THE GENERALIZABLE NATURE OF LEXICAL RETUNING

By

Scott Nelson

A THESIS

Michigan State University

in partial fulfillment of the requirements

Submitted to

for the degree of

Linguistics – Master of Arts

2019

ABSTRACT

THE GENERALIZABLE NATURE OF LEXICAL RETUNING

By

Scott Nelson

Auditory speech identification has been observed to be influenced by both lexical and visual
information. Perceptual learning experiments have used two unique paradigms to test how each of
these information sources affects the identification of ambiguous stimuli. In both cases, listeners
are more likely to identify ambiguous stimuli in the direction of the disambiguating information
they receive. It has been further argued that the resulting effects are the same and can be traced back
to the same general speech perception mechanism. Despite this claim, there have been conflicting
results in regards to generalization. Lexically induced perceptual learning has been observed to
generalize to new contexts, while visually induced perceptual learning has been observed to be
context dependent.

While the difference in these observed results could be explained by the information source
(lexical vs. visual), there are also crucial differences in the experimental designs that may offer
a better account. The training stimuli set for lexically induced perceptual learning experiments
includes many unique tokens that are presented one time each. For visually induced perceptual
learning experiments, the training set includes just one unique token presented multiple times.
Listeners therefore only receive type variation in the lexically induced perceptual learning experi-
ments. Crucially, type variation has been observed to be necessary for learning linguistic patterns
and therefore may explain the differences in observed results between the two paradigms.

This current study uses three new experiments to study the generalizable nature of lexically
induced perceptual learning. The results corroborate the idea that generalization of the effect to
new contexts is possible in lexically induced perceptual learning experiments when listeners are
trained with type variation, but when type variation is eliminated the ability to generalize the effect
to new contexts is no longer observed.

ACKNOWLEDGEMENTS

I am heavily indebted to Dr. Karthik Durvasula and would first and foremost like to extend the most
heartfelt thank you to him for all of the help and guidance he has provided me with during my time
at Michigan State University. This thesis would not have been possible without his encouragement
and support, and his technical and theoretical knowledge helped shape it into its current form. I
am lucky to have had a mentor who I could discuss technical phonetic measurements with one day
and then spend an hour discussing the philosophy of science the next. And I will always remember
that “it’s just simple physics.”

I would also like to thank Dr. Yen-Hwei Lin who, in my first ever phonology class, said to me,
“you think like a linguist.” Those words have stuck with me since that moment and her critical
advice and support have been crucial to my growth as a researcher Dr. Alan Beretta deserves a
special thank you as well for his guidance and willingness to sit on my committee.

I would also like to acknowledge Dr. Suzanne Wagner, Dr. Alan Munn, and Dr. Marcin
Morzycki for teaching me how to be a linguist both in and out of the classroom. My peers in the
graduate program also deserve a special mention. Cara Feldscher, Chad Hall, Alex Mason, Monica
Nesbit, and especially Kaylin Smith have made life more enjoyable even on the darkest days.

It would be foolish of me to sound off without thanking my parents, Bob and Kris Nelson. They
have supported me unconditionally throughout the unconventional and every-changing path that
has lead me here. Finally, I want to thank Kameron Chauvez, Shelbye Herbin, and Benjamin Linus
Nelson for being the best friend, partner, and dog during this stressful (but rewarding!!) moment
of my life.

iii

TABLE OF CONTENTS

.
.

.
.

.
.

.
.

.
.

. .
.

LIST OF TABLES .
LIST OF FIGURES .
CHAPTER 1
CHAPTER 2 BACKGROUND .

INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
2.1 Lexical retuning .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Audio-visual recalibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3 Comparison of the two paradigms . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
vi
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
1
4
4
6
8
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.3.1 Generalization .

.

.

.

CHAPTER 3 EXPERIMENT 1: REPRODUCING GENERALIZATION IN LEX-

ICAL RETUNING EXPERIMENTS . . . . . . . . . . . . . . . . . . .

3.1 Method .

.

.

.

.

.

.
3.1.1 Participants
3.1.2 Design .
.
3.1.3 Materials .
3.1.4 Procedure
.
.

.
.

.
.

.
.

3.2 Results
.
3.3 Discussion .

.

.

.
.
.
.
.

. 15
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
. 19
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

CHAPTER 4 EXPERIMENT 2: GENERALIZATION WITH NON-IDENTICAL

4.1 Method .

.

.

.

.

.

TRAINING AND TESTING CONDITIONS . . . . . . . . . . . . . . . 25
. 25
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
. 28
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

.
4.1.1 Participants
4.1.2 Design .
.
4.1.3 Materials .
4.1.4 Procedure
.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.

.
.

.
.

.
.

.

4.2 Results
.
4.3 Discussion .

.

CHAPTER 5 EXPERIMENT 3: REDUCING STIMULUS VARIATION IN LEX-

5.1 Method .

.

.

.

.

.

. 33
ICAL RETUNING . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 33
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
. 34
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

.
5.1.1 Participants
5.1.2 Design .
.
5.1.3 Materials .
5.1.4 Procedure
.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.

.
.

.
.

.
.

.

5.2 Results
.
5.3 Discussion .

.

iv

6.1 Summary of Results .
6.2 General Discussion .
.
6.3 Conclusion .
.
.
.
.
.

CHAPTER 6 DISCUSSION AND CONCLUSION . . . . . . . . . . . . . . . . . . . . 39
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
. 41
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
.
APPENDIX A LEXICAL DECISION TASK WORD LIST . . . . . . . . . . . . . 44
APPENDIX B LDT TRAINING WORD FREQUENCY INFORMATION . . . . 46

.
.
.
.

.
.

APPENDICES .

.

.

.
.

.
.

BIBLIOGRAPHY .

. .

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 48

v

LIST OF TABLES

Table 2.1: Comparison of lexical retuning and audio-visual recalibration paradigms

. . . .

8

Table A.1: Training words for Lexical Decision Tasks . . . . . . . . . . . . . . . . . . . . . 44

Table A.2: Filler words for Lexical Decision Tasks

. . . . . . . . . . . . . . . . . . . . . . 45

Table B.1: Experiment 1 LDT training word frequency data . . . . . . . . . . . . . . . . . 46

Table B.2: Experiment 2/3 LDT training word frequency data . . . . . . . . . . . . . . . . 47

vi

LIST OF FIGURES

Figure 2.1: Results adapted from Norris et al. (2003) . . . . . . . . . . . . . . . . . . . . .

Figure 2.2: Results adapted from Bertelson et al. (2003)

. . . . . . . . . . . . . . . . . . .

5

7

Figure 3.1: General Experiment Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Figure 3.2: Categorization results for the fi∼si continuum pre-test

. . . . . . . . . . . . . . 18

Figure 3.3: Phonetic categorization Psychopy screen and instructions . . . . . . . . . . . . 20

Figure 3.4: LDT Psychopy screen and instructions . . . . . . . . . . . . . . . . . . . . . . 20

Figure 3.5: Categorization results for Experiment 1 . . . . . . . . . . . . . . . . . . . . . . 22
Figure 4.1: Categorization results for the fa∼sa continuum pre-test . . . . . . . . . . . . . . 28

Figure 4.2: Categorization results for Experiment 2 . . . . . . . . . . . . . . . . . . . . .

. 30

Figure 5.1: Categorization results for Experiment 3 . . . . . . . . . . . . . . . . . . . . .

. 36

Figure B.1: Boxplots for Experiment 1 LDT training word frequency (log) . . . . . . . . . . 46

Figure B.2: Boxplots for Experiment 2 LDT training word frequency (log) . . . . . . . . . . 47

vii

CHAPTER 1

INTRODUCTION

A theory of how listeners process ambiguous speech sounds is an important part of a larger theory
of speech perception. If the ultimate goal of a listener is to reverse infer some discrete underlying
representation from the continuous speech signal (Gaskell and Marslen-Wilson, 1996, 1998; Gow,
2003; Durvasula and Kahng, 2015, 2016; Durvasula et al., 2018), receiving ambiguous input
increases the difficulty of this task. When the auditory input is sufficiently ambiguous, a listener
may search outside the auditory domain in order to find cues that may help provide disambiguating
information. Past research into these types of phenomena have shown that identification can be
affected by lexical (Ganong, 1980) or visual (McGurk and MacDonald, 1976) information. More
recent experimental paradigms have used either lexical or visual information to show that listeners’
identification responses can be predictably biased towards a specific side of a perceptual boundary
based on the disambiguating information they receive. Lexical retuning uses lexical information to
bias listeners responses in a specific direction (Norris et al., 2003), while audio-visual recalibration
relies on visual information to guide a listener towards identification (Bertelson et al., 2003). In
both cases, it is non-auditory information that ultimately leads the listeners to their identification
of the target sounds.

One question that falls from this fact is whether all forms of disambiguating information are
treated equally by listeners. Is it the case that certain types of input are given a privileged status
when decoding the speech signal? Despite relying on different information sources, lexical retuning
and audio-visual recalibration often appear to have very similar consequences. In fact, Van Linden
and Vroomen (2007) ran a series of five experiments to show that in all cases, both paradigms lead
to results that are similar to one another. Due to these findings, they claimed that both identification
strategies were part of the same underlying perception mechanism. On the surface, these results
seem to suggest that all disambiguating information may be treated equally, but a more fine-grained
analysis shows the necessity for further inquiry.

1

The validity of Van Linden and Vroomen’s (2007) claim has since been questioned due to
experimental results in both domains. A shortcoming of the original study was that, despite
running five experiments, they were unable to test every dimension in which the two experimental
paradigms had potential for variation. One conflicting dimension that their experiments did not
cover is in the realm of generalization. Using lexical retuning experiments, it has been argued
that the perceptual retuning effect is able to generalize over features (Kraljic and Samuel, 2006;
Durvasula and Nelson, 2018), across syllabic position (Jesse and McQueen, 2011), and to previously
unheard words (McQueen et al., 2006a). In contrast, audio-visual recalibration results suggest that
the recalibration effect is strongly contextually bound (Reinisch et al., 2014). Reinisch et al. (2014)
found that the phonological environment of the training and testing segments, the phonetic cues, and
the segment itself needed to be identical in order for the effect to occur, ultimately demonstrating
that generalization was not possible under their experimental conditions.

The results regarding generalization question the original claim put forth by Van Linden and
Vroomen (2007). Perhaps lexical information is more privileged than visual information when
tested over the proper dimension. Based on the data Van Linden and Vroomen (2007) had, their
conclusion was reasonable, but new evidence may require an update to the claim that lexical retuning
and audio-visual retuning are tapping into the same perceptual mechanism. Despite results over
certain dimensions looking the same, it is still possible that the two are tapping into different
levels of representation; which could therefore explain the difference in generalizability. A deeper
inspection of lexical retuning could shed greater insight on what types of representations the effect
is tapping into and whether or not these are more abstract (possibly, phonological) in nature.

