lfiI/l/l/Il/I/lllllll/l/ll M {LIBRARY % Michigan State University This is to certify that the thesis entitled EFFECTS OF DIFFERENT COMPETING SPEECH MESSAGES ON IDENTIFICATION OF SYNTHETIC SENTENCES presented by Carol Goldschmidt has been accepted towards fulfillment of the requirements for Master of Arts degreein Audiology V p / /W/MD Major professor Michael R. Chial, Ph.D. Dfie January 31. 1979 0-7639 OVERDUE FINES ARE 25¢ PER DAY PER ITEM Return to book drop to remove this checkout from your record. '-- “I“- —- J- ,3; EFFECTS OF DIFFERENT COMPETING SPEECH MESSAGES ON IDENTIFICATION OF SYNTHETIC SENTENCES by Carol Goldschmidt A THESIS Submitted to Michigan State University in partial fulfillment of the requirements for the degree of MASTER OF ARTS Department of Audiology and Speech Sciences 1979 ABSTRACT EFFECTS OF DIFFERENT COMPETING SPEECH MESSAGES ON IDENTIFICATION OF SYNTHETIC SENTENCES By Carol Goldschmidt Thirty-six bilaterally normal-hearing listeners were tested with the Synthetic Sentence Identification (881) test under six ipsilateral competing message conditions. A connected- discourse competing message was recorded by (a) the same talker who recorded the target message, (b) a different male talker, and (c) a female talker. Each of these competing messages was administered in forward and backward modes to independent groups of 12 subjects at message-to—competition ratios of -12, -18, and -24 dB. Mean percent-correct responses differed significantly as a function of (l) mes- sage-to—competition ratio, (2) talker, (3) interactions among message-to-competition ratio and talker, and (4) interactions among the direction of the competing signal and talker. ACKNOWLEDGEMENTS I consider this document to be the product of the most serious and demanding undertaking thus far in my academic career. While the completion of this Master's thesis does represent a goal achieved, the learning processes that enabled me to achieve this goal are of greater importance to me. As I reflect on my involvement with this research over the last months, I am consistently reminded of the valuable contributions made by other individuals. My thesis committee, including Drs. Michael R. Chial, Linda Lou Smith, and Leo V. Deal, provided me with needed direction and guidance throughout my thesis work. The expertise of each member is evidenced in this document, and I am most appreciative of their efforts on my behalf. The very special contributions of my thesis committee chairman, Dr. Chial, have made this experience invaluable to me. His understanding and acceptance of my goals and abilities were very supportive throughout my Master's degree program. I am grateful to those who shared the highs and lows of this thesis with me. My parents have always had a strong commitment to the goals and aspirations that led me to this point, and I hope they realize their formidable roles in my education. My entire family has shown patience and under— standing at times when I probably didn't deserve either. ii Lastly, a special thanks is due to my friends, who kept me smiling throughout. iii TABLE OF CONTENTS Page LIST or TABLES . . . . . . . . . . . . . . . . . . . . vii LIST or FIGURES. . . . . . . . . . . . . . . . . . . . ix LIST OF APPENDICES . . . . . . . . . . . . . . . . . . xi CHAPTER BACKGROUND . . . . . . . . . . . . . . . . . . . . . . 1 Introduction. . . . . . . . . . . . . . . . . . . . 1 Statement of the Problem. . . . . . . . . . . . . . 8 METHOD . . . . . . . . . . . . . . . . . . . . . . . . 10 Introduction. . . . . . . . . . . . . . . . . . . . 10 Subjects. . . . . . . . . . . . . . . . . . . . . . 11 Speech Materials. . . . . . . . . . . . . . . . . . 12 Talkers. . . . . . . . . . . . . . . . . . . . . 14 Recording Equipment. . . . . . . . . . . . . . . 14 Recording Procedures . . . . . . . . . . . . . . 15 Apparatus . . . . . . . . . . . . . . . . . . . . . 16 Calibration Procedures. . . . . . . . . . . . . . . 16 Experimental Procedures . . . . . . . . . . . . . . 18 RESULTS. . . . . . . . . . . . . . . . . . . . . . . . 24 Introduction. . . . . . . . . . . . . . . . . . . . 24 iv CHAPTER RESULTS (continued) Data Reduction. Scoring of subject responses Categorization of debriefing protocols Statistical Procedures. Synthetic Sentence Identification Data. Description. Reliability of measurement Inferential analysis Debriefing Data Description of results Summary DISCUSSION Introduction. Reliability . Test list equivalence. Test—retest consistency. Validity. The effects of message-to-competition ratio and semantlc content The effect of MCR- by Talker Difference Inter- actions. . . . . . . . . . . . . . . The effect of Semantic Content-by-Talker Difference Interactions. . . . . . Implications for Clinical Practice. Limitations of the Present Study. Findings. Page 26 26 26 27 28 28 34 41 SO SO 54 55 SS 55 SS 56 S7 57 61 61 72 74 76 CHAPTER Page DISCUSSION (continued) Conclusions . . . . . . . . . . . . . . . . . . . . 78 Suggestions for Further Study . . . . . . . . . . . 79 SUMMARY. . . . . . . . . . . . . . . . . . . . . . . . 81 APPENDICES . . . . . . . . . . . . . . . . . . . . . . 88 REFERENCES . . . . . . . . . . . . . . . . . . . . . . 115 vi Table LIST OF TABLES Page Mean percent-correct scores, standard deviations, and ranges based upon single observations for three groups of 12 normal-hearing subjects in each of six listening competing message source and direction conditions for three message-to- competition ratios . . . . . . . . . . . . . . . . 29 Test-retest correlation coefficients (Pearson r) and average coefficients for nine subjects on the synthetic sentence identification task under six listening competing message source conditions. . . 38 Analysis of variance for synthetic sentence identification indexed by percent-correct scores as a function of message-to-competition ratio, semantic content and talker. . . . . . . . . . . . 44 Newman-Keuls' specific-comparisons test results for three competing message source means (collapsed across the factor of presence vs. absence of semantic content) for each of three MCRS. Critical values are given for all possi- ble ranges (from two means apart to three means apart). A difference between any two means is significant when it exceeds the appropriate critical value. The level of significance was Pa 5 0.05. . . . . . . . . . . . . . . . . . . . . 46 Newman-Keuls' specific-comparisons test results for three MCRS (collapsed across the factor of presence vs. absence of semantic content) for each of three competing message sources (S, D , D ). Critical values are given for all possible ranges (from two means apart to three means apart). A difference between any two means is significant when it exceeds the appropriate critical value. The level of significance was Pa 5 0.05. . . . . . . . . . . . . . . . . . . . . 47 vii Table (Cont.) Page 6. 10. Newman-Keuls' specific-comparisons test results for three competing message sources (collapsed across the factor of MCR) for each of two levels of semantic content (F vs. B). Critical values are given for all possible ranges (from two means apart to three means apart). A difference between any two means is significant when it exceeds the appropriate critical value. The level of signi- ficance was Pa 3 0.05. . . . . . . . . . . . . . . 48 Frequency table of first responses by 36 normal- hearing subjects in response to debriefing strategy questions. . . . . . . . . . . . . . . . 51 Cross tabulation of responses by normal-hearing subjects reporting on perceiVed distractiveness of competing messages on the synthetic sentence ' identification task. . . . . . . . . . . . . . . . 53 Psychometric function slopes compared from data reported in three studies. . . . . . . . . . . . . 59 Newman-Keuls' specific-comparisons test results for six competing message types and sources for three separate MCRs. Critical values are given for all possible ranges (from two means apart to Six means apart). A difference between any two means is significant when it exceeds the appro- priate critical value. The level of significance was Pa 5 0.05 . . . . . . . . . . . . . . . . . . 