm meL
m.NO OO OO m.OO ON Lm NmN LL Om N<NL
0.00 mO OO OON mN Nm m.NL O Nm N<OL
O.mm mm Om N.mO OL mm O.Nm OL NN N<OO
m. LO Ow NO m.ON mL OO O.Nm mN LN N<wO
m.NO OO NO Om m Om O.wL N mm L<NO
mmw mm OO O.NN mL mm N. LO m NL L<OO
0.00L LO LO O.mw LO LN 0.0N OO Om L<mO
0.00 NO LN L.OO mm mm LOO mL NN L<mO
8280 6\° 82.80 L88. 8880 °\o 8280 L88. 8280 6\6 8880 L88. Loo Sam
.LmLqum «8:84— .88...—

LhmomL as... 5:35.88 2:5 .55

126

APPENDIX E

Data Figures

127

5:680 02544»

Hmom
_

 

 

 

 

I m.O

I m.N

 

 

128

 

 

02m

82280 50m

FmOm
_

 

 

 

 

 

IOO
IOw

...oo_

 

 

ONL

109.1103 .IQQIIIHN

129

REFERENCES

130

 

REFERENCES

Alain, C., Cortese, F., & Picton, T. (1999). Event-related brain activity associated with
auditory pattern processing. Neuroreport, 10, 2429—2434.

American National Standards Institute (ANSI). (1996). Speciﬁcations for audiometers
(ANSI 83.6-1996). New York: Author.

Aslin, R., Saffran, J ., & Newport, E. (1998). Computation of conditional probability
statistics by 8 month-old infants. Psychological Science, 9, 321-324.

Atienza, M., Cantero, J ., Grau, C., Gomae, C., Dominguex-Marin, E., & Escera, C.
(2003). Effects of temporal encoding on auditory object formation: a mismatch
negativity study. Cognitive Brain Research, 16, 359-371.

Bailey, D. & Wolery, M. (1992). Teaching Infants and Preschoolers with Disabilities.
New York: Merrill.

Barker, J. & Cooke, M. (1999). Is the sine wave speech cocktail party worth attending?
Speech Communication, 27, 159-174.

Besson, M. & Schon, D. (2003). Comparison between language and music. In I. Peretz
and R. Zatorre (Eds), The cognitive neuroscience of music (pp.269-293). London:
Oxford University Press.

Bradley, L. & Bryant, P. (1985). Rhyme and reason in reading and spelling. Ann Arbor:
University of Michigan Press.

Brattico, E., Winkler, 1., Naatanen, R., Paavilainen, P. & Tervaniemi, M. (2002).
Simultaneous storage of two complex temporal sound patterns in auditory
sensory memory. Neuroreport, 13, 1747-1751.

Breitenstein, C. & Knecht, S. (2003). Spracherwerb und statistisches Lemen [Language
acquisition and statistical learning] [Abstract]. Nervenarzt, 74, 133-143.

Brent, M. (1999). Speech segmentation and word discovery: a computational
perspective. Trends in Cognitive Sciences, 3, 294-301.

Bryan, W. & Harter, N. (1899). Studies on the telegraphic language: The acquisition of a
hierarchy of habits. Psychological Review, 6, 345-375.

Bumham, D., Earnshaw, L., & Quinn, M. (1987). The development of the categorical
identiﬁcation of speech. In B. McKenzie and R. Day (Eds.), Perceptual
development in early infancy: Problems and Issues (pp. 237-275). Hillsdale, NJ:
Lawrence Erlbaum Associates.

131

Canﬁeld, R. & Haith, M. (1991). Young infants’ visual expectations for symmetric and
asymmetric stimulus sequences. Developmental Psychology, 27, 198-208.

Carroll, J. Davies, P. & Richman, B. (1971). Word frequency book. New York:
American Heritage Publishing Co, Inc.

Chase, W. & Simon, H. (1973). The mind’s eye in chess. In W.G. Chase (Ed.), Visual
information processing (pp. 215-281). Oxford, England: Academic Press.

Cheour, M., Ceponiene, R., Lehtokoski, A., Luuk, A., Allik, J ., Alho, K. & Naatanen, R.
(1998). Development of language-speciﬁc phoneme representations in the infant
brain. Nature Neuroscience, 1, 351-353.