The option to place a distinction between high-level lexical information and low-level visual
information is enticing, but a counter argument to this source-based story can be made if we consider
the stimuli used in these retuning/recalibration studies. In lexical retuning experiments, the training
stimuli is a set of lexical items, each of which the listener hears once. In audio-visual recalibration
experiments, there is typically only one training item that a listener is exposed to multiple times.
The key here is that even when listeners are exposed to the same number of total stimuli, the number

2

of contexts in which they are exposed to the ambiguous input varies between the two paradigms.
It is possible that the reason that learners in the audio-visual recalibration paradigm are unable to
generalize the effect is because they internalize the ambiguity as an idiosyncratic pronunciation
of that single string. Contextual variation has be shown to be important in generalizing linguistic
patterns (Gerken and Bollt, 2008; Denby et al., 2018). It is therefore important to see what effect
stimulus variation has on lexical retuning before exploring representations further. In order to do
this, I will use a series of new experiments to further test the generalizable nature of the lexical
retuning effect.

The remainder of this thesis will be formatted as follows. In chapter 2 I will give an explanation
and background of lexical retuning and audio-visual recalibration. This will include detailed
descriptions of the experimental paradigms as well as a comparison between the two. Part of this
chapter will cover generalization as it relates to the overall perceptual learning literature. This
will help to set up the motivations for the experiments described in chapter 3, chapter 4, and
chapter 5. Chapter 6 will contain a general conclusion and recap what role each experiment will
have on theories of speech perception and our overall understanding of perceptual retuning and
recalibration.

3

CHAPTER 2

BACKGROUND

2.1 Lexical retuning

The lexical retuning paradigm was developed by Norris et al. (2003) in order to test the effect
of lexical feedback on adapting to an unusual speaker. In this original set of experiments, they
predicted that, because of the bias that gets induced vis-à-vis the “ganong effect” (Ganong, 1980),
listeners’ identification responses to an f∼s continuum would vary depending on in which words
a listener heard an ambiguous blend of [f] and [s] ([? f s]).1 For their experiment, they created a
list of 40 Dutch words containing either /f/ or /s/ (20 of each) that were non-minimal pairs for the
opposing sound (e.g., witlof “chicory”; naaldbos “pine forest”) and a blended 41-step continuum
of [f] and [s] sounds spliced onto an [E] vowel. They chose 14 steps from the continuum to use
in a pre-test. The results of the pre-test were used to find the maximally ambiguous point on the
continuum and five steps to be used in the testing portion of the main experiment. The fricative
portion of the maximally ambiguous point was then used as [? f s] in the training portion of the
experiment.

The overall design of the experiment was a lexical decision task (LDT) followed by phonetic
categorization of the five steps identified in the pre-test. The LDT was broken into three groups.
Each list had 100 real words and 100 nonce words. Of the 100 real words, 40 of them were the words
that contained either /f/ or /s/. One group of participants heard [? f s] in the /f/-words, a second group
heard it in /s/-words, and a final group heard it in phonotactically licit Dutch nonce-words. The
group who heard [? f s] in nonce words also heard the regular /f/ and /s/ words with their standard
unambiguous pronunciations.

It was predicted that participants who heard the ambiguous segment in words normally con-
taining /f/ would give more “f” responses on the phonetic categorization task, especially for the

1Hereafter, I will refer to a sound ambiguous between two sounds [X] and [Y] as [?XY ].

4

points on the segment that were already ambiguous. Similarly, listeners who heard [? f s] in words
that normally contained /s/ would do the opposite and give more “s” responses. The third group
who heard the ambiguous sound in nonce words were predicted to act as a control group. Their
responses were therefore predicted to lie in the middle of the other two groups.

Results from their first experiment are shown below in Figure 2.1. “?f” represents the group
that heard [? f s] in /f/-words and “?s” represents the group that heard [? f s] in /s/-words. The “non”
refers to the control group that heard [? f s] in nonce words. The x-axis is step number on the f∼s
continuum and the y-axis is how many “s” responses were given. The “?f” group gave significantly
fewer “s” responses and then “?s” group gave significantly more “s” responses. There was no
statistically significant difference between the control group and either of the other groups.

Figure 2.1: Results adapted from Norris et al. (2003)

These results suggest that lexical information can modulate the way in which ambiguous stimuli
are categorized. Using a second experiment in their original study, Norris et al. (2003) confirm that
their results cannot be explained by selective adaptation (Eimas and Corbit, 1973; Samuel, 1986). 2
With both of these results in hand, it is proposed that during the speech perception process listeners
are getting lexical feedback not for on-line recognition but instead for learning (Norris et al., 2003,
2Selective adaptation refers to the phenomenon where repeated presentation of a stimulus results

in temporary insensitivity to that stimulus.

5

p. 233-234).3 Therefore, this paradigm and the resulting effect have been frequently referred to as
“perceptual learning” (Norris et al., 2003; Eisner and McQueen, 2005, 2006; Kraljic and Samuel,
2005, 2006; Samuel and Kraljic, 2009). Since the original study, other segments beyond fricatives
have also been shown to work within the paradigm. Stops (Kraljic and Samuel, 2006), liquids
(Scharenborg et al., 2011), and lexical tones (Mitterer et al., 2011), and phonotactic information
(Cutler et al., 2008) have all been shown to elicit and effect using lexical retuning experiments,
suggesting that the effect is generally not something specific to certain acoustic cues, but rather an
adaptation response to ambiguous, but categorizable, speech.

If lexical retuning is modeled as a learning effect then it is fairly clear how it may lead to
higher level abstraction/generalization. The representation for the two segments in competition are
constantly being updated, and the sampling distribution for an individual changes depending on
which words they heard the ambiguous segment in. A listener learns in the sense that they need to
learn the distribution for the speaker they are hearing (Kleinschmidt and Jaeger, 2015).

2.2 Audio-visual recalibration

During the same time frame that the lexical retuning paradigm was beginning, the audio-
visual recalibration paradigm was developed by Bertelson et al. (2003). Here, it was knowledge
of the McGurk effect (McGurk and MacDonald, 1976) that drove its creation. McGurk and
MacDonald (1976) observed that when a listener gets visual (lipread) information that contradicts
the auditory input, their identification of the auditory segment can be modulated by visual (lipread)
information. 4 Using this general phenomenon, Bertelson et al. (2003) tested what would happen
3While the use of on-line here may seem non-standard, the basic idea is that they argue strongly
against models of speech perception with lexical feedback in (Norris et al., 2000) and therefore
need to make this differentiation between recognition and learning in order to maintain their strict
feedforward model. They believe that the learning happens over time and therefore is qualitatively
different than on-line feedback for recognition.

4The effect is more prevalent when using a non-labial segment for the visual presentation
accompanied by a labial auditory segment than the opposite scenario. This is briefly noted by
Bertelson et al. (2003), but because they were interested in the “aftereffects” (learning/recalibration
rather than initial response), they did not find it worrisome for their experiment.

6

when participants are presented ambiguous audio alongside unambiguous visual input using a
nine-step b∼d continuum and audio/video recordings of a male native Dutch speaker pronouncing
/aba/ and /ada/.

Each participant took a pre-test prior to the main portion of the experiment to find their individual
maximally ambiguous point ([?bd]) along the continuum. The audio token corresponding to [?bd]
and the two steps ±1 step from the token were used in the main experiment. [?bd] paired with
unambiguous visual cues were used as training items and all three steps were used in a post-training
phonetic categorization task (two alternative forced choice; “aba” or “ada”). The main experiment
consisted of 16 blocks. Each block consisted of eight exposure trials where participants were
presented with one of the two visual cues paired with [?bd], followed by six phonetic categorization
tasks (two each of the three steps determined by the pre-experiment). Eight of the blocks used the
/aba/ visual cue and the remaining eight used the /ada/ visual cue. Listeners were therefore exposed
to both training segments.

Figure 2.2: Results adapted from Bertelson et al. (2003)

Results from Bertelson et al.’s (2003) initial study on audio-visual recalibration are presented in
Figure 2.2. The “B” visual adaptor group corresponds to participants who saw unambiguous visual
input of a /b/ simultaneously with the ambiguous [?bd] audio segment. The “D” visual adaptor

7

group saw /d/ visual input, along with the ambiguous sound. Because the phonetic categorization
task consisted of the most ambiguous sound on a /b/∼/d/ continuum for each individual participant
and the steps ± 1 from that step on the continuum, the graph below is normalized across speakers
and does not constitute any one specific area of the continuum. Speakers who saw the /d/ visual
input were more likely to respond “d” afterwards and speakers who saw /b/ visual input were more
likely to respond with “b”.

The results presented demonstrate that the identification of ambiguous auditory input can be
biased by visual information. Bertelson et al. (2003) interpret these results fairly generally and
conclude that when cross modal biases occur, recalibration of some sort must occur.

2.3 Comparison of the two paradigms

While the two paradigms use different methods to disambiguate the ambiguity of auditory
stimuli, they appear at a cursory glance to be inducing a similar perceptual shift. In both cases,
when presented with an unambiguous anchor (either non-minimal pair lexical item or clear visual
presentation), listeners adapt their responses in the direction of the unambiguous anchor. Despite
this similarity, there are a number of dimensions in which the two strategies vary and lead to
skepticism about whether or not these are parts of the same general speech perception mechanism.
These differences are outlined in Table 2.1 and discussed further below.

Lexical Retuning

Audio-Visual Recalibration

Processing Strategy
Speech Type(s)
Stability
Presence of opposing phoneme
Experiment Design
Table 2.1: Comparison of lexical retuning and audio-visual recalibration paradigms

Clear or Ambiguous
Relatively short lasting

Relatively long lasting

Between Subjects

Within Subjects

Top-down
Ambiguous

Yes

Bottom-up

No

The first way to compare the two is with their processing strategies. While it can be argued that
the use of visual information/lip-reading is a bottom-up strategy, it’s clear that the use of lexical
information relies on top-down processing. This difference could be what leads to variation between
the two paradigms in the ability of their effects to generalize. There is also a strong difference

8

in what types of speech can elicit each effect. Lexical retuning can only work on ambiguous
speech. If clear speech is used then there is no need for any disambiguation strategy. On the other
hand, McGurk effects can alter the perception of clear speech (McGurk and MacDonald, 1976).
This suggests that visual information is potentially tapping into a different underlying mechanism,
potentially one that is autonomous. Further support for the autonomous view comes from Baart
and Vroomen (2010) who observed that the effects of audiovisual recalibration persisted even when
participants were simultaneously given tasks that put a strain on working memory. These findings
put audiovisual recalibration in a similar category as selective adaptation which has also shown to
be an automatic process that is unaffected by memory load (Samuel and Kat, 1998). Because Norris
et al. (2003) ran a second experiment to show that lexical retuning effects could not be explained by
selective adaptation, this gives additional substance to the hypothesis that audiovisual recalibration
and lexical retuning are not tapping into the same general perceptual mechanism. If the two effects
target unique spots along the perceptual stream then this would once again give credence to why
one allows for generalization and the other does not.