62 viii Figure LIST OF FIGURES Performance-intensity functions generated by eight normal-hearing subjects when identi- fying synthetic sentences, as taken from Dirks and Bower (1969). . . . . . . . . Block diagram of experimental apparatus for identification of synthetic sentences Flow chart of experimental procedures Mean percent—correct scores based upon Single observations for three groups of 12 normal- hearing subjects in each of six listening competing message source and direction condi- tions as a function of message-to-competition ratio . . . . . . . . . . . . . . . . . . . . Mean percent-correct scores based upon single observations for 12 normal-hearing subjects for competing message conditions of forward- same, forward-different , and forward-different2 as a function of messagA-to-competition ratio Mean percent-correct scores based upon single observations for 12 normal-hearing subjects for competing message conditions of backward- same, backward-different , and backward-different2 as a function of message-to-competition ratio Mean percent-correct scores based upon single observations for 12 normal-hearing subjects for competing message conditions of forward- same and backward-same as a function of message-to-competition ratio . . . . . . . . . Mean percent-correct scores based upon single observations for 12 normal-hearing subjects for competing message conditions of forward- different and backward-different as a function 6f message-to-competitioA ratio . . ix Page 17 19 31 32 33 35 36 Figure (cont.) 9. 10. 11. 12. 13. Mean percent-correct scores based upon single observations for 12 normal-hearing subjects for competing message conditions of forward- different and backward-different as a function 8f message-to-competitiog ratio Pearson product-moment correlations based upon test-retest data for three normal-hearing subjects for six listening competing message source conditions at a message-to-competion ratio of -12 dB . . . . . . . . . . . . Pearson product-moment correlations based upon test-retest data for three normal-hearing subjects for six listening competing message source conditions at a message-to-competion ratio of -18 dB. . . . . . . . . . . . . Pearson product-moment correlations based upon test—retest data for three normal-hearing subjects for six listening competing message source conditions at a message-to-competion ratio of ~24 dB. . . . . . . . . . . . . . . . Mean Pearson product-moment correlations based upon test-retest data for three groups of three normal~hearing subjects across listening conditions as a function of message-to-compe- tition ratio . . . . . . . . . . . . . . Page 37 39 40 42 43 LIST OF APPENDICES Appendix Page A. Subject screening protocol . . . . . . . . . . . 88 B. Pilot Study. . . . . . . . . . . . . . . . . . . 89 C. Synthetic Sentence Identification Test Lists used as stimulus materials for the primary message (Chial, 1978). . . . . . . . . . . . . . 100 Transcript of competing message passage. . . . . 102 Standardized listener instructions provided to subjects before the experimental session and informed consent release form. . . . . . . . 107 Subject debriefing protocol. . . . . . . . . . . 109 Matrices of experimental conditions and listening conditions depicting order of presentation . . . 110 Individual subject scores (percent-correct) for 36 normal-hearing subjects on the synthetic sentence identification task for six competing message source and direction conditions. . . . . 112 xi BACKGROUND INTRODUCTION The analytic study of speech understanding utilizes a variety of verbal materials including nonsense syllables, phonetically-balanced monosyllables, and dissyllables (spon- daic and otherwise). These materials differ in redundancy and in the type and amount of information they provide about a listener or communication system. Analysis of these materials has been provided by Fletcher and Steinberg (1929); Fletcher (1929); Miller, Weiner, and Stevens (1946); Hud— gins, and others (1947); and Egan (1948). For assessment of auditory function, techniques were reported by Walsh and Silverman (1946), Hirsh (1952), and Silverman and Hirsh (1955). Speech intelligibility tests using isolated syllables and word materials cannot focus directly on temporal process- ing factors because the stimuli are too brief. To assess temporal processing function, it is necessary to use materials of sufficient duration to allow for alteration of the temporal characteristics of speech messages (Jerger, 1970b). Sentence length materials, if constrained in length, are more suitable for study of temporal factors. Miller (1951, p. 791) stated that sentences should be further restricted, because the amount of information relayed by a sentence varies with the 2 degree of contextual constraints. Several sentence lists have been devised (Fletcher and Steinberg, 1929; Hudgins, and others, 1947; Davis and Silverman, 1960), but they do not meet these criteria. Miller (1951, p. 790) discussed the generation of sen- tential materials based on statistical properties of Eng- lish. The method of producing sentences or approximations of English was structured by the relative frequencies of occurrence of words or sequences of words. This procedure of generating artificial sentences provided a basis for assessment devices. Speaks and Jerger (1965) and Jerger (1970b) described a different approach to sentence tests. The Synthetic Sen- tence Identification (SSI) test describes a listener's abil- ity to identify grammatically constrained sentences of fixed length. It departs from traditional test paradigms in four ways. First, it uses syntactically constrained units as stimuli. Second, the SSI test uses a limited number of explicitly designated alternatives (i.e., closed message set). Third, learning or practice effects can be determined with relative ease for a message set. Finally, experimenter error can be minimized through automated data acquisition procedures. Jerger and Hayes (1976) used synthetic sentences with a competing message for hearing aid evaluations, reasoning that sentential stimuli more closely approximate everyday communi- cation experiences. Greenberg (1975) also investigated 3 hearing aid performance by presenting synthetic sentences in noise with low levels of harmonic distortion to normal lis- teners through an ear-level hearing aid. Results indicated that levels of harmonic distortion as low as 3% adversely affected identification of synthetic sentences. Edgerton and others (1977) compared preferred hearing aids determined with the SSI test and clients' subjective rating of hearing aids with results of an independent signal detection task. Subject biases did not affect the results of aided SSI tests. In addition, Speaks and Jerger (1967) applied the statistical- decision model to the synthetic sentence identification task, and reported that normal listeners generated receiver oper- ating characteristic curves consistent with the model. Again, it was possible to describe SSI performance that was free of the listener's criterion. Jerger (1970b) studied the diagnostic application of the SSI to listeners with retrocochlear disorders. Test materials were degraded by various means in an attempt to differentiate between brain stem and temporal lobe lesions. The lesion sites were best differentiated by effects of ipsilateral and contralateral competing messages on the SSI test. The experimental procedure employed for the SSI test follows that of traditional articulation testing (Fletcher and Steinberg, 1929; Egan, 1948). SSI test results are des- cribed with a performance-intensity (P-I) function. It has been found that the roughly sigmoidal P-I function for the 4 SSI is steep, rising to a 90-100% level at sufficiently high intensities (Speaks, Jerger and Jerger, 1966). In quiet, the SSI task was found to be too easy to differentiate among abilities to understand speech (Jerger, Speaks, and Trammell, 1968). Several modifications of the SSI procedures failed to adequately decrease the slope of the P-I function (Speaks, Jerger and Jerger, 1966; Speaks, 1967; Speaks and Karmen, 1967; Speaks, Karmen and Benitez, 1967; Jerger, Speaks and Trammell, 1968). Low-pass filtering (fC = 500 Hz) had no appreciable effect on the slope and shape of the SSI P-I fun- ction nor did the addition of masking noise. Dirks and Wil- son (1969) observed similar results using synthetic sentences and white noise in a sound field condition. A further variation of the SSI test includes introduc- tion of a competing speech message to increase test diffi- culty (Speaks, Karmen and Benitez, 1967; Jerger, Speaks and Trammell, 1968; Dirks and Bower, 1969; Garstecki and Mulac, 1974). Speaks, Karmen and Benitez (1967) reported the random masking and disruptive features of a single-talker competing message decreased the slope of the P-I function more than con- tinuous noise. At a constant message-to-competition ratio (MCR), theintensity levels of the messages were varied consid- erably with no significant effect on performance. Dirks and Bower (l969)found that the masking