Christophe, A., Dupoux, E., Bertoncini, J., & Mehler, J. (1994). Do infants perceive
word boundaries? An empirical study of the bootstrapping of lexical acquisition.
Journal of the Acoustical Society of America, 95, 1570-1580.

Cohen, J. (1988). Statistical power analysis for the behavioral sciences (2nd ed.).
Hillsdale, N.J.: L. Erlbaum Associates.

Cole, R., & Jakimik, J. (1980). A model of speech perception. Hillsdale, NJ: Lawrence
Earlbaum Associates.

Coney, J. (2002). Lateral asymmetry in phonological processing: Relating behavioral
measures to neuroimaged structures. Brain and Language, 80, 355-365.

Conroy, C. (2003). The effects of four variables on the intelligibility of synthesized
sentences [Abstract]. Journal of the Acoustical Society of America, 114, 2366.

Conway, C., & Christiansen, M. (2001). Sequential learning in non-human primates.
Trends in Cognitive Sciences, 5, 539-546.

Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped how
humans reason? Studies with the Wason selection task. Cognition, 31, 187-276.

Craik, F. & Lockhart, R. (1972). Levels of processing: A framework for memory
research. Journal of Verbal Learning and Verbal Behavior, 11, 671-684.

Cutler, A., & Carter, D. (1987). The predominance of strong initial syllables in the
English vocabulary. Computer Speech and Language, 2, 133-142.

Cutler, A., & Norris, D. (1988). The role of strong syllables in segmentation for lexical
stress. Journal of Experimental Psychology: Human Perception and
Performance, 14, 113-121.

Davis, G. (2000). Aphasiology: Disorders and clinical practice. Boston: Allyn & Bacon.

Eimas, P., Siqueland, E., Jusczyk, P., & Vigorito, J. (1971). Speech perception in

132

 

infants. Science, 171, 303-306.

Elrnan, J. & Bates, E. (1997). Acquiring language [Letter to the editor]. Science, 276,
1180.

Felleman, D. & Van Essen, D. (1991). Distributed hierarchical processing in the primate
cerebral cortex. Cerebral Cortex, 1, 1-47.

Fiser, J ., & Aslin, R. (2002). Statistical learning of higher-order temporal structure from
visual shape sequences. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 28, 458-467.

Fitch, R., Miller, S. & Tallal, P. (1997). Neurobiology of speech perception. Annual
Review of Neuroscience, 20, 331-353.

Francis, A. (1998). Perceptual learning of synthetic speech [Abstract]. Journal of the
Acoustical Society of America, 104, 1835.

Fritch, 8., Large, N., Pisoni, D. (2000). Perception of wordlikeness: Effects of segment
probability and length on the processing of nonwords. Journal of Memory and
Language, 42, 481-496.

Galotti, K. (1999). Cognitive Psychology In and Out of the Laboratory. Boston:
Wadsworth Publishing Company.

Gobet, F ., Lane, P., Crocker, S., Cheng, P., Jones, G., Oliver, 1. & Pine, J. (2001).
Chunking mechanisms in human learning. Trends in Cognitive Science, 5, 236-
243.

Gomez, R. & Gerken, L. (1999). Artiﬁcial grammar learning by l-year-olds leads to
speciﬁc and abstract knowledge. Cognition, 70, 109-135.

Gomez, R. & Gerken, L. (2001). Infant artiﬁcial language learning and language
acquisition. In M. Tomasello & E. Bates (Eds.), Language development: The
essential readings (pp. 42-48). Malden, MA: Blackwell Publishers Ltd.

Goodsitt, J ., Morgan, J ., & Kuhl, P. (1993). Perceptual strategies in prelingual speech
segmentation. Journal of Child Language, 20, 229-252.

Grifﬁths, T. (2003). The neural processing of complex sounds. In I. Peretz and R. Zatorre
(Eds), The cognitive neuroscience of music (pp. 167-177). London: Oxford
University Press.

Grosjean, F. (1985). The recognition of words after their acoustic offset: Evidence and
implications. Perception & Psychophysics, 38, 299-310.