Another area in which empirical results for each paradigm have conflicted is stability or how
long each effect lasts. Vroomen et al. (2004) observed that the recalibration effect induced by
audio-visual information was relatively short lasting. After hearing 50 exposure stimuli, the effect
wore off after only 6 post-training tests. In the realm of lexical retuning, independent results suggest
that the effect was relatively long lasting. Kraljic and Samuel (2005) observed that it could last up
to 25 minutes while Eisner and McQueen (2006) found the effect to last for 12 hours regardless if
the participant slept or not. One could argue for an account of these results based on where along
the perceptual stream each effect is being integrated. The long-lasting nature of the lexical retuning
effect could be due to it affecting a higher-level representation, while the audio-visual information
may simply bias the recognition process. In other words, the visual information changes the way a
sound is perceived at the input, but the lexical information cannot alter the perception until later on
down stream after the high level information has been obtained.

Additionally, there is the learning aspect to consider. Lexical retuning experiments were

9

originally developed to see how speakers adapted to novel accents. Norris et al. (2003) argued
that higher level information could be used to provide feedback if learning was the ultimate goal.5
Someone pronouncing a word strangely may necessitate learning of the speaker’s idiosyncratic
mappings from surface to underlying forms, but conflicting audio and visual information does
not require this since it does not require the same type of mapping. Speculating on this may be
unnecessary, though, as there are other confounds between the previously described experiments
(such as silence following the training) that may also explain the difference in results.

There are a few general design differences between lexical retuning and audio-visual recalibra-
tion experiments as well. The presence or absence of the opposing phoneme in training conditions
is one of them. Lexical retuning experiments include both segments from the continuum while
audio-visual recalibration experiments have only the phoneme being tested presented in each spe-
cific training-testing block. Additionally, because the design of the audio-visual recalibration
experiments is blocked in a way that participants are exposed to both phonemes as the ambiguous
segment, this means that they are constantly reinterpreting the same ambiguous sound as both
discrete categories in a way that participants of lexical retuning experiments are not.

There also has been variation in the segment type and syllabic position used throughout the
two paradigms with only brief mentions throughout the literature suggesting that this could be
causing the difference in results. Audio-visual recalibration experiments present the same training
token throughout the entire block. Since they are only hearing the ambiguous auditory segment
in one environment, it is possible that participants may just consider the ambiguity to be unique
to that string. Due to the design of lexical retuning experiments, participants hear the ambiguous
segment in multiple words. The stimulus variation in lexical retuning may lead listeners to the
conclusion that this is a general principle of the speaker and not just an idiosyncratic pronunciation
of one string. If this is the case, then it may explain why lexical retuning experiments have shown
5As discussed previously in footnote 3, this was primarily due to their Merge model of speech
perception which was strictly feedforward (Norris et al., 2000) The learning aspect was added
to differentiate between different types of on-line feedback. They ultimately argue that on-line
feedback has no benefit to perception and is only used when learning is involved.

10

generalization effects in ways that audio-visual recalibration have not.

To test how lexical retuning and audio-visual recalibration performed head to head, Van Linden
and Vroomen (2007) created a series of five experiments where the same ambiguous segment [?tp]
was used in both tasks. In both cases, the ambiguous segment appeared in coda position. The five
experiments that were run can be broken down as follows: Experiment 1 resulted in both paradigms
having similar size effects; Experiment 2 found that the effect dissipated relatively quickly for both
paradigms; Experiment 3 observed that the presence of the opposing phoneme increased the size
of the effect for both paradigms; Experiment 4 found that a three minute silence period between
training and testing did nothing to affect the size/stability of the effect; Experiment 5 resulted in
listeners who were trained on both of the phoneme categories not performing any differently than
those who were trained on only one category.6

Their takeaway from all experiments is that lexical retuning and audiovisual recalibration show
similar aftereffects and therefore possibly represent the same general mechanism. This was taken
as the standard view and even echoed in Samuel and Kraljic’s (2009) general overview of the
phenomenon. The problem with assuming that the results from Van Linden and Vroomen’s (2007)
study indicate an overall similarity between the two perceptual effects is that they only tested certain
dimensions. As mentioned previously, there have been additional conflicting results in regards to
generalization within both paradigms. These must be addressed so that we can better evaluate the
overall nature of perceptual recalibration.

2.3.1 Generalization

Generalization in lexical retuning experiments has been shown in a few different ways. Going back
to the original experiment (Norris et al., 2003), one can make the argument that lexical retuning
has always had the effect of promoting generalization. In this experiment, the training word lists
6Some of these results are in conflict with prior results in the literature. Some possibilities for
this conflict may be syllabic position or consonant type. Due to the different labs working within
each paradigm, there was variation in the types of segments and syllabic position used for prior
comparisons. These new results from Van Linden and Vroomen (2007) may be observed due to
the better control of these two variables.

11

that were used contained a multitude of different vowel-fricative bigrams while the testing string
was always from an [Ef]∼[Es] continuum. Only one of the training words contained an [E] vowel.
Therefore, literally only one instance of the training stimuli matching the test stimuli is enough to
induce the lexical retuning perceptual shift, or listeners were able to learn a generalized pattern that
was independent of the environment of the ambiguous segment7.

Additionally, Jesse and McQueen (2011) observed that when listeners were trained with the
ambiguous segment in coda position, they transferred the effect when tested on segments from
a [fø]∼[sø] continuum. These results suggest a generalization to a position-independent repre-
sentation. Finally, generalization to a featural level has been shown for both stops (Kraljic and
Samuel, 2006) and fricatives (Durvasula and Nelson, 2018). In both instances, training and testing
environments were different, suggesting that the retuning effect was operating over a more abstract
representation than just auditory features.

In the realm of audio-visual recalibration the opposite has been found. Reinisch et al. (2014)
observed no evidence of generalization in a series of three experiments that suggested that both
training and testing environments needed to be identical in order to induce the perceptual re-
calibration. They interpret their results to mean that the recalibration is specific to phonemes,
acoustic cues, and the phonological context, but state that, “This conclusion rests, however, on the
assumption that the visually-guided recalibration reflects a general speech-perception mechanism
(an assumption empirically supported by Van Linden and Vroomen (2007)” (p. 104). The general
speech perception mechanism alluded to here is one in which audio-visual retuning is the same as
lexical retuning. A driving factor for this thesis is to see whether or not this assumption should
continue to be made. I will use the lexical retuning paradigm to pursue this further.

There are two issues that I am specifically going to look at relating to this assumption. Van Lin-
7A potentially more interesting question is whether or not listeners would be able to generalize
the pattern to an environment they didn’t hear despite hearing multiple environments in the training
stage. In other words, if they hear the ambiguous segment in multiple environments but not every
environment they’re aware of, would they generalize to the non-trained environments? Depending
on how listeners pattern, this could be a way to further pursue questions of representation and
possibly even learnability

12

den and Vroomen (2007) did not include any experiments that would test whether or not the two
paradigms resulted in generalization in their comparison study. To see if indeed the two paradigms
use the same general speech perception mechanism, one needs to establish that both of them result
in a similar degree of generalization under similar conditions. One way to make this comparison
is to use an experimental setup similar to Reinisch et al. (2014), but within the lexical retuning
paradigm. In their experiment, they strictly controlled the phonetic environments of the training
and testing conditions. No such study using the lexical retuning paradigm has explicitly done
the same thing. Furthermore, one possible reason that lexical retuning experiments have resulted
in more generalizability than audio-visual recalibration experiments is the increased amount of
stimulus variation present. The lack of such stimulus variation in audio-visual recalibration may
lead participants to view the ambiguous sound as idiosyncratic to the particular string they hear it
in. Probing if generalization is present in lexical retuning experiments in the absence of stimulus
variability would allow us to understand if such differences in stimulus variability could be the
source of the observed discrepancy between the two paradigms.

If we look outside of the perceptual learning paradigm, it has been shown that linguistic
generalizations require type experience and not token experience in order to be learned. Gerken
and Bollt (2008) present results that suggest that 9-month old infants are able to generalize a
constraint that says heavy syllables should be stressed to novel stimuli when presented with three
unique training items, but failed to do so when presented with one unique training item presented
multiple times. Denby et al. (2018) draw a similar conclusion using artificial language learning
experiments. They looked at listeners abilities to learn gradient phonotactic patterns and found that
contextual variability (type frequency) significantly affected learning while the number of times an
exemplar was repeated (token frequency) had no such effect. Both of these results taken together
suggest that in order to generalize linguistic patterns, type experience is needed.

Three experiments relating to these issues will be used to 1) more carefully test the generalizable
nature of lexical retuning and 2) see what effect the type vs. token experience has on it. Experiment 1
is primarily a replication of Norris et al. (2003) with a slight modification of the control group. This

13

is used to corroborate the idea that lexical retuning can be learned from many conditions in training
to one condition in testing. Experiment 2 is an attempt to replicate some of the training/testing
environments used in Reinisch et al. (2014), but within the lexical retuning paradigm.
In this
experiment, the training condition is limited to one vowel and is unique from the vowel used in the
testing condition. This is used to see if lexical retuning shows generalization effects when strictly
controlling the training and testing environment. In this sense it may give a stronger indication
at the overall generalizable nature of the lexical retuning effect. Experiment 3 uses a similarly
restricted training/testing environment as Experiment 2, but reduces the training stimuli set in
order to see how the absence of type variation affects generalization. This, even more closely than
Experiment 2, replicates the training and testing that participants underwent in Reinisch et al.’s
(2014) experiments. The results from all three experiments will give us a better idea about how
and why the generalization facts vary between lexical retuning and audio-visual recalibration.

14

CHAPTER 3

EXPERIMENT 1: REPRODUCING GENERALIZATION IN LEXICAL RETUNING

EXPERIMENTS

In Experiment 1 the goal is to confirm that the baseline generalization that is seen in Norris et al.’s
(2003) experiment can be replicated. This experiment looks at training conditions with multiple
unique vowels as well as variation in syllabic position. The testing block of the experiment has
the target fricative in onset position. Since it was shown that lexical retuning can transfer its effect
from coda to onset but not vice versa (Jesse and McQueen, 2011), this allows for the use of both
syllabic positions and therefore increases the set of potential words to use in the training stimuli.
One significant difference between Experiment 1 and the Norris et al. (2003) experiment is in
the control group. In their experiment, Norris et al. (2003) include nonce words containing an
ambiguous segment and normal words containing [f] and [s] in their control group’s LDT wordlist
and argue that because they do not form real words then the retuning effect will not occur since
there is no lexical information to tap into. Since this was never clearly confirmed, there is a
possibility that the presence of the ambiguous segment still had an influence on the control group
despite appearing in nonce words. To err on the side of caution, the control group for Experiment
1 contains no instances of an ambiguous segment, but rather just the clear [f] and [s] tokens and
the same filler words as the test group. This will hopefully control for any subtle effect of the
ambiguous segment for those in the control group.

The other primary departure from previous work using lexical retuning is the addition of a pre-
LDT phonetic categorization task while also keeping the post-LDT phonetic categorization task.
This allows for before-after comparisons to be made within both the control and test groups, as
well as the traditional ‘After”-“After” comparison that is typically used in lexical retuning analysis.
I also add a second between-group comparison which is the difference from before to after for both
groups and argue that this is a more insightful comparison.

15

3.1 Method

3.1.1 Participants

Seventy-three undergraduate students from Michigan State University (Mean Age = 21; 57 female,
4 unreported gender) received either course credit or a small monetary reimbursement ($10) for
participating in the study. All speakers identified as American English speakers and did not report
any hearing problems.