effect of a single-talker competing message on the SSI was not due to disruptive features of semantic content of the competing message. The competing 5 message was reproduced in the forward (CMF) and backward (CMB) mode, with the reverse mode used to reduce the semantic content of the competing message. Dirks and Bower presented the primary and competing messages at varied message-to-com- petition ratios by changing the intensity of the competing message (Mfll= 0 dB re: sound pressure level of primary mes- sage). Following extensive practice listening, eight normal listeners identified a single set of ten synthetic sentences under different listening conditions. The first experiment examined possible effects of sem- antic content on the identification of synthetic sentences. When the competing message is reversed, the temporal charac- teristics of the forward message are maintained, while the semantic component has been removed (Meyer-Eppler, 1950). Dirks and Bower presented ten synthetic sentences with either a CMF or a CMB at each of four MCR's. The primary or target signal was presented at a sound pressure level (SPL) of 40 dB (re: 20pPa) and the competing message was varied so that MCR's of -12, -18, -24, and -30 dB (re: sound pressure level of primary message) were obtained. P-I functions were gener— ated for each subject, and mean percent-correct responses were tabulated across subjects. These are seen in Figure I. Results indicated that P-I functions for synthetic sentences were very similar when the identification task was performed with either a forward or reversed competing message. Thus, the distrac— tive features of a competing message do not significantly affect synthetic sentence identification when the same speaker I T I T" '00- CMF o-——o " 90_ CMB A---A 80- 70‘- 60- 50 - 40- 30- 20" IOh- '- PERCENT CORRECT IDENTIFICATION l L I l -30 - 24 -I8 -I2 SICIIIIAL[COMPELIII‘IIGB MESSAGE RATIO I Figure 1. P-I functions generated by eight normal-hearing srbjects (four trials each) when identifying synthetic sentences with a forward conpeting message (CMF) and a backward carpeting message (CMB) as a function of signal- to-competing message ratio (Dirks and Bower, 1969) . 7 is heard on both the target and competing message. Subse- quent experiments by Dirks and Bower also revealed that per- formance did not differ significantly when a foreign language competing message was used in the forward and backward mode. It was concluded that distractional factors linked to seman- tic content of the competing message did not measurably effect the P-I functions for the identification of synthetic sentences. When there was an intensity and time difference between the primary and competing message, listeners attended to the differences in loudness and temporal patterns to achieve higher percent-correct scores under a same talker condition. When there was no difference in the intensity and temporal pattern of the two signals, identification of synthetic sentences became more difficult and produced lower scores. It is known that the human auditory system can selective- ly attend to a specific auditory signal in the presence of other competing signals. A listener can "tune in” a signal and appropriately process it, while simultaneously and consci- ously "tuning out" other signals. The phenomenon is labeled the "cocktail party" effect (Cherry, 1954; 1966; Hardy, 1959; MacLean, 1959; Kaiser and David, 1960; Mitchell and others, 1971). While the "cocktail party" effect is essentially binau- ral, it is conceivable that the ability to auditorally select and process one signal from others might also be present in a monaural situation, such as with the 881. The "cocktail party” effect demonstrates a listener's choice of a desired target 8 message from a variety of messages as long as the competing messages do not suppress the desired target message (Kaiser and David, 1960). In an actual cocktail party situation, a listener may rely on cues such as voice quality differences to accomplish attentional focus on a chosen target message. Dirks and Bower (1969) did not observe effects due to semantic content when the same talker was used for both pri- mary and competing messages, but the issue of semantic effect has not yet been resolved for the case in which different talkers produce the primary and competing messages. It is possible that perceived qualitative differences between two speakers (primary vs. competing) may alter a listener's abil- ity to identify a primary message. Such information would be useful if for no reason other than face validity: in real listening situations, a single talker cannot Simultaneously produce two separate signals. The primary message may pro— duce varying amounts of "temporal overlap" when paired with similar and different competing messages. Further, the effects of semantic content may differ from what was found by Dirks and Bower when two different talkers produce the primary and competing messages. STATEMENT OF THE PROBLEM This study sought to determine whether normal-hearing subjects would perform differently in a synthetic sentence identification task when the same and different talkers were 9 used to produce the primary and competing messages. The following questions were asked. 1. Does mean normal listener performance in an SSI task vary significantly as a function of the sim- ilarity of the competing message source to the primary message source? Does mean normal listener performance in an SSI task vary significantly as a function of the presence vs. absence of semantic content in the competing message? Does mean normal listener performance in an SSI task vary significantly as a function of inter— actions among (a) the similarity of primary and competing message sources, and (b) the presence vs. absence of semantic content in the competing message? Does mean normal listener performance in an SSI task vary significantly as a function of inter- actions among (a) the Similarity of primary and competing message sources, and (b) message-to-competition ratio? Does mean normal listener performance in an SSI task vary significantly as a function of message-to- competition ratio? METHOD INTRODUCTION Synthetic sentence materials provide a useful tool for investigating temporal factors of speech understanding. The SSI test presents listeners with syntactically controlled items in a closed message set. In order to adequately dif- ferentiate among abilities to process sentential stimuli, a competing message is added to the sentence Signal. Various message-to—competition ratios are achieved by altering the intensity of the competing message and maintaining the sen- tence signal at a constant level. Dirks and Bower (1969) found that when the competing message was presented in both the forward and reverse modes, normal listeners performed in a similar fashion under both listening conditions when the same talker read both the synthetic sentences and the compet- ing message. The present study addressed the issues of talker similarity, presence vs. absence of semantic content in the competing message, and message-to-competition ratio singularly and in combination. SUBJECTS Thirty-six bilaterally normal-hearing subjects served in this investigation. The mean subject age was 22.2 years 10 11 and ranged from 20 to 30 years. The sample included thirty females and six males. Subjects were required to pass a routine audiometric screening prior to participation in the experiment. Normal listeners had air-conduction hearing threshold levels (HTL) no poorer than 15 dB (re: ANSI 53.6-1969) at audiometric test frequencies between 250 and 4000 Hz. Spondee thresholds (STS) were measured using the procedure suggested by Martin and Stauffer (1975) and were found to be commensurate with air—conduction thresholds (15 dB HTL or less). All subjects had speech discrimination scores of 90% or better on the Northwestern University Audi- tory Test, No. 6 (presented at 40 dB SL, Tillman and Carhart, 1966). Listeners presented Type A tympanograms (Jerger, 1970a) and bilateral acoustic reflexes within the normal range of 70-100 dB HTL (Jepson, 1963). Subjects did not demonstrate bilateral acoustic reflex decay at test frequen- cies of 500 and 1000 Hz. Listeners reported no otologic or familial history of hearing loss. Participants were naive with respect to synthetic sentence material. Subjects were required to achieve 100% accuracy during practice with ten commercially recorded synthetic sentences presented in quiet before further participation in this study. An audiometric screening form is given in Appendix A. SPEECH