Hauser, M., Newport, E., & Aslin, R. (2001). Segmentation of the speech stream in a
non-human primate: statistical learning in cotton-top tamarins. Cognition, 78,

133

 

B53-B64.

Hayes, J. & Clark, H. (1970). Experiments on the segmentation of an artiﬁcial speech
analogue. In J. Hayes (Ed.), Cognition and the Development of Language (pp.
221-223). New York: John Wiley & Sons.

Hohne, E., & Jusczyk, P. (1994). Two-month-old infants’ sensitivity to allophonic
differences. Perception and Psychophysics, 56 613- 623.

Hollich, G.J., Hirsh-Pasek, K., & Golinkoff, RM. (2000). Breaking the language barrier:
An emergentist coalition model for the origins of word learning. Monographs of
the Society for Research in Child Development, 65 (3).

Houston, D., Jusczyk, P., Kuijpers, C., Coolen, R., & Cutler, A. (2000). Cross-language
word segmentation by 9-month-olds, Psychonomic Bulletin & Review, 7, 504-
509.

Hu, C.-F. (2003). Phonological Memory, Phonological Awareness, and Foreign
Language Word Learning. Journal of Learning Language, 53, 429-462.

Huettel, S., Mack, P., & McCarthy, G. (2002). Perceiving patterns in random series:
dynamic processing of sequence in prefrontal cortex. Nature Neuroscience, 5,
485-490.

Hunt, R. & Aslin, R. (2001). Statistical learning in a serial reaction time task: Access to
separable statistical cues by individual learners. Journal of Experimental
Psychology: General, 130, 658-680.

Irnan, R. (1994). A data-based approach to statistics. Belmont, CA: Duxbury Press.

Johnson, E., & Jusczyk, P. (2001). Word segmentation by 8-month-olds: When speech
cues count more than statistics. Journal of Memory and Language, 44, 548-567.

Johnson, J. & Newport, E. (1991). Critical period effects on universal properties of
language: The status of subjacency in the acquisition of a second language.
Cognition, 39, 215-258.

Jourdain, R. (1997). Music, the brain and ecstasy. New York: William Morrow & Co.
J usczyk, P. (1997). The discovery of spoken language. Cambridge, MA: MIT Press.

Jusczyk, P. (1998). Using the headtum preference procedure to study language. In C.
Rovee-Collier, L. P. Lipsitt, & H. Hayne (Eds.), Advances in infancy research
(Vol. 12, pp. 188-204). Stanford, CT: Ablex.

Jusczyk, P. (1999). How infants begin to extract words from speech. Trends in Cognitive
Science, 3, 323-328.

134

 

Jusczyk, P., & Aslin, R. (1995). Infants’ detection of sound patterns of words in ﬂuent
speech. Cognitive Psychology, 29, 1-23.

J usczyk, P., & Hohne, E. (1997). Infants’ memory for spoken words. Science, 277,
1984-1986.

Jusczyk, P., Hohne, E., & Bauman, A. (1999). Infants’ sensitivity to a110phonic cues for
word segmentation. Perception & Psychophysics, 61, 1465-1476.

J usczyk, P., Houston, D., & Newsome, M. (1999). The beginnings of word segmentation
in English-learning infants. Cognitive Psychology, 39, 159-207.

Kaufman, A., & Kaufman, N. (1990). Kaufman brief intelligence test. Circle Pine, MN:
American Guidance Services.

Kersten, D. & Yuille, A. (2003). Bayesian models of object perception. Current Opinion
in Neurobiology, 13, 150-158.

Kirkham, N., Slemmer, J ., & Johnson, S. (2002). Visual statistical learning in infancy:
evidence for a domain general learning mechanism. Cognition, 83, B35-B42.

Kluender, K., Diehl, R., & Killeen, P. (1987). Japanese quail can learn phonetic
categories. Science, 237, 1195-1197.

Kubovy, M. & Van Valkenburg, D. (2001). Auditory and visual objects. Cognition, 80,
97-126.

Kuhl, K., Williams, K., Lacerda, F., Stevens, K., & Lindblom, B. (1992). Linguistic
experience alters phonetic perception in infants by 6 months of age. Science,
255, 606-608.