3.1.2 Design

The experiment consists of three blocks. First, participants will be given a phonetic categorization
task. They will hear various steps from a blended continuum of [f] and [s] segments that precedes
an [i] vowel. This will be used as a baseline measurement for analysis. A training phase follows in
the form of an auditory lexical decision task (LDT) which is used to introduce participants to the
ambiguous segment that will trigger the perceptual learning. The LDT is then followed by a second
phonetic categorization task, identical in form to the baseline test. The flow of the experiment is
visualized below in Figure 3.1.

Phonetic

Categorization

Lexical Decision

Task

Phonetic

Categorization

Figure 3.1: General Experiment Design

By adding the “Before” phonetic categorization block, it allows for both within- and between-
subjects analysis. Traditionally, lexical retuning experiments have looked at between-subjects
comparisons of post-LDT phonetic categorization tasks with no reference to a starting baseline
(Norris et al., 2003; Eisner and McQueen, 2005, 2006; Kraljic and Samuel, 2005, 2006; McQueen
et al., 2006b; Cutler et al., 2008; Sjerps and McQueen, 2010; Jesse and McQueen, 2011; Scharenborg
et al., 2011; Reinisch and Holt, 2014; Mitterer et al., 2011, 2013, 2016). In these experiments,

16

it was tacitly assumed that each group was starting from the same baseline.
Including both a
“Before” and “After” categorization removes this assumption and better controls for any sampling
idiosyncrasies. Each participant is assigned to one of two groups: FISItest, FISIcontrol, which will
be discussed in more detail below.

3.1.3 Materials

The lexical decision task uses a 150 word list containing 75 English words and 75 phonotactically
licit English nonce words. Thirty-four of the English words are training items, while the remaining
41 and all 75 nonce words are used as filler. All 34 training tokens contain either an [f] or an [s]
(each segment distributed equally) and do not form a minimal pair when replaced with the opposing
segment (e.g., beef, sing). All the training words used were monosyllabic and contained one of
nine different vowels. Nine of the 17 words for each segment had the crucial segment in the onset
(word initial) position of the word. Test words were controlled for frequency using the Subtlexus
corpus (Brysbaert and New, 2009). While the [f]-words were less frequent (12.85/million) than the
[s]-words (20.77/million), no statistically significant difference between the log frequencies was
found [t(28.3) = 0.48, p = 0.64]. The remaining 114 filler words contain no instances of [f s v z].
The real word fillers range in syllable count from 1-3 while the nonce word fillers range from 1-4.
The 150 words for the LDT were spoken by a female native American English speaker from
Michigan. Each word was read aloud into a Logitech 980186-0403 microphone (frequency repsonse
100Hz–16kHz; -67dBV/ubar, -47dBV/Pascal ±-4dB) in a quiet room and recorded directly into
Praat (Boersma and Weenink, 2016) at a sampling rate of 44.1 KHz. The speaker also recorded
tokens of [fi] and [si]. These were used to make the continuum: fi∼si. The creation process was
done as follows: first, the tokens of [fi] and [si] were manually annotated in Praat (Boersma and
Weenink, 2016) to mark the fricative and vowel portions of the token. From here, the entire process
was automated using scripting. To make each continuum, equal amounts of the fricative portion
of each token was spliced out (165 ms; normalized to 50dB). The amplitudes of the f/s pairs were
then blended in 41 equal steps (e.g., step 1 was 100% [f] and 0% [s]; step 2 was 97.5% [f] and

17

2.5% [s]; . . . step 41 was 0% [f] and 100% [s]). The continuum was then re-spliced back onto the
vowel portion from the [fi]. It is well known that that formant transition information in the vowel
acoustics is a cue for place of articulation (Delattre et al., 1955, 1962), therefore this method may
introduce a slight bias towards “f” responses, but in previous studies it has been shown to elicit
normal response functions (Norris et al., 2003; McQueen et al., 2006b; Durvasula and Nelson,
2018)

The continuum was then used in a pre-test to find the most ambiguous step. Of the 41 total
steps, 14 were chosen to use for phonetic categorization. These included steps 1 & 41 which were
the 100% [f] and 100% [s] tokens, respectively, as well as every other step between steps 7-29.
This meant that the majority of the tokens that participants categorized were from the ambiguous
portion of each continuum. Thirteen American English speakers (Mean Age = 20.9 years; 8 female,
1 unreported gender) from Michigan State University separate from those in the main experiment
participated in the pre-test for partial class credit.

Figure 3.2: Categorization results for the fi∼si continuum pre-test

Participants were tested using PsychoPy (Peirce et al., 2019). Each participant heard the 14
steps from the continuum four times each in random running order. A two alternative forced choice
paradigm was used. After hearing a sound, the participant was instructed to use a computer mouse
to indicate whether the sound they had just heard contained either “f” or “s”. The mean response

18

for each step was calculated to create an identification response function. This resulted in the
sigmoidal function expected for consonant segments (Liberman et al., 1957). The point closest
to which the response function intersected with the 50% response rate was then interpreted as the
most ambiguous point on the continuum. For the fi∼si continuum, the 50% response rate lay in
between steps 17 and 19 and therefore the fricative portion of step 18 was used to create [? f s−i].
This was then used as a replacement for all of the [f] sounds in the LDT for the FISItest training
group.

Praat (Boersma and Weenink, 2016) was once again used to manipulate audio stimuli, this time
to create an altered versions of the LDT wordlist. The FISItest version of the list had all of the [f]
tokens in the LDT replaced with [? f s−i]. The FISIcontrol wordlist was unaltered and therefore had
no instances of an ambiguous token. To make the altered wordlist, the fricative portion of each word
containing [f] was manually annotated and marked at points of zero crossing. Scripting was then
used to automatically remove the original frication portion of the word and replace it with [? f s−i].
The returned wordlist was then manually checked for naturalness based on author-intuition.

3.1.4 Procedure

The main experiment was also performed using PsychoPy (Peirce et al., 2019). Up to 12 participants
were tested in the lab simultaneously. Prior to the experiment, participants were verbally instructed
that they will be doing three tasks on the computer using various response mechanisms - the phonetic
categorization (first and third tasks) requires the clicking of a mouse while the lexical decision task
(second task) will require them to use the keyboard to give a yes/no response. Participants were
also verbally instructed to answer as quickly and accurately as possible and to remain seated and
quiet until everyone in the room had completed all three tasks. Participants wore over the ear
headphones (Plantronics .Audio 355; 20Hz-20kHz response) throughout the entire experiment and
specific instructions were visually presented to them on the screen before and during each task.

The phonetic categorization tasks were the same format as the pre-test described above: each
iteration contained the same 14 set of steps played four time each in random running order.

19

Participants were randomly placed into one of the two groups. Both the FISItest and FISIcontrol
groups were played the same fi∼si continuum. As before, participants were given a two alternative
forced choice task (“f” or “s”) and instructed to use a computer mouse to click which sound they
hear.

Figure 3.3: Phonetic categorization Psychopy screen and instructions

Upon completion of the first phonetic categorization task, participants underwent the lexical
decision task. They were instructed through PsychoPy (Peirce et al., 2019) that they will now be
hearing a series of words, one at a time, and have to decided whether or not the word they heard
is a real English word. Additionally, they were instructed now to use the computer keyboard’s ‘a’
and ‘l’ keys to answer “no” or “yes” to the question, “Is this an English word?” An ‘a’ response
corresponded to “no” and an ‘l’ response corresponded to “yes”. This information was constantly
on screen as reference for participants.

Figure 3.4: LDT Psychopy screen and instructions

If a participant did not respond within 3.5 seconds of the onset of the sound, a new sound

20

was played and no response was recorded. Each word was presented in random running order.
Depending on the group that an individual participant is assigned to, they were played the corre-
sponding LDT list as described above. After completion of the LDT, participants were given the
same phonetic categorization task that they did prior to it. The only difference from the previous
phonetic categorization is that a new random running order of the stimuli was used.

3.2 Results

In the original lexical retuning experiments, Norris et al. (2003) set a criteria of 50% accuracy
rate of the ambiguous segment in the LDT in order to keep a participants’ data for analysis. This
was to ensure that participants were recognizing it as a quality exemplar of whatever segment it
was replacing. In this experiment, the criteria was expanded. Participants were required to score
50% or higher in accuracy for both of the training segments ([s] and either [f] or [? f s]), as well as
have an overall score of 50% or higher. Two participants ended up being removed from analysis.
Overall, both groups had accuracy rates of 90%.

The within-subjects analysis was performed first. A one-tailed, paired t-test showed that there
was a statistically significant decrease in alveolar responses from the “Before” phonetic categoriza-
tion task to the “After” phonetic categorization task for the FISItest group [MeanDiffBefore-After = -
0.0873, t(35)=-4.8, p<.001]. This suggests that participants in the FISItest group altered their
categorization of the fi∼si after hearing [? f s] in words normally containing [f] during the LDT. A
one-tailed, paired t-test showed that there was also a statistically significant decrease in alveolar
responses from the “Before” phonetic categorization task to the “After” phonetic categorization task
for the FISIcontrol group [MeanDiffBefore-After = -0.0393, t(34)=-2.46, p<.01]. This is unexpected
as there was no alteration to this group’s LDT. One explanation for this result is the fact that the
vowel portion of all of the phonetic categorization tokens was taken from a token of [fi]. It was
predicted beforehand that if there was any bias, that it would be in the direction of “f” responses.
These results seem to bear out that sentiment. The categorization functions for both of these
comparisons can be seen in the upper left and upper right graphs of Figure 3.5.

21

Figure 3.5: Categorization results for Experiment 1. Upper left is FISIcontrol before vs after
comparison; Upper right is FISItest before vs after comparison; Bottom left is the after responses of
FISIcontrol vs the after responses FISItest; Bottom right is the difference between after and before
responses between the two groups.

Since both of the within-subjects analyses proved to be significant, between-subjects tests were
run. The first test was between the responses of the “After” phonetic categorization task for both
groups. This is traditionally the comparison that is made in lexical retuning experiments. A one-
tailed Welch test showed that there was a statistically significant difference in alveolar responses
between the FISItest and FISIcontrol groups [MeanDiffTest-Control = 0.053, t(67.9)=1.77, p=0.04].1
This result is slightly surprising as previously it has been observed that there was no difference
between control and test groups (Norris et al., 2003). The design of this experiment also allows
for a second between-subjects analysis. While the “After”-“After” comparison gives insight into
1Welch tests were used for the between subjects analysis since they do not assume homogeneity

of variance between samples.

22

what participants are doing based on the experimental group they were assigned to, it does not
take into account that different samples of the population may have different starting points for the
phonetic categorization. The difference between the “After” results and the “Before” results can
be taken, for both the test group and the control group, and used as comparison between the two
groups. A one-tailed Welch test showed that there was also a statistically significant difference in
alveolar responses between the two groups [MeanDiffTest-Control = 0.048, t(68.1)=1.99, p=0.026].
The functions for both of these comparisons are in the bottom left and bottom right graphs of Figure
3.5.