MATERIALS Eighty synthetic sentences were selected from lists gen- erated by Chial (1978). These sentential materials were 12 composed following the procedures of Miller (1951) and Speaks and Jerger (1965) and were used for two reasons. First, alternative and equivalent SSI sets were needed to avoid stimulus learning effects that might accompany subject training. Such learning effects might have constituted a Type II bias. Second, real clinical situations probably do not involve the lengthy training used by Dirks and Bower. Thus, it was desirable that subjects were similarly "naive” in the stimuli used in any given experimental condition. The present research used third-order approximations to real sentences based on the conditional probabilities of word-triplets. The lists generated by Chial were constructed using the following procedures. A word-pair was randomly selected from the Thorndike and Lorge count of the 1000 most common words (Thorndike and Lorge, 1944). An individual selected a third word to follow the original pair. Next, the first word of this triplet was masked and a second indi- vidual selected a new "third" word. This process continued (using a new individual for each subsequent "third" word) until a sentence was constructed that contained at least seven words. Chial (1978) used the following seven criteria to maintain homogeneity among the items in the lists generated: 1. Each sentential sequence included seven words, each of which must be from the Thorndike and Lorge (1944) corpus. l3 2. Each sentential sequence had between eight and ten syllables. 3. No sentence segment existed within a given senten- tial sequence. 4. No sentential sequence contained repetition of words. 5. Each sentential sequence had no meaning as determined by three different readers. 6. No two sequences in a given set of ten began with the same word or phoneme. 7. No two sequences in a given set of ten ended with the same word or phoneme. Third-order approximations are closer to "real” English sentences than are first- and second-order approximations because the sequence of words follows Specific rules of syn- tax more closely than randomly selected single words or word- pairs. Eight sentence lists (each of which fulfilled the criteria enumerated above) were generated by combining indi- vidual sentences from those developed by Chial (1978). A pilot study (see Appendix B ) produced the Six lists used in the experiment. These Six lists are reported in Appendix C. The competing message was taken from a passage published in The Reader's Digest (Hauser, 1973, freely edited). A transcript of this passage is given in Appendix D. The same reading was used for each of the six competing messages 14 developed (three recorded in the forward mode and three recorded in the backward mode). Talkers The primary and competing messages were recorded by three talkers whose speech differed qualitatively. They are described as follows: 1. One adult male with a General American dialect recorded both primary and competing messages in a conversational style. 2. One adult male with a General American dialect recorded the competing message in a radio broad- cast style. 3. One adult female with a residual Southern dialect recorded the same competing message recorded by the other two talkers in a conversational style. Therefore, Talker l differed from Talker 2 along the dimen- sion of speaking style, and Talker 3 differed from Talkers l and 2 along the dimensions of pitch and dialect. Recording Equipment All stimuli were recorded in a sound-treated chamber (IAC, Model 402) using a dynamic microphone (Electrovoice RE-l6), low-noise cassette tape (3M 176 Tenzar) and a cas- sette tape recorder (Nakamichi Series II 700). Talkers monitored overall voice level via a remote VU meter and ear- phones. The signal was routed from the microphone to an 15 audio mixer (Teac 2) and an electonic switch (Coulbourn 884-4) before reaching the cassette tape recorder to mute inter-stimulus noise levels. Talkers were alerted visually by a cue lamp to the timing sequence provided by control logic devices (Coulbourn Instruments, Inc.). Cassette recordings were made using Dolby noise reduction techniques. Recording Procedures Six lists of synthetic sentences were recorded by one male speaker (Talker l). The same competing message was recorded by each of the three talkers. Each forward com- peting message (CMF) was then rerecorded in the backward mode (CMB) to generate the remaining three competing mes- sages. Each competing message (CMF) was rerecorded on a reel-to-reel recorder (Ampex AG 500) using professional mastering tape (Ampex 631). This recording was then played in reverse and dubbed onto run tapes using a cassette tape recorder and low-noise cassette tapes. A 1000 Hz level calibration tone preceded each recorded sentential and continuous discourse message. This tone was equated in level to the root-mean-square (RMS) amplitudes of the primary and competing messages, as monitored by a true RMS voltmeter (Bruel and Kjaer Model 2606). l6 APPARATUS SSI stimuli were reproduced by one cassette tape recorder (Nakamichi Series II 700) and the competing mes- sage signal was reproduced by another cassette tape recor- der (Marantz 5020). The outputs of the two tape recorders were routed to separate input channels of a speech audio- meter (Grason-Stadler Type 162). The two signals were then mixed and routed to each of two earphones (TDH 49 with MX4l/AR cushions) mounted in separate TC-89 headbands. A dummy earphone/cushion was placed in each headband. Thus, two subjects could be tested Simultaneously. The speech audiometer facilitated mixing and level control of each signal so that different MCRs could be established. The cassette recorder which reproduced the sentence signal was linked to a digital—to-analog converter (CoulbournfiS77-08). This device triggered a set of logic devices (Coulbourn Instruments, Inc.) that activated subject cue lamps when a sentence signal was present. Figure 2 depicts this apparatus. CALIBRATION PROCEDURES The equipment was checked for absolute level output, frequency response, amplitude linearity, and amplitude dis- tortion. A 1000 Hz sinusoid was introduced at the input to the audiometer, and the output was measured with an octave- band sound level meter (Bruel and Kjaer Model 4144) by means of a NBS 9A, 6-cc coupler (Bruel and Kjaer Model 4152). 17 .moucoucom umuocucxm mo :oHumumwfiucoc_ pew m:umpmaam Hmucoewpoaxo we Emhwmmw xuofim .N beam“; ”8:8 ufis I Gama nwémm + III 05" 35:8 322:8 .312 "E. -8433: 2! E Sflauflh U / 71 I/ in .6 H - / /.‘L 0min£ n uuonfluw 7 / U593 39: 8933 l/ 7/ f} cassfls W WIT 5...... , Lafifia canned: blur:— - can... «a: 0333 ~ “0636 LSXQa algae flHHHuATIJ i 7//E¥/// ffZZf/jf/ 18 Reference level calibration was accomplished with a sound level calibratOr (Bruel and Kjaer Model 4320). The fre- quency response of the audiometer-earphone system was sim- ilarly measured using a swept sine wave source (Bruel and Kjaer Model 1024), pressure microphone, acoustic coupler, sound level meter, and graphic level recorder (Bruel and Kjaer Model 2305). Amplitude linearity and harmonic dis- tortioniwne checked following the procedure outlined in ANSI 53.6-1969. The frequency response of the tape cassette recorders was determined using a calibration tape (Standard Tape Lab- oratory, Inc., No. 26) and a swept signal produced by a sine generator (Bruel and Kjaer Model 1024). Before each subject pair was tested, the output level of the speech audiometer was calibrated electronically by measuring the RMS voltage produced at the terminals of the earphones by the recorded level calibration tone. EXPERIMENTAL PROCEDURES Figure 3 summarizes the sequence of experimental pro- cedures. Subjects were run in teams of two members each. The two subjects were seated in a sound—treated chamber (IAC Model 1200A). The initial part of the test session was devoted to practice. Two pre-recorded lists of the SSI (Auditec; 1978) were used for practice sessions. By using pre-recorded SSI stimuli for practice, subjects were fam- iliarized with the task and type of stimuli. Subjects did 19 .mouzonOHQ Hancoeweogxo mo “peso 30~u .m opswwu 20 not practice with the specific experimental message sets. Participants listened to standardized task instructions before identifying SSI stimuli. One list of ten synthetic sentences (SSI) was presented in quiet to the right ear at a sound pressure level (SPL) of 40 dB (re: 20u Pa; Dirks and Bower, 1969). A second list of ten synthetic sentences was presented, at a SPL of 40 dB with a competing message (CMF) at