Kuhl, P. & Miller, J. (1975). Speech perception by the chinchilla: Voiced-voiceless
distinction in alveolar plosive consonants. Science, 190, 69-72.

Kuhl, P. & Padden, D. (1982). Enhanced determinability at the phonetic boundaries for
the voicing feature in macaques. Perception and Psychophysics, 32, 542-550.

Kulynych, J ., Vladar, K., Jones, D., & Weinberger, D. (1994). Gender differences in the
normal lateralization of the supratemporal cortex: MRI surface-rendering
morphometry of Heschl’s gyrus and planum temporale. Cerebral Cortex, 4, 107-
118.

Lake, D. & Bryden, M. (1976). Handedness and sex differences in hemispheric
asymmetry. Brain and Language, 3, 266-282.

Lambe, E. (1999). Dyslexia, gender, and brain imaging. Neuropsychologia, 37, 521-536.

Liebenthal, E., Binder, J ., Piorkowski, R., & Remez, R. (2003). Short-term

135

 

reorganization of auditory analysis induced by phonetic experience. Journal of
Cognitive Neuroscience, 15, 549-558.

Lovaas, I. (1977). The autistic child: language development through behavior
modification. New York: Irvington Publishers.

Lovett, M. (2002). Problem solving. In D. Medin & H. Pashler (Eds.), Stevens ’
Handbook of Experimental Psychology, Third Edition, Volume Two: Memory and
Cognitive Processes (pp. 317-362). New York: John Wiley & Sons.

Maillart, C., Schelstraete, M.-A., & Hupet, M. (2004). Phonological representations in
children with SLI: A study of French. Journal of Speech, Language, and
Hearing Research, 4 7, 187-198.

Marslen-Wilson, W. D. (1987). Parallel processing in spoken word recognition,
Cognition, 25, 71-102.

Mattys, S., & Jusczyk, P. (20013). Phonotactic cues for segmentation of ﬂuent speech by
infants. Cognition, 78, 91-121.

Mattys, S., & Jusczyk, P. (2001b). Do infants segment words or recurring contiguous
patterns? Journal of Experimental Psychology: Human Perception and
Performance, 27, 644-655.

Mattys, S., J usczyk, P., Luce, P., & Morgan, J. (1999). Phonotactic and prosodic effects
on word segmentation in infants. Cognitive Psychology, 38, 465-494.

McClure, K., Ferreira, F ., & Bisanz, G. (1996). Effects of grade, syllable segmentation,
and speech of presentation on children’s word-blending ability. Journal of
Educational Psychology, 88, 670-681.

McQueen, J. (1998). Segmentation of continuous speech using phonotactics. Journal of
Memory and Language, 3 9, 21-46.

McQueen, J ., Norris, D., & Cutler, A. (1994). Competition in spoken word recognition:
Spotting words in other words. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 20, 621-638.

Morgan, J. (1996). A rhythmic bias in preverbal speech segmentation. Journal of
Memory and Language, 35, 666-688.

Murray, S., Kersten, D., Olshausen, B., Schrater, P., & Woods, D. (2002). Shape
perception reduces activity in human primary visual cortex. Proceedings of the
National Academy of Science, 99, 15164-15169.

Naatanen, R. (1992). Attention and brain function. Hillsdale, NJ: Lawrence Erlbaum
Associates.

136

 

Naatanen, R. (1995). The mismatch negativity: A powerful tool for cognitive
neuroscience, Ear & Hearing, 16, 6-18.

Naatanen, R., Schroger, E., Karakas, S., Tervaniem, M., & Paavilainen, P. (1993).
Development of a memory trace for complex sound in the human brain.
Neuroreport, 4, 503-506.

Nemoy, E. & Davis, S. (1954). The correction of defective consonant sounds. Magnolia,
MA: Expression.

Newport, E. & Aslin, R. (2004). Learning at a distance 1. Statistical learning of non-
adjacent dependencies. Cognitive Psychology, 48, 127-162.

Nittrouer, S. (2001). Challenging the notion of innate phonetic boundaries. Journal of
the Acoustical Society of America, I 10, 1598-1605.

Norris, D., McQueen, J ., & Cutler, A. (1995). Competition and segmentation in spoken-
word recognition. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 21, 1209-1228.