3.3 Discussion

The results from Experiment 1 show that listeners give an increased number of “f” responses to
a fi∼si continuum after hearing [? f s−i] in place of [f] tokens in an LDT. Since the [f] tokens were
in both coda and onset position, and in a multitude of different vowel environments, it is possible
that a generalization of the perceptual learning has occurred. But a statistically significant shift
in the same direction was shown for the FISIcontrol group as well. This indicates that there is
some confound causing a shift from the “Before” phonetic categorization to the “After” phonetic
categorization. It is possible that this is due to the formant transition information present in the
fi∼si continuum. For this reason, a comparison between the groups was necessary.

The “After”-“After” comparison that has typically been made when analyzing these types of
experiments also showed a statistically significant difference in the expected direction (fewer “f”
responses for the FISItest group). When looking at the results in the lower left graph of Figure 3.5,
it can be seen that much of the difference between the two groups appears to be in the last three
steps of the continuum. We expect the variation to be more present in the middle of the continuum.
An inspection of the lower right graph of Figure 3.5 shows this. This graph shows the difference
in responses within each group from the “Before” to the “After” phonetic categorization tasks.
There is steady separation between the two groups starting at step 11 and going all the way through
the rest of the continuum. Step 13 is the point in all the other graphs where the response rates

23

are in between 0-1 and therefore where we would predict less stable responses (i.e., this is where
we should see the effects of the lexical retuning). The statistics confirm that there is a significant
difference in the expected direction when comparing the two groups this way.

The overall takeaway from Experiment 1 is that there is some evidence for generalization within
lexical retuning experiments. Participants were trained with the ambiguous token in a multitude
of different vowel environments, but tested in only one. That being said, the training set of words
for the LDT did include two words where [? f s] was in coda position after [I] and one time where
it followed [i]. It also included two words where [? f s] was in onset position before [I]. So 5/17
words in the training set contained the same vowel (or one that is very similar) as in phonetic
categorization testing. Therefore it could be argued that the perceptual learning followed from this
smaller subset. To make the claim that it is in fact generalization to new contexts, stricter training
and testing conditions are required. In Experiment 2, I will further probe the generalizable nature
of lexical retuning by seeing whether the learning effect can generalize to a context that is not
present in the training block.

24

CHAPTER 4

EXPERIMENT 2: GENERALIZATION WITH NON-IDENTICAL TRAINING AND

TESTING CONDITIONS

Experiment 1 replicated previous results, and showed that the lexical retuning effect could be learned
from many training environments and be used in one testing environment. It is hypothesized that
this is because listeners learn a pattern and generalize it; however, there was some overlap in
training/testing vowel environments, which places the generalization claim under dispute.
It is
possible that the subset of training words where the vowel context matched the testing condition
may have been enough to induce the shift in categorization seen in Experiment 1.

Recall that context dependency was more strictly tested using the audio-visual recalibration
paradigm. There it was shown that the perceptual learning effect was context dependent. One of the
findings from Reinisch et al.’s (2014) study is that when participants are trained in the environment
of an [i] vowel, they are unable to generalize the effect when tested on an [a] vowel. (i.e., /aba/
or /ada/ for training did not induce the perceptual recalibration shift for an [ibi]∼[idi] continuum).
There has been no study using the lexical retuning paradigm that has imposed as strict training and
testing environmental restriction. For this reason, Experiment 2 uses the lexical retuning paradigm
to see whether participants trained with an ambiguous segment in the environment of an [i] or [I]
vowel will show lexical retuning effects when tested on a [fa]∼[sa] continuum. Since there is no
overlap in vowel context between the training and testing conditions, a shift in the categorization by
the test group would more explicitly support generalization of the effect over a context independent
representation.

4.1 Method

4.1.1 Participants

Seventy-eight undergraduate students from Michigan State University (Mean Age = 19.5; 47
female,1 unreported gender) received either course credit or a small monetary reimbursement ($10)

25

for participating in the study. All speakers identified as American English speakers and did not
report any hearing problems.

4.1.2 Design

The overall design for Experiment 2 was identical to that in Experiment 1. Experiment 2 has
the same three part set up of pre-training phonetic categorization, training vis-à-vis an auditory
lexical decision task, and post-training phonetic categorization. There are some more, however: the
phonetic categorization continuum is now a fa∼sa continuum and the training words in the lexical
decision task are now limited to words that contain an [i] or [I] vowel next to the target fricatives.
These will be more explicitly outlined below. For this experiment, participants were assigned to
one of two groups: FASAtest or FASAcontrol.

4.1.3 Materials

An LDT was used with 150 words split into 75 English words and 75 phonotactically licit English
nonce words. The 116 filler words (41 real/75 nonce) from Experiment 1 were once again used
in the Experiment 2 list. The 34 training items are new. Since the goal of this experiment is to
observe how the lexical retuning effect behaves when the phonological environment for the training
and testing vary, all the training items have the crucial segments that are situated next to an [i] or
[I] vowel. Expanding the criteria beyond just the [i] vowel was necessary in order to obtain a large
enough training sample. The choice of [I] was due to it being the most phonetically similar segment
in American English (Hillenbrand et al., 1995). Phonologically, the [i]–[I] distinction is typically
embodied over a contrast of ATR or length. In this study, the goal is to use phonological distance
as a comparison and therefore only the height and front-back features are of interest. Therefore,
we can use the [A] vowel for contrast since it is maximally different from [i I] on both of those
dimensions.

All 34 training tokens once again contain either an [f] or an [s] (each segment distributed equally)
and do not form a minimal pair when replaced with the opposing segment. For both segments,

26

13 of the tokens have the crucial segment in onset position. It appears in coda position for the
remaining four. Each segment has 6 disyllabic tokens and 11 monosyllabic tokens. All disyllabic
tokens have the crucial segment in onset position. Eight of the words for each segment contain [I]
and are all monosyllabic (6 onset, 2 coda). The training tokens are controlled for frequency using
the Subtlexus corpus Brysbaert and New (2009). Due to the limited number of words matching
the criteria, there is a mismatch in frequencies. The [s]-words are more frequent (47.56/million)
than the [f]-words (19.39/million), but no statistically significant difference is found between the
log frequencies of the two groups [t(28.67) = -0.35, p = 0.73].

The new 34 words for the LDT were spoken by the same female native American English
speaker from Michigan, as in Experiment 1. Each word was read aloud into a Logitech 980186-
0403 microphone (frequency repsonse 100Hz–16kHz; -67dBV/ubar, -47dBV/Pascal ±-4dB) in
a quiet room and recorded directly into Praat (Boersma and Weenink, 2016) at a sampling rate
of 44.1 KHz. The speaker also recorded tokens of [fa], and [sa] to make the fa∼sa continuum.
Stimulus creation was done the same way as Experiment 1. The created fa∼sa continuum was
used in a new pre-test to find the most ambiguous step. Eight American English speakers (Mean
Age = 20.5 years; 2 female) from Michigan State University separate from those in the main
experiment participated in the pre-test for partial class credit. Participants heard the same 14 steps
as in Experiment 1 (1,7,9,11,13,15,17,19,21,23,25,27,29,41) The mean response for each step was
once again calculated to create the identification response function. and resulted in the expected
sigmoidal function. Figure 4.1 shows these results below.

For the fa∼sa continuum, step 16 was chosen as the most ambiguous segment. According to
the pre-test results, the 50% response rate should be between steps 13 and 15, but due to the sharp
rise and then drop between steps 13 and 17, multiple steps were tested against personal intuition in
order to confirm that the segment chosen was maximally ambiguous. Steps 14 and 16 were tested
due to their placement on the resulting categorical function from the pre-test. If we look at the area
between steps 13 and 17 without step 15, then these both appear in a region that could feasibly be
the 50% response rate. Step 18 was also tested as a comparable due to the result from the fi∼si

27

continuum. Ultimately, step 16 was chosen to create the [? f s−a] sound that would replace all of
the [f] sounds in the LDT for the FASAtest group.

Figure 4.1: Categorization results for the fa∼sa continuum pre-test

Praat (Boersma and Weenink, 2016) was once again used to create an altered versions of the LDT
wordlist. The FASAtest version of the list had all of the [f] tokens in the LDT replaced with [? f s−a].
The FISIcontrol wordlist was unaltered and therefore had no instances of an ambiguous token. To
make the altered wordlist, the same method was used as Experiment 1. The resulting wordlist was
then manually checked for naturalness based on author-intuition. While it is potentially worrisome
since it has been shown that identification of the vowel following a fricative can be correctly made
from acoustic information in the frication portion of the signal (Yeni-Komshian and Soli, 1981;
Soli, 1981; McMurray and Jongman, 2016), previous experiments have used a similar method with
no noticeable problem (Norris et al., 2003; Eisner and McQueen, 2005, 2006; McQueen et al.,
2006a,b; Durvasula and Nelson, 2018).

4.1.4 Procedure

The general procedure is the same as outlined in Experiment 1 above. The few differences are
outlined in the following paragraph. Participants were randomly assigned into one of two groups:
FASAtest or FASAcontrol. For the phonetic categorization tasks, both groups categorized the fa∼sa

28

continuum. For the LDT, participants heard the list that corresponded to their assigned group
as described above. Psychopy (Peirce et al., 2019) was once again used and all other procedural
methods were exactly the same as Experiment 1.

4.2 Results

The same criteria as Experiment 1 was used to exclude any participants from analysis. Three par-
ticipants failed to identify target segments accurately and were therefore removed. The FISIcontrol
group had an overall response accuracy rate of 91%, while the FASAtest group’s accuracy was
lower at 87%.

A within-subjects analysis was again performed first and a one-tailed, paired t-test showed that
there was once again a statistically significant decrease in alveolar responses from the “Before”
phonetic categorization task to the “After” phonetic categorization task for the FASAtest group
[MeanDiffBefore-After = -0.065, t(27)=-3.19, p<.01]. This suggests that participants in this group
altered their categorization of the fa∼sa continuum after hearing [? f s−a] in the context of [i] and
[I] vowels in words that normally contained [f] during the lexical decision task. Unlike Experiment
1, the control group in this experiment did not show a shift towards the labiodental side of the
continuum. A one-tailed, paired t-test showed that there was no decrease in alveolar responses
from the “Before” phonetic categorization task to the “After” phonetic categorization task for the
FASAtest group [MeanDiffBefore-After = -0.03, t(39)=-1.29, p=0.10]. The categorization functions
for both of these comparisons can be seen in the upper left and upper right graphs of Figure 4.2.