an MCR of ~20 dB (re: sound pressure level of primary message). The same talker read both the synthetic sentences and the competing message. The subjects had five seconds to identify the sentences on response sheets placed in front of them. Listeners were required to correctly identify each stimulus item presented on the first SSI list, thus insur- ing that they understood the listening task. Dirks and Bower (1969) had their subjects respond via pushbuttons. This was not possible in the present study: instead, written response sheets were used. Response sheets were checked for accuracy before the experimental session com~ menced. The experimental session immediately followed the practice session. Twelve subjects were randomly assigned to three listening groups, corresponding to MCRS of ~12, ~18 and ~24 dB (re: the SPL of the primary message). All three of the MCRS were used by Dirks and Bower (I969). The two more negative values were within the linear portion of the 21 of the performance-intensity (Pl) function reported by Dirks and Bower and reproduced above on page 6. Subjects listened to six lists of synthetic sentences and six competing mes- sages at one MCR. Three lists of synthetic sentences were presented to each subject with a forward competing message (CMF) and three lists of synthetic sentences were presented with a reversed competing message (CMB). The three CMF and three CMB conditions included listening conditions of (a) a male talker for primary and competing message, (b) a male talker for the primary message and a different male talker for the competing message, and (c) a male talker for the pri~ mary message and a female talker for the competing message. The synthetic sentences were presented at a SPL of 40 dB and the SPL of the competing message was varied to obtain the three MCRS. A typed list of ten synthetic sentences (ordered differently from the recorded sentential message) was placed before each subject. Each of the printed sen- tences was assigned a number (#1 ~ #10) for identification purposes. A subject cue lamp was mounted in front of each listener and lighted each time a sentence was read. After each sentence was heard, the subject had five seconds to respond by entering on a response sheet the number of the sentence just heard. A different response form was used for each of the six test lists. A five-minute break divided administration of the first three and the last three condi- tions. 22 Each subject was debriefed in an attempt to gather data concerning use of listening strategies. This effort consti- tuted less formal procedures to determine what, if any, listening strategies were employed. The experimenter asked and noted subject responses to the following questions: 1. Did you find yourself adopting any sort of listen- ing strategy during this session? 2. If so, why? 3. If so, what was your strategy? 4. If so, how effective do you think your strategy was? 5. In general, how distracting was (were) the com- peting message(s)? 6. Which competing message or messages were least distracting? 7. Which competing message or messages were mpg; distracting? A subject debriefing protocol is shown in Appendix F. Listening conditions were counterbalanced within MCR conditions. The serial order of CMF and CMB were alternated. Direction of the first competing message presentation (for~ ward vs. backward) was counterbalanced. The six sentential sets were randomly assigned to competing message conditions at each MCR. Further randomization of test lists occurred within each MCR group, and six subjects received one order of list presentation; the remaining six subjects received a second order of list presentation. Appendix G depicts count- erbalancing and randomization schedules for the experiment. 23 Three subjects were randomly selected from each MCR group to provide test-retest data pertaining to experimental effect. The nine subjects returned for a second practice and experimental session between two and five days following the first session. Listeners received identical experimental procedures during both listening sessions. RESULTS INTRODUCTION The present research was conducted to evaluate the performance of normal listeners in a synthetic sentence identification task when the same and different talkers provided the primary and competing messages. Thirty-six bilaterally normal listeners were separated equally into three message-to-competition ratio groups (~12 dB, ~18 dB, and ~24 dB re: sound pressure level of the primary message) before participating in the identification task. Subjects practiced with commercially recorded synthetic sentences (SSI) prior to identifying the experimental stim- uli. Listeners were then presented with six lists of syn- thetic sentences read by a male talker (Talker 1). Each sentence list was ipsilaterally paired with a different connected discourse competing message. The competing mes- sages were recorded by (a) the same male talker who recorded the sentences, (b) a different male talker, and (c) a female talker and were reproduced in both forward and reverse modes. Listeners identified each sentence heard by recording the list item identification number on a response sheet. Sub- jects heard each sentential and continuous discourse signal 24 25 at a constant message-to-competition ratio (MCR) as deter- mined by random assignments to MCR groups. At the conclusion of the identification task, subjects were debriefed to obtain subjective information concerning use of listening strategies. Nine subjects were randomly selected to participate in a second experimental session to generate test-retest data. Three subjects from each MCR group received identical experi- mental procedures during the second session. The following experimental questions were 1. Does mean normal listener performance task vary significantly as a function arity of the competing message source message source? 2. Does mean normal listener performance task vary significantly as a function asked: in an SSI of the simil- to in of the primary an SSI the presence vs. absence of semantic content in the competing message? 3. Does mean normal listener performance in an 881 task vary significantly as a function of interactions among (a) the similarity of primary and competing message sources, and (b) the presence vs. absence of semantic content in the competing message? 4. Does mean normal listener performance in an SSI task vary significantly as a function of interactions among 26 (a) the similarity of primary and competing message sources, and (b) message—to-competition ratio? 5. Does mean normal listener performance in an SSI task vary significantly as a function of message-to- competition ratio? DATA REDUCTION Subject performance on the synthetic sentence identifi- cation task was assessed with percent-correct responses. Outcomes of debriefing procedures were determined by categor- izing responses to direct questions and by statistically sum- marizing responses to items that employed Likert-type response scales. Scoring of subject responses Individual subject data were generated by determining percent-correct scores for each of the Six sentence lists. Ten percent of the 216 subject response forms were randomly drawn and subjected to a second scoring to provide a check on the accuracy of data reduction. This check indicated 100% agreement between the first and second scoring. Categorization of debriefing protocols Data from listener debriefing protocols were categori- zed according to three different procedures. First, subject responses to questions regarding listening strategies were evaluated with a content analysis to identify frequency of occurrence of specific strategies. Second, Likert scale 27 responses concerning the effectiveness of listening strate— gies were summarized with means for each strategy. Last, subjective responses about the perceived distractiveness of competing messages were tabulated in the form of frequency distributions. STATISTICAL PROCEDURES Listener performance was evaluated with percentage of sentences correctly identified for each listening condition. Means, standard deviations, and ranges were computed for all listening conditions. Individual subject scores are given in Appendix F. Analysis of these data followed a mixed effects, three- way analysis of variance (ANOVA) with repeated measures on the factor of presence versus absence of semantic content and on the factor of primary message talker similarity (Linton and Gallo, 1975, pp. 246-259). The remaining factor (MCR) was nested, non-repeated measure. The analysis per- mitted evaluation of the