Pantev, C., Oostenveld, R., Engelien, A., Ross, B., Roberts, L., & Hoke, M. (1998).
Increased auditory cortical representation in musicians. Nature, 392, 811-814.

Pascual-Leone, A. & Walsh, V. (2001). Fast backprojections from the motion to the
primary visual area necessary for visual awareness. Science, 292, 510-512.

Pugh, K., Shaywitz, B., Shaywitz, S., Constable, R., Skudlarski, P., Fulbright, R.,
Bronen, R., Shankweiler, D., Katz, L., Fletcher, J ., & Gore, J. (1996). Cerebral
organization of component processes in reading. Brain, 119, 1221-1238.

Ramus, F., Hauser, M., Miller, C., Morris, D., & Mehler, J. (2000). Language
discrimination by human newborns and by cotton-top tamarin monkeys. Science,
288, 349-351.

Rauschecker, J. (1998). Cortical processing of complex sounds. Current Opinion in
Neurobiology, 8, 516-521.

Reber, A. (1989). Implicit learning and tacit knowledge. Journal of Experimental
Psychology: General, 1 18, 219-235.

Reynolds, M., Bond, 2., Fucci, D. (1996). Synthetic speech intelligibility: Comparison
of native and non-native speakers of English. Augmentative and Alternative
Communication, 12, 32-36.

Ritter, W., Deacon, D., Gomes, H., Javitt, D., & Vaughan, H. (1995). The mismatch
negativity of event-related potentials as a probe of transient auditory memory: a
review. Ear & Hearing, 16, 52-67.

137

 

Roy, D. & Pentland, A. (2002). Learning words from sights and sounds: a computational
model. Cognitive Science, 26, 113-146.

Saffran, J. & Wilson, D. (2003). From syllables to syntax: Multilevel statistical learning
by lZ-month-old infants. Infancy, 4, 273-284.

Saffran, J. (2001). Words in a sea of sounds: The output of statistical learning. Cognition,
81, 149-169.

Saffran, J. (2002). Constraints on statistical language learning. Journal of Memory and
Language, 47, 172-196.

Saffran, J. (20033). Statistical language learning: Mechanisms and constraints. Current
Directions in Psychological Science, 12, 110-1 14.

Saffran, J. (2003b). Mechanisms of musical memory in infancy. In I. Peretz and R.
Zatorre (Eds), The cognitive neuroscience of music (pp. 32-41). London: Oxford
University Press.

Saffran, J ., Aslin, R., & Newport, E (1996). Statistical learning by 8-month-old infants.
Science, 274, 1926-1928.

Saffran, J., Johnson, E., Aslin, R., & Newport, E. (1999). Statistical learning of tone
sequences by infants and adults. Cognition, 70, 27-52.

Saffran, J ., Newport, E., & Aslin, R. (1996). Word segmentation: The role of
distributional cues. Journal of Memory and Language, 35, 606-621.

Sanders, L., Newport, E., & Neville, H. (2002). Segrnenting nonsense: An event-related
potential index of perceived onsets in continuous speech. Nature Neuroscience,
5, 700-703.

Sasaki, T., Suetomi, D., Nakajima, Y., & ten Hoopen, G. (2002). Time-shrinking, its
propagation, and gestalt principles. Perception & Psychophysics, 64, 919-931.

Sato, Y., Yabe, H., Hiruma, T., Sutoh, T., Shinozaki, N., Nashida, T., & Kaneko, S.
(2000). The effect of deviant stimulus probability on the human mismatch
process. Neuroreport, 11, 3703-3708.

Scientiﬁc Learning Corporation. (1998) Fast For WordTM [Computer Software].
Berkeley, CA: Author.

Shanna, J ., Angelucci, A., & Sur, M. (2000). Induction of visual orientation modules in
auditory cortex. Nature, 404, 841-847.

Shaywitz, B., Shaywitz, S., Pugh, K., Constable, R., & Skudlarski, P. (1995). Sex
differences in the functional organization of the brain fro language. Nature, 3 73,

138

 

607-609.