Looking at the between-subjects comparisons, the standard “After”-“After” comparison for
lexical retuning experiments resulted in no difference between the groups. A one-tailed Welch test
showed no significant difference between the “After” responses of each group [MeanDiffTest-Control = -
0.04, t(76)=-1.07, p=0.87]. The final comparison to be made is that of the difference between the
two groups. This likely best captures the effect of training on the classification of a continuum as
it has a baseline to compare against. It additionally allows for a normalization of the two groups.
A one-tailed Welch test showed that there was no statistically significant difference in alveolar

29

Figure 4.2: Categorization results for Experiment 2. Upper left is FASAcontrol before vs after
comparison; Upper right is FASAtest before vs after comparison; Bottom left is the after responses
of FASAcontrol vs the after responses FASAtest; Bottom right is the difference between after and
before responses between the two groups.

responses between the two groups when testing the entire continuum [MeanDiffTest-Control = 0.07,
t(55.3)=2.34, p=0.011]. Using the entire continuum for this analysis is rather conservative as there
is no expected change at the tails. A priori, a decision was made to also run comparisons using the
most ambiguous step and the two steps surrounding it (±1).1. A one-tailed Welch test showed that
there was a statistically significant difference in the difference in alveolar responses between the
two groups when testing this subset of the continuum [MeanDiffTest-Control = 0.11, t(71.7)=1.97,
p=0.026]. This suggests that the FASAtest group gave more labiodental responses after the lexical
1Recall that step 16 was determined to be the most ambiguous step in the pre-test and used as
[? f s−a]. Since step 16 was not used in the main experiment, a separate analysis on the “Before”
responses of the FASAcontrol group was used. This showed that step 17 was the most ambiguous,
so steps 15,17,19 were used

30

decision task and therefore were able to generalize the learning effect from the context of high-front
vowels to low-back vowels. These results are shown in the bottom left and bottom right graphs of
Figure 4.2

4.3 Discussion

Experiment 2 shows that lexically-induced perceptual learning persisted under strict training and
testing conditions. These results expand on those from Experiment 1 and show that generalization
occurs even when the training and testing environments overlap. Listeners who heard [? f s−a] in
words that normally contained [f] during the LDT were more likely to respond “f” when categorizing
a fa∼s continuum afterwards. Unlike Experiment 1, the within-subjects comparisons differed from
group to group. The FASAtest group showed a significant difference in responses from before to
after in the expected direction. In Experiment 1, both groups showed this difference and it was
thought that there was a confounding effect, possibly from the formant transitions cues present in
the ambiguous segment. The lack of difference here could be a result of using a different ambiguous
token and continuum than the ones from Experiment 1. The vowel used in the fA∼sA continuum
may have weaker formant transition cues than the fi∼si vowel. This is not to suggest that the type
of vowel used ([i] or [A]) is the locus of variation, but rather a byproduct of the two sets of stimuli
being created at separate times. It is possible that during the construction process, the two vowel
tokens were spliced at different time periods within the full [fi] or [fA] tokens such that they included
different amounts of formant transition information. Ultimately, this variation between the control
groups before-after responses from Experiment 1 to Experiment 2 is not super important as it gets
washed out when looking at the difference of differences between the control and test groups.

The between group comparison showed the crucial difference. If we were to limit the analysis
to just the ‘After”-“After” comparison, as has previously been done in both lexical retuning and
audio-visual recalibration experiments, then it would appear as if the effect is no longer able to
generalize when we strictly control the training and testing conditions. This matches what was
found in Reinisch et al.’s (2014) study and adds a new dimension to Van Linden and Vroomen’s

31

(2007) story that these two paradigms tap into the same general perception mechanism. The
comparison developed in Experiment 1 of the within group difference from before to after shows
its importance here. Visually, it is very noticeable in the bottom right graph of Figure 4.2. There is
clear separation in the before to after results for both groups and it once again happens in exactly
the predicted section of the continuum.

This results is important because it not only has theoretical importance in regards to the
generalizable nature of lexical retuning, but also methodological and analytical importance. The
addition of the “Before” phonetic categorization in the experimental design allowed for better
control of the sampling populations and subsequently a better analysis by comparing the amount
of change between groups after training rather than just their responses after training. In the future
when performing perceptual learning experiments, this comparison should allow for clearer insights
into how listeners are changing their categorization habits in response to the training they receive.
On the flip side, it is possible that previous experiments have missed certain insights by not making
this comparison. Overall, the results from Experiment 2 confirm that lexically-induced perceptual
learning (i.e., lexical retuning) is able to generalize to new contexts even when the training context
is strictly reduced. Comparing this to the results presented in Reinisch et al. (2014), it suggests
that lexical information is privileged in a way that auditory information is not (at least when it
comes to perceptual learning). There is still an alternative explanation for the difference in results.
The current lexical retuning experiment was designed in a way to provide type experience to
listeners, while audio-visual recalibration experiments provide token experience. In order to fully
comprehend whether the generalizable nature is domain specific or not, the type/token distinction
needs to be further tested. Experiment 3 will do so.

32

EXPERIMENT 3: REDUCING STIMULUS VARIATION IN LEXICAL RETUNING

CHAPTER 5

One area that has not been addressed in the literature is the stimulus variation within each paradigm.
Audio-visual recalibration experiments present the same string (typically some sort of VCV nonce
word) continuously in order to induce the perceptual shift. Lexical retuning experiments present
multiple, unique words throughout the LDT. In the former case, there is no variation, while the
latter includes within-experiment stimulus variation. This experiment will test what happens if
you remove the within-experiment stimulus variation from a lexical retuning experiment. Like
Experiment 2, it will test whether training in the context of an [i] vowel will lead to retuning effects
in a fa∼sa continuum. The difference here is that now instead of 17 different training words, the
LDT will contain 1 training word that gets repeated 17 times.

5.1 Method

5.1.1 Participants

Seventy-two undergraduate students from Michigan State University (Mean Age = 20.4; 54 female,
4 unreported gender) received either course credit or a small monetary reimbursement ($10) for
participating in the study. All speakers identified as American English speakers and did not report
any hearing problems.

5.1.2 Design

The overall design for Experiment 3 is identical to that in Experiment 2. Experiment 3 has the
same three part set up of pre-training phonetic categorization, training vis-à-vis an auditory LDT,
and post-training phonetic categorization. The only change made is to the LDT and will be
outlined below. For this experiment, participants were assigned to one of two groups: FASA2test
or FASA2control.

33

5.1.3 Materials

The LDT for Experiment 3 uses a subset of the stimuli used in Experiment 2. The overall number
of unique words used in the LDT for Experiment 3 is reduced to eight. Of the eight words, two are
training items and the remaining six are filler. Two of the filler items are real English words and the
other four are phonotactically licit English nonce words. For the training items, the [f]-word is more
frequent (logFreq=3.21) than the [s]-word is (logFreq=2.64). Both training items are disyllabic and
have the crucial segment in the onset position (word initial). Putting the training segment in onset
position means there is a match for position between training and testing and therefore predicts a
stronger likelihood that the retuning will occur. Using disyllabic words gives more information
to the listener to identify it as a real English word and therefore gives a better chance that the
ambiguous token will be recognized as a normal token of [f].

The filler items are all either mono- or di-syllabic. Two forms of the list were created. The
FASA2test list was sampled from the FASA LDT list from Experiment 1 and therefore had the word
containing [f] replaced with [? f s−a]. The FASA2control list was identical to the FASA2test list, but
it had the normal pronunciation for all words, including the word containing [f]. Therefore, the
only difference between the two lists was in the single [f]-containing training token. The materials
for the phonetic categorization tests are identical to the ones used in Experiment 2. Since the LDT
materials are a subset of the ones created in Experiment 2, it was unnecessary to run a pre-test to
find the most ambiguous segment to use for replacement in the LDT training words.

5.1.4 Procedure

The general procedure is the same as previous experiments. The few differences are outlined in the
following paragraph. Participants were randomly assigned into one of two groups: FASA2test or
FASA2control. For the phonetic categorization tasks, both groups categorized the fa∼sa continuum.
For the LDT, participants heard the list that corresponded to their assigned group. Since the LDT
lists for this experiment only contains 8 total words, each word was now presented 17 times each.
17 is chosen because that is the number of unique training words containing [f] or [s] in Experiment

34

1. Participants therefore heard 136 tokens in random running order during the lexical decision task.
While the overall number of words participants heard was reduced by 14, they did hear the same
number of training tokens between Experiment 2 and Experiment 3. Psychopy (Peirce et al., 2019)
was once again used and all other procedural methods were the same as Experiments 1 and 2.

5.2 Results

The same criteria as in the previous experiments was used to exclude any participants from
analysis. Five participants failed to identify target segments accurately and were therefore re-
moved. Both groups had lower overall accuracy rates than the previous experiments (88% for
FASA2controland 85% for FASA2test), but this was brought down primarily by the nonce words.
For the training words containing the target fricatives, both groups had greater than 95% accuracy.
Both groups showed no difference in alveolar responses from “Before” to “After”. A one-
tailed, paired t-test showed that there was no statistically significant decrease in alveolar responses
from the “Before” phonetic categorization task to the “After” phonetic categorization task for the
FASA2test group [MeanDiffBefore-After = -0.032, t(31)=-1.25, p=0.11] and for the FASA2control
group [MeanDiffBefore-After = -0.013, t(34)=-1.21, p=0.12]. This suggests that in both cases there
was no observable change in alveolar responses after the lexical decision task for either group. What
is of note here is the fact that the FASA2test group did not show a statistically significant change
despite hearing the same overall number of [? f s−a] tokens (17) as the test group in Experiment 2.
The categorization functions for both of these comparisons can be seen in the upper left and upper
right graphs of Figure 5.1.

Looking at the between-subjects comparisons, the ‘After”-“After” results do show a shift in the
expected direction, but not enough to be statistically different between the groups. A one-tailed
Welch test showed that there was no statistically significant difference in the percentage of alveolar
responses between the FASA2test and FASA2control groups in the “After” phonetic categorization
task [MeanDiffTest-Control = 0.064, t(61.5)=1.47, p=0.07]. As discussed in previous chapters, the
difference comparison is likely the better comparison to make. A one-tailed Welch test showed that

35

Figure 5.1: Categorization results for Experiment 3. Upper left is FASA2control before vs after
comparison; Upper right is FASA2test before vs after comparison; Bottom left is the after responses
of FASA2control vs the after responses FASA2test; Bottom right is the difference between after and
before responses between the two groups.

there was no statistically significant difference in the difference of alveolar responses from “Before”
to “After” between the two groups [MeanDiffTest-Control = 0.02, t(41.1)=0.702, p=0.24]. This result
seems to corroborate the idea that the lexical retuning effect did not occur for the participants in
the FASA2test group. The between-subjects results are shown in the bottom left and bottom right
graphs of Figure 5.1.

5.3 Discussion

The results from Experiment 3 question whether the locus of variation in the generalizabil-
ity of perceptual learning effects is the modality in which listeners receive their disambiguating
information. It was hypothesized that because lexical information is considered to be high-level

36

information, that it would allow for generalization due to tapping into the representational structure
of the segments. The low-level visual information not tapping into the representational structure
was subsequently proposed for why Reinisch et al. (2014) observed no generalization effects using
an audio-visual recalibration paradigm. This experiment brings into question that hypothesis.

In Experiment 2, it was observed that listeners could generalize to a new context in the test-
ing condition when the training condition contained one unique vowel environment presented in
multiple unique words. Experiment 3 kept the same testing condition and the same single vowel
environment for the training condition, but repeated the same word multiple times. In other words,
it traded type experience for token experience to better match the conditions used in audio-visual
recalibration experiments. The results suggest listeners were no longer able to generalize the lexical
retuning effect, despite hearing the same number of raw tokens in the training condition for both
experiments. This suggests a possible reanalysis of why Reinisch et al. (2014) were unable to find
any type of generalization effects. It may not be the case that visually-guided perceptual learning
is specific to phonemic contrast, cues, and contexts as they claim (Reinisch et al., 2014, p. 102),
but rather that generalization of the effect requires type experience. Because the experiments were
designed in a way that did not allow for type experience, it is now unclear as to what is the exact
cause of the lack of generalization.