significance of mean differences as a function of the main effects of a) message-to-competition ratio, b) competing message direction, and c) competing mes- sage talker, as well as three two-way interactions and one three-way interaction. In addition to the usual ANOVA sum- mary data, computations were made of exact estimates of the probability of Type I error for observed F-ratios (for F-rat- ios significant at or beyond Paso.05) and, w2, a strength of association index. This last statistic indexes the 28 proportion of variance in the dependent variable that can be "explained by” the significant independent variables (main effects or interactions). Where appropriate by virtue of significant F—ratios, specific comparisons among means were accomplished using the Newman-Keuls' specific comparisons test (Linton and Gallo, 1975, pp. 347-352). Test-retest reliability of subject performance was assessed by computing Pearson product-moment correlation coefficients (Linton and Gallo, 1975, pp. 347-352) based upon data for three subjects from each MCR group. Correlation coefficients were determined for each of the six listening conditions (Paso.20). Listening condition coefficients from each MCR were transformed to z-scores, summed, then divided by the number of correlations. This average z-score was retransformed to a correlation, yielding an average correl- ation across six listening conditions at each MCR. SYNTHETIC SENTENCE IDENTIFICATION DATA Description Table 1 summarizes subject performance for the identifi- cation task. For the ~12 dB and ~18 dB MCR groups, the back- ward-same competing message listening condition yielded the highest mean scores (83.3%, 76.7%, respectively), whereas the backwards-different1 competing message listening condition produced the lowest mean scores (68.3%, 61.7%, respectively). The forward-same listening condition also yielded a mean score of 68.3% for the ~12 dB MCR. In the ~24 dB MCR group, the 29 Table 1. Mean percent-correct scores, standard deviations, and ranges based upon single observations for three groups of 12 normal-hearing subjects in each of six listening competing message source and direction conditions for three message-to-competition ratios. LISTENING CONDITION* FS FD FD BS BD BD 1 2 1 2 “CR % 68.3 80.0 80.0 83.3 68.3 82.5 -12 dB 8.0. 24.4 13.5 19.5 21.5 28.0 20.5 (N = 12) Range 90.0 40.0 60.0 60.0 90.0 70.0 % 75.8 70.8 69.2 76.6 61.7 67.5 ~18 dB 5.0. 23.1 33.2 23.5 27.7 27.2 27.7 (N = 12) Range 80.0 100.0 90.0 90.0 100.0 100.0 % 53.3 50.0 46.7 67.5 41.7 36.7 —24 dB 5.0. 16.7 22.6 20.2 16.6 12.0 20.2 (N = 12) Range 60.0 70.0 70.0 60.0 50.0 70.0 *"F" denotes forward, "B" denotes backward. S refers to the condition where the same talker produced the primary and competing messages. D designates different talkers for two messages: subscript 1 denotes different male talkers; D2 denotes female talker. 30 highest mean score (67.5%) was again produced by the backward— same listening condition, whereas the lowest mean score with the backward-different2 competing message listening condition. Figure 4 summarizes relations among the six competing message listening conditions. For each MCR group, the back- ward-same listening condition offered the highest scores. Average subject performance was Similar for the ~12 dB and ~18 dB MCR groups; performance levels declined for the ~24 dB MCR group. Figures 5 and 6 illustrate the relations between mean scores for the six listening conditions. Figure 5 depicts subject performance with three competing messages presented in the forward mode. Mean scores ranked (from high to low) as follows: same talker on the competing message, different male talker on the competing message, and female talker on the competing message. This was the case for both the ~18 dB and ~24 dB MCR groups but not for the ~12 dB MCR group. For the latter MCR group, the female talker and different male talker competing messages had equal performance levels, with the same talker competing message at a lower performance level. Figure 6 describes the relations among subject per- formance and three competing messages presented in the back- ward mode. For the ~12 dB and ~18 dB MCR groups, scores ranked (from high to low) as follows: same talker on the competing message, female talker on the competing message, and different male talker on the competing message. For the ~24 dB MCR group, competing message sources ranked as follows: 31 .owumu :ofiuwuonsou-ou-ommmmme mo :ofluucsm a ma maofluflvcou :ofiauuhfiw was oUASOm ommmmoa mcfiuoqaou wcflcoumfla xflm we some a“ muuonnsm w:wpmo:-~msho: NH mo museum omega pom meoHum>homno oqwcwm com: woman mohoum pumhuou-u:ouuun can: .v unswem axe- n 6: 5.3- n 62 82. u 6: 88 a 888 d 88 %%%%%S %%Sfi%5 %%fl%%$ § BEAN PERCENT-(DIRECT stES 100 90 80 70 6O 50 40 30 20 10 32 I I T L a - 'I I— - KEY: same talker on (M H - different male talker on GM H “ female talker on (M H l l _L -12dB -18dB -24dB MIR Figure 5. Mean percent-correct scores based upon single observations for 12 normal-hearing subjects for carpeting message conditions of forward-same, forward-different , and forward-different2 as a function of message-to-conpetition ratio. 33 100 I I T 90 - .4 801- - \3 O I l O‘ o I I «b O I I MEAN PERCENT- CORRECT SCORES 8 | l M O 1 I N o I I KEY: same talker on QVI H 10- different male talker on (N H - female talker on GM ‘ ‘ I l l G ~12dB ~18dB ~24dB MIR Figure 6. Mean percent-correct scores based upon single observations for 12 normal-hearing subjects for competing message conditions of backward-same, backward-different , and backward-different2 as a function of nesslge-to-competition ratio. 34 same talker on the competing message, different male talker on the competing message, and female talker on the competing message. Figures 7, 8, and 9 show the effects of semantic content on mean scores. When the talker was the same for both stim- uli (Figure 7), the backward competing message consistently produced higher means across the three MCR groups. When dif- ferent male talkers were used (Figure 8), the forward compet~ ing message yielded higher means at each MCR. When a female talker produced the competing message (Figure 9), listener performance was very close for the ~12 dB and ~18 dB MCR groups, whereas for the ~24 dB MCR group, performance levels were higher with the forward competing message. Reliability of measurement Test-retest correlation coefficients were computed for each of six competing message conditions within each MCR group (see Table 2). Significant correlations were those which exceeded 0.48 at Paso.20. For the ~12 dB MCR group (see Figure 10), correlations ranged from +0.56 (backward-different1 competing message) to -0.76 (forward-same competing message) and yielded a mean r of +0.37. This average denotes low reliability across all listening conditions for ~12 dB MCR. Figure 11 depicts correlations generated by listeners in the ~18 dB MCR group. In this instance, coefficients ranged from +0.99 (backward-different competing message) to ~0.70 1 3S 100 —T I I 90 - .. 80 _ J g 70 .. .. 8 U) B 60 h- - 8 50— _ E g 30L. J 20-~ ~ KEY: 10- same talker on CMF H .. same talker on CMB H o J l l ’l ~12dB ~18dB ~24dB MIR Figure 7 . Mean percent-correct scores based upon single observations for 12 neural-hearing subjects for competing message conditions of forward-same and backward-same as a function of message-to- conpetition ratio. 36 100 90.. _ \l o I l O\ O I J .5 C I A m PERCENT-(DRRBCI‘ SCORES 8 I I 30 - .. 20 - .. KEY: 10 _ different male talker on CMF H + different male talker on (NB I. o l l l ~12dB ~18dB . ~24dB MIR Figure 8. Mean percent-correct scores based upon single observations for 12 normal-hearing snbjects for conpeting message conditions of forward- different and backward-differentl as a function of messagé-to-conpetition ratio. 100 37 ,O‘ \1 on :0 o o o o MEAN PERCENT - mRRECT SCORES U1 o 40 —- .4 30 — _ 20 >— _ KEY: 10 _ female talker on CMF H fl female talker on (NB H 0 l I L ~ 12dB - 18dB - 24dB MIR Figure 9. khan percent correct scores based upon single observations for 12 normal-hearing subjects for competing message conditions of forward~ different and backward-different2 as a function of messag ~to~conpetition ratio. 