Simon, H. & Gobet, F. (2000). Expertise effects in memory recall: Comment of Vicente
and Wang (1998). Psychological Review, 107, 593-600.

Simos, P. 8; Molfese, D. (1997). Electrophysiological responses from a temporal order
continuum in the newborn infant. Neurophsychologia, 35, 89-98.

Steinschneider, M., Schroeder, C., Arezzo, J ., & Vaughan, H. (1995). Speech-evoked
activity in primary auditory cortex: Effects of voice onset time.

Electroencephalography and Clinical Neurophysiology: Evoked Potentials, 92,
30-43.

Sur, M. & Leamey, C. (2001). Development and plasticity of cortical areas and
networks. Nature Reviews Neuroscience, 2, 251-262.

Terrance, H. (1991). Chunking during serial learning by a pigeon: 1. basic evidence.
Journal of Experimental Psychology: Animal Behavior Processes, I 7, 81-93.

Tervaniemi, M. (2003). Musical sound processing: EEG and MEG evidence. In I. Peretz
and R. Zatorre (Eds), The cognitive neuroscience of music (pp. 294-309).
London: Oxford University Press.

Torgesen, J ., & Morgan, S. (1990). Phonological synthesis tasks: A developmental,
functional, and componential analysis. In H. Swanson & B. Keogh (Eds.),
Learning disabilities: Theoretical and research issues (pp. 263-276). Hillsdale,
N.J.: L. Erlbaum Associates.

Torgesen, J ., Morgan, 8., & Davis, C. (1992). Effects of two types of phonological
awareness training on word learning in kindergarten children. Journal of
Educational Psychology, 84, 364

Tu, Z. & Zhu, S-C. (2002, April). Parsing images into region and curve processes. In A.
Heyden, G. Sparr, M. Nielsen, P. J ohnasen (Eds.), Proceedings of the 7'”
European Conference on Computer Vision. Berlin, Heidelberg: Springer-Verlag.
Abstract retrieved January 21, 2004, from http://link.springer.de/link/service/
series/0558/tocs/t2352.htm#toc2352.

Vaz Pato, M., Jones, S., & Sprague, P. (2002). Mismatch negativity to single and
multiple pitch-deviant tones in regular and pseudo-random complex tone
sequences. Clinical Neurophysiology, 113, 519-527.

Vitevitch, M., & Luce, P. (1999). Probabilistic and neighborhood activation in spoken
word recognition. Journal of Memory and Language, 40, 374-408.

Vollmeyer, R., Burns, B., & Holyoak, K. (1996). The impact of goal speciﬁcity and
systematic of strategies on the acquisition of problem structure. Cognitive

139

 

Science, 20, 75- 100.

Wagner, R., Torgesen, J ., & Rashotte, C. (1994). Development of reading-related
phonological processing abilities: New evidence of bi-directional causality ﬁom a
latent variable longitudinal study. Developmental Psychology, 30, 73-87.

Wagner, R., Torgesen, J ., & Rashotte, C. (1999). Comprehensive test of phonological
processing. Austin, TX: PRO—ED.

Wagner, R., Torgesen, J ., Rashotte, C., Hecht, S., Barker, T., Burgess, S., Donahue, J ., &
Garon, T. (1997). Changing relations between phonological processing abilities
and word-level reading as children develop from beginning to skilled readers: A
5-year longitudinal study. Developmental Psychology, 33, 468-479.

Winkler, I. & Czigler, I. (1998). Mismatch negativity: Deviance detection or the
maintenance of the "standard." Neuroreport, 9, 3809-3813.

Winkler, I., Cowan, N., Csepe, V., Czigler, I., & Naatanen, R. (1996). Interactions
between transient and long-tenn auditory memory as reﬂected by mismatch
negativity. Journal of Cognitive Neuroscience, 8, 405-415.

Winkler, I., Schroger, E., & Cowan, N. (2001). The role of large-scale memory
organization in the mismatch negativity event-related brain potential. Journal of
Cognitive Neuroscience, 13, 59-71.

Woodcock, R. & Johnson, M. (1989, 1990). Woodcock-Johnson psycho-educational
battery-revised. Allen, TX: DLM Teaching Resources.

140

 

    

 
 

C E V

l[will];[gllglllilllllll