That is not to say that there still may not be a difference in generalizability between the two
paradigms. What this experiment has shown, in conjunction with the previous experiments, is
that type experience is necessary to learn and generalize the perceptual learning effect in lexical
retuning experiments. This may just be a general property of learning linguistics generalizations
as it has also shown to be beneficial in teaching infants stress constraints (Gerken and Bollt, 2008)
as well as in phonotactic learning experiments (Denby et al., 2018). It is therefore necessary to
see how type experience affects generalization in audio-visual recalibration experiments in order
to confirm that the difference between it and the lexical retuning paradigm can be explained by this
general learning principle. Future experiments beyond the scope of this thesis may be beneficial in
this regard.

37

There are two potential outcomes for this hypothetical experiment. On one hand, if an
audiovisual-recalibration experiment is designed to allow for type variation in the training stimuli,
and no generalization to new environments is found, this would suggest that it is the modality of
the disambiguating information that is the cause of the previously observed variation in results.
In other words, lexical information leads to generalization while visual information does not. If
generalization is observed in this new experiment, it would suggest that the two disambiguating
modalities are being treated equally and the previously observed differences are likely due to the
type/token distinction in the training stimuli. Regardless of the outcome, the results should set an
important precedence for further experiments using either of the perceptual learning paradigms.

38

CHAPTER 6

DISCUSSION AND CONCLUSION

6.1 Summary of Results

This thesis investigated the generalizable nature of perceptual learning in speech using the
lexical retuning paradigm developed by Norris et al. (2003). More specifically, it sought out why
the two paradigms housed under the perceptual learning tent (lexical retuning and audio-visual
recalibration (Bertelson et al., 2003)) were showing different results in relation to generalization
despite earlier claims that both were part of the same general perception mechanism (Van Linden
and Vroomen, 2007). Three new experiments were run to confirm that lexical retuning allowed for
generalization to new contexts both under broad (Experiment 1) and more restricted (Experiment
2) training conditions, as well as what effect type and token experience have on generalization
(Experiment 3).

The results from Experiment 1 showed that generalization of the perceptual learning effect in
lexical retuning experiments was possible when using a multitude of different training environ-
ments and testing on an environment that listeners only heard once throughout the training phase.
Listeners were more likely to respond “f” on a phonetic categorization task of a fi∼si continuum
after hearing an ambiguous segment [? f s−i] in place of [f] in words during a lexical decision task
than listeners who heard normal [f] in the same words. Experiment 1 also presented new method-
ological and analytical results for perceptual learning studies. Previously, comparisons of phonetic
categorization responses after exposure to an ambiguous segment within both paradigms have been
made between two groups. This tacitly assumes a homogenized beginning state for all individuals.
The addition of a “Before” phonetic categorization task to the experiment design allowed for a
comparison of differences from before to after between the two groups. In Experiment 1, it was
this comparison of the differences that most explicitly displayed the effect that the lexically-guided
training provided by the ambiguous segment had on the test group. Experiment 1 was therefore

39

important not only to corroborate past results and give more supporting evidence for generalization
within lexical retuning, but also to show that methodologically there are places in which we can
improve the experimental/analytical methods used within these perceptual learning paradigms.

Experiment 2 observed that lexical retuning could still lead to generalization, even in more
strictly controlled training and testing conditions. Unlike Experiment 1 where the set of training
words contained multiple unique vowel environments, Experiment 2’s training set only included
words that contained either an [i] or an [I] vowel. Listeners were then tested on a fa∼s continuum.
Once again, listeners were more likely to respond “f” on a phonetic categorization task the fa∼sa
continuum after hearing an ambiguous token [? f s−a] in place of [f] in words during a lexical
decision task than listeners who heard normal [f] in the same words. This result was especially
clear when looking at the difference from “Before”-“After” between the test and the control group.
The importance of this comparison was also made clear with Experiment 2. The ‘After”-“After”
comparison between the two groups showed no statistically significant difference and in fact showed
more alveolar responses for the test group (opposite of what is predicted). Overall, the results from
experiment challenge what was shown with audio-visual recalibration experiments, mainly that the
phonological environment needs to be identical in training and testing conditions in order to see an
effect (Reinisch et al., 2014).

Experiment 3 was used to test whether or not type experience could explain why lexical retuning
experiments allowed for generalizability. The training and testing environments for Experiment
3 were identical to the ones used in Experiment 2, but the overall number of training items was
reduced. In Experiment 2, listeners heard 34 unique training words, 1 time each (17 containing
[? f s−a]; 17 containing [s]). In Experiment 3, listeners heard 2 unique training words, 17 times
each (1 containing [? f s−a], 1 containing [s]). This was done to attempt to match the training
experience that participants typically receive in audio-visual recalibration experiments. In those
experiments, the same training stimuli is presented over and over again. Results from Experiment 3
show that listeners were no longer able to generalize their perceptual training when presented with
the same ambiguous stimuli multiple times rather than multiple, unique stimuli. In this case, there

40

was no statistically significant difference between the test and control groups and suggests that type
experience is necessary for generalization within lexical retuning.

6.2 General Discussion

One idea presented early on in this thesis was to probe whether or not different types of
disambiguating information were treated differently by the speech perception mechanism. Based
on the results presented here, I can not conclusively answer this question. What has been shown
is that when the training and testing conditions are more closely matched for the two perceptual
learning experimental paradigms, both unambiguous visual and unambiguous lexical information
do appear to show similar results, specifically in the realm of generalizability. The results from
Experiment 3 most explicitly showed that removing type variability in the training set resulted
in a lack of generalizability within the lexical retuning paradigm, just as with the audio-visual
recalibration paradigm. This suggests that it would be hasty to attribute previously observed
differences in generalizability between the two paradigms to different underlying mechanisms.
However, it should be pointed out that in order to complete the argument, it is now crucial to see
what happens when the audio-visual recalibration paradigm is altered to contain type variation as
in Experiments 1 and 2. I leave this as a demonstration for further research.

In general, the ability for generalization within the lexical retuning paradigm supports a view of
speech perception involving abstract representations (i.e., phonemes, features). This view has been
questioned by more recent experimental results (Reinisch et al., 2014; Mitterer et al., 2016, 2018),
but the results presented in this thesis put them into question. The generalizable nature of perceptual
learning has up to this point been almost exclusive to the lexical retuning paradigm, where it has been
shown to generalize over position (Jesse and McQueen, 2011), features (?Durvasula and Nelson,
2018), and to new words (McQueen et al., 2006a; Sjerps and McQueen, 2010). Reinisch et al.’s
(2014) provide the strongest argument against generalization using the audio-visual recalibration
paradigm, but this thesis has identified a critical difference between the two experiment styles that
In this regard, the lack of generalization shown in
may have lead to the contradictory claims.

41

Reinisch et al. (2014) could be taken as a result of no type experience in the training stimuli and
not assumed to be a result of the lack of abstract representations within speech perception.

6.3 Conclusion

An important question within speech perception research is what do listeners do when presented
with ambiguous input? This thesis has shown that the answer to that is largely dependent on the
type of disambiguating information they receive. What is most clear is that there appears to be a
type/token distinction. When listeners hear ambiguity in only one specific type, they consider it to
be an idiosyncrasy and do not generalize it to other types, even when hearing multiple repetitions
of the token. This is an alternative explanation to previous perceptual learning results in the
audio-visual recalibration domain (Reinisch et al., 2014) that claimed that the effect was highly
constrained to conditions that completely matched the testing environment. The necessity of type
variation has also been shown in other learning domains (Gerken and Bollt, 2008; Denby et al.,
2018) and suggests that this may not entirely be specific to perceptual learning paradigms. It is
therefore necessary to see how type experience interacts with visually-guided perceptual learning
in order to tease apart the roles that different modalities have on the speech perception mechanism.
Overall this thesis has made three contributions. First, it has corroborated the idea that lexical
retuning can result in generalization. This was observed under broad conditions in Experiment 1
and more strict conditions in Experiment 2. Second, it has shown that type experience is crucial
for the generalization to occur. Third, it has argued for a change to the methodology and analysis
of perceptual learning experiments by adding the “Before” phonetic categorization task to the
experimental design. This subsequently allowed for better insight into how the ambiguous token
presented in the lexical decision task affects the categorical representation of the segments under
question by allowing for a comparison of the differences within test and control groups.

42

APPENDICES

43

APPENDIX A

LEXICAL DECISION TASK WORD LIST

Training Words

Experiment 1
truss
bluff
chess
chef
cliff
bliss
less
deaf
deuce
poof
whiff
kiss
geese
beef
boss
cough
such
fudges
sect
soup
sash
say
sip
silt
soon
sulk

felt
food
fab
fade
fig
fish
fool
full

Experiment 2/3
kiss
cliff
geese
reef
thief
piece
bliss
whiff
silk
film
filth
sick
sim
fish
sing
fill
sip
fifth
fib
seek
seem
fiend
seagull
feeble
female
seated
seeing
fetal
seeker
fever
seeping
fiji
feline
seething

Table A.1: Training words for Lexical Decision Tasks. The bolded subset indicate Experiment
3.

44

Real

Filler Words

Nonce

truck
tomb
right
loot
wet
trim
pen
lick
mole
mall
twin
lake
line
rat
while
will
day
drip
globe
toy

building

rowboat

tuba
laker

daydream

butter
little
data
liquor
maker
treaty
trooper
toilet
lemur
baking
typing
boycott
pillow
yellow
jupiter
lump

weg
yilk
queng
yoog
ledmon
troplund
yoogul
cratbool
pingloon
kotlim

pendalimt
cridamlo

himp
rone
relnt
trad
waip
ligbee
yoomake
rekmate
magmilnt
mitger
wiblee
midaran
lopenad
doonagort
rivathrog

noopalingdo

poginito
dewagelin
krebdingolee
tridalondit
plagtorin

plone
twek
trag
toin
rone

poindool
metroy
deemlond
ratwell
petnole
hapekt

wegtamlip
dripnalog
himnapile
podaling
habapithloo
habalipdee
blegondanit
bindilamok
radoobling
trongolemp

heen
hoong
gelp
roip
giblo
wegring
talkibd
retlid

bidbowk
lempok
libmadoop
kebnoti
yoodalin
drataloob
datoonmeg

boondatribnok
hundapegnort

kililitbi
kinudiplo
kablingno
hapooli

Table A.2: Filler words for Lexical Decision Tasks. All 116 words were used in Experiments 1
and 2. The bolded subset were used in Experiment 3

45

APPENDIX B

LDT TRAINING WORD FREQUENCY INFORMATION

B.1 Experiment 1

SYLLABLES
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1

POSITION WORD FREQ/MIL LOGFREQ WORD FREQ/MIL LOGFREQ
coda
coda
coda
coda
coda
coda
coda
coda
onset
onset
onset
onset
onset
onset
onset
onset
onset
Table B.1: Experiment 1 LDT training word frequency data