38 Table 2. Test-retest correlation coefficients (Pearson r) and "average” coefficients for nine subjects on the synthetic sentence identification task under Six listening competing message source conditions. LISTENING CONDITION rs F01 F02 BS 801 802 Xr MCR (N = 3) -12 dB r ~.76** +0.28 +0.11 -0.20 +0.56** +0.12 +0.37* (N = 3) ~18 dB , r +0.87** -0.70** -0.20 -0.20 +0.99** +0.90** +0.80** (N = 3) -24 dB * r +0.99** +0.69** +0.80** +0.87** +0.28 +0.76** +0.84** *r to z transformations interpolated from Fisher's z transformation (Bruning and Kintz, 1977, pp. 250-251). **Significant beyond Pas0.20 level (df = l; rcritical = 0'480)' 39 100 100 I 1 I I L L H N H N 83 E E 22 52. EB LISTENING CONDITION Figure 10. Pearsonproduct-nnment correlations based upon test-retest data for three mural-hearing subjects for six listening competing message source conditions at a message-to-competition ratio of ~12 dB. 40 100 80.. 60— 409- 20- 20- 60.. _ 100 I 4 I 4 I 4 H N v-4 N E E E 28 E E LISTENING CONDITION Figure 11. Pearson product-mm correlations based upon test-retest data for three normal-hearing subjects for six listening competing message source conditions at a message-to-competition ratio of ~18 dB. 41 (forward-different1 competing message), with a mean of +0.80. This indicates high reliability from Trial 1 to Trial 2. Correlation coefficients for the ~24 dB MCR group are shown in Figure 12. The r scores ranged from +0.99 (forward- same competing message) to +0.28 (backward-different compet- 1 ing message) and yielded a mean r of +0.84. This was the only case where all coefficients were of positive sign and, except for the backward-different1 competing message, revealed marked reliability for the remaining five listening conditions. For the ~24 dB MCR group, the correlation mean truly describes the Strength of relationship between Trial 1 and Trial 2 data. Mean correlation coefficients across listening condi- tions are described in Figure 13. As the MCR became more negative, the reliability increased considerably. This indi- cates that generally, the ~18 dB and ~24 dB MCR conditions were more consistent than the ~12 MCR condition. Statistical analysis Table 3 summarizes the results of a mixed effects, three-way analysis of variance of synthetic sentence identi- fication data. Significant F-ratios exceed critical values at the 0.05 level. The main effect of talker difference was significant, indicating that the three competing message talkers differed significantly from one another. The main effect of semantic content was not significant in this anal~ ysis. The talker difference-by-semantic content interaction 42 100‘ I I I I I 20 — - 4o *- a 60 — a 80 - _ 100 ‘ ‘ 4 1 i ' H N H QN E E E E E on LISTENING CONDITION Figure 12. Pearson prodmt-moment correlations based won test-retest data for three normal-hearing sub'ects for six listening competing message source con itions at a message-to-competition ratio of ~24 dB. 100 43 80 60 40 20 20 40 6O 80 L 100 I L - lZdB ~ 18dB - 24dB FER Figure 13. khan Pearson product-unment corre1ations based upon test-retest data for three groups of three normal-hearing subjects across listening conditibns as a function of message-to-competition ratio. 44 Table 3. Analysis of variance for synthetic sentence identification indexed by percent-correct scores as a function of message-to-competition ratio, semantic content, and talker. SOURCE DF 55 Ms Fobs. Fa'.05 p(F)obs. w? Iglgfifigti) 35 79800.0 IL‘R (Message- _4 to-competItion 2 30186.11 15093.10' 10.04** 3.30* 6.3 x 10 0.356 ratio) ERROR 1 33 49613.90 1503.50 Yézu§gcfs) 180 64333.33 ERROR 165 52736.11 Sgn§::T?"tic 1 46.30 46.30 0.12 4.15* 7 x 3'1 MLR x sc 2 256.48 128.24 0.34 3.30* 7.2 x 10‘1 ERROR 2 33 12630.56 382.74 I (Talker) 2 3108.33 1554.17 6.10** 3.14* 4.1 x 10’3 0.018 MIR 1: r 4 3538.89 884.72 3.46“ 252* 1.3 x 10’2 0.017 ERROR 3 66 16886.11 255.85 sc x r 2 3623.15 1811.57 5.15** 3.148 8.5 x 10'3 0.020 MCR 11 SC 11 r 4 1024.07 256.02 0.73 2.52* 5.8 x 10'1 ERROR 4 66 23219.44 351.81 TOTAL 215 144133.33 *Interpolated from F-table (Winer, 1971, pp. 864-869). **Significant beyond the Pu- 0.05 level of significance. 45 was significant, indicating that talker differences were not independent of semantic content. The talker difference-by— MCR interaction was also significant. This suggests that differences among talkers were not independent of MCR and vice versa. In addition, the main effect of MCR was found to be Significant. The remaining interactions were not significant. Comparisons among individual means were made with the Newman-Keuls' specific-comparisons test. Paired comparisons were made for three percent-correct means representing same, differentl, and different2 competing message source condi- tions for each MCR and vice versa. Table 4 describes the following Significant differences (Th3 0,05)between means: S versus D2 and D1 versus D2 for ~12 dB MCR; S versus D2 and S versus D1 for ~18 dB MCR; and 8 versus D1 and S versus D for ~24 dB MCR. 2 Table 5 depicts comparisons of mean values of ~12 dB, ~18 dB and ~24 dB MCR groups for each competing message source. Significant differences were found between all means tested except for the ~12 dB MCR mean versus the ~18 dB MCR mean for the same talker condition. Comparisons of means for presence and absence (F vs. B) of semantic content and for the three competing message source conditions are shown in Table 6. Significant differ- ences were only found with the absence of semantic content between the D2 and D1 competing messages. All remaining com- parisons were nonsignificant at the 0.05 level. 46 Table 4. Newman-Keuls' specific-comparisons test results for three competing message source means (collapsed across the factor of presence vs. absence of semantic content) for each of three MCRS. Critical values are given for all possible ranges (from two means apart to three means apart). A difference between any two means is significant when it exceeds the appropriate critical value. The level of Significance was Peso.05. MEANS 02 S 01 [~12 dB MCR] (81.25) (75.83) (74.20) * * 02 5.42 7.05 cv3 .40 S 1.63 cv2 .32 D1 MEANS S 02 01 [-18 dB MCR] (76.25) (68.33) (66.25) S 7.92* 10.0* cv3 .40 02 2.08 cv2 32 D1 MEANS s 01 02 [—24 dB MCR] (60.42) (45.83) (41.67) S 14.59* 18.75* cv3 .40 01 4.16 cv2 .32 D2 *Denotes significant comparisons. 47 Table 5. Newman-Keuls' specific-comparisons test results for three MCRS (collapsed across the factor of presence vs. absence of semantic content) for each of three competing message sources (S, D , D ). Critical values are given for all possible rangAs (from two means apart to three means apart). A difference between any two means is significant when it exceeds the appropriate critical value. The level of significance was PaS0.05. MEANS ~18 dB MCR ~12 dB MCR ~24 dB MCR [S] (76.25) (75.83) (60.42) ~18 dB MCR 0.42 15.83* CV3 .40 ~12 dB MCR 15.4l* CV2 .32 ~24 dB MCR MEANS ~12 dB MCR ~18 dB MCR -24 dB MCR [D1] (74.20) (66.25) (45.83) ~12 dB MCR 7.95* 28.37* CV3 .40 ~18 dB MCR 20.42* CV2 .32 -24 dB MCR MEANS -12 dB MCR ~18 dB MCR ~24 dB MCR [D2] (81.25) (68.33) (41.67) ~12 dB MCR 12.92* 39.58* CV3 .40 -18 dB MCR 26.66* CV2 .32 -24 dB MCR *Denotes significant comparisons. 48 Table 6. Newman-Keuls' specific-comparisons test results for three competing message sources (collapsed across the factor of MCR) for each of two levels of semantic content (F vs. B). Critical values are given for all possible ranges (from two means to three means apart). A difference between any two means is significant when it exceeds the appropriate critical value. The level of significance was PoS0.05. MEANS D S D 1 2 [F] (66.90) (65.83) (65.27) 01 1.07 1.63 cv3 = 6.40 S 0.56 cv2 = 5.32 D2 MEANS S 02 01 [Bl (75.83) (62.22) (57.22) 5 13.61* 18.61* cv3 = 6.40 02 5.0 CV2 = 5.32 D1 IDenotes significant comparisons. 