0.33
7.45
3.14
111.1
2.86
121.16
1.59
124.29
291.22
0.73
25.2
1.14
1639.78
5.1
0.33
257.65
0.75

truss
chess
bliss
less
deuce
kiss
geese
boss
such
sect
soup
sash
say
sip
silt
soon
sulk

1.26
2.58
2.21
3.75
2.17
3.79
1.91
3.8
4.17
1.58
3.11
1.77
4.92
2.42
1.26
4.12
1.59

bluff
chef
cliff
deaf
poof
whiff
beef
cough
fudge
felt
food
fab
fade
fig
fish
fool
full

6.22
11.88
21.57
14.53
2.16
2.49
19.71
8.78
3.63
119.82
154.43
0.61
5.61
1.22
83.49
89.33
166.9

2.5
2.78
3.04
2.87
2.05
2.11
3
2.65
2.27
3.79
3.9
1.51
2.46
1.8
3.63
3.66
3.93

Figure B.1: Boxplots for Experiment 1 LDT training word frequency (log)

46

B.2 Experiments 2/3

SYLLABLES
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2

POSITION WORD FREQ/MIL LOGFREQ WORD
coda
coda
coda
coda
onset
onset
onset
onset
onset
onset
onset
onset
onset
onset
onset
onset
onset

kiss
geese
piece
bliss
silk
sick
sim
sing
sip
seek
seem
seagull
seated
seeing
seeker
seeping
seething

3.04
2.31
3.09
2.11
3.52
2.37
3.63
3.35
2.99
1.82
2.16
1.94
3.21
1.78
3.01
1.88
1.67

cliff
reef
thief
whiff
film
filth
fish
fill
fifth
fib
fiend
feeble
female
fetal
fever
fiji
feline

21.57
4
24.27
2.49
65.25
4.53
83.49
43.94
19.2
1.27
2.8
1.69
31.61
1.16
19.94
1.47
0.9

FREQ/MIL LOGFREQ
121.16
1.59
124.49
3.14
9.78
165.43
1.1
97.59
5.1
18.31
139.82
1.22
8.55
109.63
0.92
0.37
0.45

3.79
1.91
3.8
2.21
2.7
3.93
1.76
3.7
2.42
2.97
3.85
1.8
2.64
3.75
1.68
1.3
1.38

Table B.2: Experiment 2/3 LDT training word frequency data. All words were used in
Experiment 2. The bolded subset were used in Experiment 3.

Figure B.2: Boxplots for Experiment 2 LDT training word frequency (log)

47

BIBLIOGRAPHY

48

BIBLIOGRAPHY

Baart, M. and Vroomen, J. (2010). Phonetic recalibration does not depend on working memory.

Experimental brain research, 203(3):575–582.

Bertelson, P., Vroomen, J., and De Gelder, B. (2003). Visual recalibration of auditory speech

identification: a mcgurk aftereffect. Psychological Science, 14(6):592–597.

Boersma, P. and Weenink, D. (2016). Praat: doing phonetics by computer [Computer program].

Version 6.0.19, retrieved 13 June 2016 from http://www.praat.org/.

Brysbaert, M. and New, B. (2009). Moving beyond kučera and francis: A critical evaluation of
current word frequency norms and the introduction of a new and improved word frequency
measure for american english. Behavior research methods, 41(4):977–990.

Cutler, A., McQueen, J. M., Butterfield, S., and Norris, D. (2008). Prelexically-driven perceptual

retuning of phoneme boudaries.

Delattre, P. C., Liberman, A. M., and Cooper, F. S. (1955). Acoustic loci and transitional cues for

consonants. The Journal of the Acoustical Society of America, 27(4):769–773.

Delattre, P. C., Liberman, A. M., and Cooper, F. S. (1962). Formant transitions and loci as acoustic

correlates of place of articulation in american fricatives. Studia linguistica, 16(1-2):104–122.

Denby, T., Schecter, J., Arn, S., Dimov, S., and Goldrick, M. (2018). Contextual variability and
exemplar strength in phonotactic learning. Journal of Experimental Psychology: Learning,
Memory, and Cognition, 44(2):280.

Durvasula, K., Huang, H.-H., Uehara, S., Luo, Q., and Lin, Y.-H. (2018). Phonology modulates
the illusory vowels in perceptual illusions: Evidence from mandarin and english. Laboratory
Phonology: Journal of the Association for Laboratory Phonology.

Durvasula, K. and Kahng, J. (2015).

Illusory vowels in perceptual epenthesis: The role of

phonological alternations. Phonology, 32(3):385–416.

Durvasula, K. and Kahng, J. (2016). The role of phrasal phonology in speech perception: What

perceptual epenthesis shows us. Journal of Phonetics, 54:15–34.

Durvasula, K. and Nelson, S. (2018). Lexical retuning targets features.

Annual Meetings on Phonology, volume 5.

In Proceedings of the

Eimas, P. D. and Corbit, J. D. (1973). Selective adaptation of linguistic feature detectors. Cognitive

Psychology, 4(1):99 – 109.

49

Eisner, F. and McQueen, J. M. (2005). The specificity of perceptual learning in speech processing.

Attention, Perception, & Psychophysics, 67(2):224–238.

Eisner, F. and McQueen, J. M. (2006). Perceptual learning in speech: Stability over time. The

Journal of the Acoustical Society of America, 119(4):1950–1953.

Ganong, W. F. (1980). Phonetic categorization in auditory word perception. Journal of experimental

psychology: Human perception and performance, 6(1):110.

Gaskell, M. G. and Marslen-Wilson, W. D. (1996). Phonological variation and inference in lexical
access. Journal of Experimental Psychology: Human perception and performance, 22(1):144.

Gaskell, M. G. and Marslen-Wilson, W. D. (1998). Mechanisms of phonological inference in
speech perception. Journal of Experimental Psychology: Human Perception and Performance,
24(2):380.

Gerken, L. and Bollt, A. (2008). Three exemplars allow at least some linguistic generalizations: Im-
plications for generalization mechanisms and constraints. Language Learning and Development,
4(3):228–248.

Gow, D. W. (2003). Feature parsing: Feature cue mapping in spoken word recognition. Perception

& Psychophysics, 65(4):575–590.

Hillenbrand, J., Getty, L. A., Clark, M. J., and Wheeler, K. (1995). Acoustic characteristics of
american english vowels. The Journal of the Acoustical society of America, 97(5):3099–3111.

Jesse, A. and McQueen, J. M. (2011). Positional effects in the lexical retuning of speech perception.

Psychonomic Bulletin & Review, 18(5):943–950.

Kleinschmidt, D. F. and Jaeger, T. F. (2015). Robust speech perception: recognize the familiar,

generalize to the similar, and adapt to the novel. Psychological review, 122(2):148.

Kraljic, T. and Samuel, A. G. (2005). Perceptual learning for speech: Is there a return to normal?

Cognitive psychology, 51(2):141–178.

Kraljic, T. and Samuel, A. G. (2006). Generalization in perceptual learning for speech. Psychonomic

bulletin & review, 13(2):262–268.

Liberman, A. M., Harris, K. S., Hoffman, H. S., and Griffith, B. C. (1957). The discrimination
of speech sounds within and across phoneme boundaries. Journal of experimental psychology,
54(5):358.

McGurk, H. and MacDonald, J. (1976). Hearing lips and seeing voices. Nature, 264(5588):746–

748.

50

McMurray, B. and Jongman, A. (2016). What comes after/f/? prediction in speech derives from

data-explanatory processes. Psychological science, 27(1):43–52.

McQueen, J. M., Cutler, A., and Norris, D. (2006a). Phonological abstraction in the mental lexicon.

Cognitive Science, 30:1113–1126.

McQueen, J. M., Norris, D., and Cutler, A. (2006b). The dynamic nature of speech perception.

Language and speech, 49(1):101–112.

Mitterer, H., Chen, Y., and Zhou, X. (2011). Phonological abstraction in processing lexical-tone

variation: Evidence from a learning paradigm. Cognitive Science, 35(1):184–197.

Mitterer, H., Cho, T., and Kim, S. (2016). What are the letters of speech? testing the role of
phonological specification and phonetic similarity in perceptual learning. Journal of Phonetics,
56:110–123.

Mitterer, H., Reinisch, E., and McQueen, J. M. (2018). Allophones, not phonemes in spoken-word

recognition. Journal of Memory and Language, 98:77 – 92.

Mitterer, H., Scharenborg, O., and McQueen, J. M. (2013). Phonological abstraction without

phonemes in speech perception. Cognition, 129(2):356–361.

Norris, D., McQueen, J. M., and Cutler, A. (2000). Merging information in speech recognition:

Feedback is never necessary. Behavioral and Brain Sciences, 23(3):299–325.

Norris, D., McQueen, J. M., and Cutler, A. (2003). Perceptual learning in speech. Cognitive

Psychology, 30(2):1113–1126.

Peirce, J., Gray, J. R., Simpson, S., MacAskill, M., Höchenberger, R., Sogo, H., Kastman, E., and
Lindeløv, J. K. (2019). Psychopy2: Experiments in behavior made easy. Behavior research
methods, pages 1–9.

Reinisch, E. and Holt, L. L. (2014). Lexically guided phonetic retuning of foreign-accented
speech and its generalization. Journal of Experimental Psychology: Human Perception and
Performance, 40(2):539.

Reinisch, E., Wozny, D. R., Mitterer, H., and Holt, L. L. (2014). Phonetic category recalibration:

What are the categories? Journal of phonetics, 45:91–105.

Samuel, A. G. (1986). Red herring detectors and speech perception:

adaptation. Cognitive psychology, 18(4):452–499.

In defense of selective

Samuel, A. G. and Kat, D. (1998). Adaptation is automatic. Perception & psychophysics, 60(3):503–

510.

51

Samuel, A. G. and Kraljic, T. (2009). Perceptual learning for speech. Attention, Perception, &

Psychophysics, 71(6):1207–1218.

Scharenborg, O., Mitterer, H., and McQueen, J. M. (2011). Perceptual learning of liquids. In Inter-
speech 2011: 12th Annual Conference of the International Speech Communication Association,
pages 149–152.

Sjerps, M. J. and McQueen, J. M. (2010). The bounds on flexibility in speech perception. Journal

of Experimental Psychology: Human Perception and Performance, 36(1):195.

Soli, S. D. (1981). Second formants in fricatives: Acoustic consequences of fricative-vowel

coarticulation. The Journal of the Acoustical Society of America, 70(4):976–984.

Van Linden, S. and Vroomen, J. (2007). Recalibration of phonetic categories by lipread speech
versus lexical information. Journal of Experimental Psychology: Human Perception and Per-
formance, 33(6):1483.

Vroomen, J., van Linden, S., Keetels, M., De Gelder, B., and Bertelson, P. (2004). Selective
adaptation and recalibration of auditory speech by lipread information: dissipation. Speech
Communication, 44(1-4):55–61.

Yeni-Komshian, G. H. and Soli, S. D. (1981). Recognition of vowels from information in fricatives:
Perceptual evidence of fricative-vowel coarticulation. The Journal of the Acoustical Society of
America, 70(4):966–975.

52