49 Strength of association between the independent variables and dependent variable (percent-correct) was assessed by calculating w? for significant effects. The MCR factor yielded an w2 value of 0.356, indicating that differences among the three MCRS accounted for 35.6% of the score variance. For the talker factor, 02 was 0.018, indi- cating that only 1.8% of the variance in scores could be attributed to differences in talker. The w2 value for the MCR-by—talker interaction was 0.017 or 1.7%, again account- ing for a small amount of variance in scores because of this interaction. For the semantic content-by-talker interaction, O? was 0.020, indicating that 2% of the variance in scores could be accounted for by this interaction. Outcomes of the inferential analysis revealed that the factor of MCR most significantly affected normal listener performance in an SSI task. In all instances, mean scores decreased as MCR became more negative. This main effect yielded the most significant F-ratio, highest value of w2, and frequent significant mean differences. The main effect of talker difference was significant, as mean scores did vary across competing message source conditions. The sig- nificant interaction of MCR and talker difference can be attributed mainly to the MCR main effect, since Figures 5 and 6 describe only one transverse interaction apiece. The factor of semantic content did not produce a significant F-ratio, although interactions with the other two main 50 effects were significant. No significant differences in mean scores were observed as a function of presence or absence of semantic content. DEBRIEFING DATA Description of results The content analysis of subject responses concerning listening strategies is described in Table 7. After a lis- tening session was completed, subjects were asked if they had consciously developed a listening strategy during the course of the experiment and, if so, to describe that strat- egy in their own words. Subjects were not prompted to respond with any specific exemplary strategies. The most frequently reported strategy was that of listening for the first or last words in test sentences (N = 15). The next most frequent strategy was that of listening for subject- determined key words (N = 9). The remaining strategies were employed by single subjects or by small groups. A total of 24 subjects claimed that they used a Single strategy during the listening session, 12 subjects reported using two strat- egies concurrently, and one subject claimed the use of four strategies to identify the synthetic sentences. Table 7 also summarizes Likert scale data by which sub- jects rated the "mean effectiveness" of their listening strategies. The scale consisted of five equal-appearing intervals placed onzacontinuum ranging from "very ineffective” to "very effective." Responses for each listening strategy Table 7. Frequency table Of first responses by 36 normal- hearing subjects in response to debriefing strategy questions. STRATEGY DESCRIPTOR FREQUENCY MEAN JUDGED STANDARD EFFECTIVENESS RATING DEVIATION {2:2t;§¥dfir5t or 15 2. 93 0.80 "Listened for vowels” l 3.0 - "Listened for /s/” 2 3.0 1.41 ”Tune out speaker" I 3.0 - ”Listened for one word" 3 3.7 (1.58 "Scanned and reread 2 2.5 2.12 sentences "Listened for key words" 9 2.78 0.83 "Random pauses of CM" 1 3.0 ' "No strategy" 2 ' ' _36—. R53.10(n=34) 0.93 52 were averaged to produce the mean ratings. The Strategy of listening for one word was judged to be most effective (mean of 3.7) by three subjects. The strategy rated as most inef- fective was that of scanning and rereading the sentences (mean of 2.5). The most effective strategy had the smallest standard deviation (0.58), whereas the least effective strategy had the largest standard deviation (2.12). Subjects were also asked to indicate which competing message was subjectively the most distracting and which was the least distracting. These judgments are summarized in Table 8. Responses concerning the least distracting and most distracting competing messages were tabulated only if the response contained information about both talker and direction. No attempt was made to secure judgments that distinguished between the two male talkers. There was greater discrepancy among subjects' responses regarding the least distracting competing message than for the most distracting competing message. The forward-male competing messages were most often rated as least distracting (N = 14). The remaining competing messages were designated less often as the least distracting competing message. Sub- ject selection of the most distracting competing message ‘wasevenly'distributed across the four combinations of direc- tion and source. This suggests that judgments about the most distractive competing message were not influenced in any systematic way by attributes of the messages. 53 Table 8. Cross tabulation of responses by normal-hearing_ subjects reporting on perceived distractiveness Of competing messages on the synthetic sentence identification task. LEAST DISTRACTING COMPETING MESSAGE Direction Forward Backward Male 14 3 2:17 Talker Female 7 8 2:15 Z=21 E=ll MOST DISTRACTING COMPETING MESSAGE Direction Forward Backward Male 8 6 £=l4 Talker Female 6 9 Z=15 54 SUMMARY Mean percent—correct scores were highest for the backward-same competing message listening condition for the ~12 dB, ~18 dB, and ~24 dB MCR groups. Subject performance levels were lower for the -24 dB MCR group than for either the ~12 dB or ~18 dB MCR groups. Comparison of means gener- ated by all six competing messages indicated that the same talker condition yielded the highest mean scores for both directions of competing message presentation. Mean test- retest correlation coefficients averaged across listening conditions for each MCR were +0.37 (~12 dB), +0.80 (~18 dB), and +0.84 (~24 dB): test-retest consistency was greatest for the most negative MCR. Significant F-ratios were obtained for main effects of talker difference and MCR. Significant main effect interactions were observed for the talker difference-by-semantic content and talker difference- by-MCR interactions. Approximately 36% of the variance in mean scores could be attributed to MCRS. The remaining significant factors yielded very small strength of associa— tion values. The listening strategy most often selected by subjects was that of identifying the first or last word of a test sentence. The forward male (same or different) com- peting message was rated as least distracting; no single competing message was consistently chosen as the most distracting. DISCUSSION INTRODUCTION The review and interpretation of the present findings focuses on several areas. The reliability of experimental results is discussed in terms of inter-list equivalence and test-retest consistency. Present outcomes are compared with previous findings. Possible clinical applications of the synthetic sentence identification task are discussed. Limi- tations of experimental method and design are examined and the experimental questions are discussed in the context of those limitations. Finally, recommendations are offered for future research with the synthetic sentence listening task. RELIABILITY Test List Equivalence Results of the pilot study conducted to evaluate inter- list reliability failed to provide unambiguous statistical evidence about which lists should be retained or discarded for experimental use. Therefore, the assumption was made that the stimulus materials were equivalent across lists. Two sentence lists (4 and 6) were arbitrarily discarded from the original group of eight lists because they Offered the lowest mean performance across subjects and the lowest absolute mag- nitude of test-retest correlation coefficients. SS 56 Each sentence list fulfilled the requirements for list homogeneity as described in Chapter II. Thus, the stimulus items were assumed homogenous across lists in terms of grammatical, lexical, and phonemic constraints. The present research used six different lists of syn- thetic sentences. Dirks and Bower (1969) and Jerger and Hayes (1976) employed only one list Of ten synthetic sentences as stimuli, organized in several different random orders. Subjects heard repeated lists containing identical sentences; no controls were implemented to avoid stimulus learning effects. The use of six different sentence lists in the pre~ sent study allowed subjects to be similarly "naive" about the stimuli used for each listening condition. Test-Retest Consistency Correlation coefficients for each listening condition within each message-to-competition ratio (MCR) indicated that the highest reliability was found for the ~24 dB MCR group (average r = +0.84). The average r observed for the ~18 dB MCR (+0.80) also indicated high reliability from Trial 1 to Trial 2, whereas the average coefficient for the ~12 dB MCR group (+0.37) indicated low reliability. Thus, as the MCR became more negative, subject performance became more consistent across trials. This outcome probably derives in part from the fact that the correlation coefficient assumes a linear underlying continuum, and the fact that the ~12 dB MCR listening condition approximated the asymptotic 57 portion of the psychometric function for this task. VALIDITY The Effects of Message-to-Competition Ratio and Semantic Content Present findings confirmed the findings of Dirks and Bower (1969) that as MCR became more negative, subject per- formance declined. The use of the three MCRS (~12 dB, ~18 dB and ~24 dB) partially describes listener performance in both studies. Dirks and Bower also presented the sentence lists at an MCR of ~30 dB. Specific comparisons of mean percent-correct scores from the present research to those of Dirks and Bower (1969) repli- cated Dirks and Bower's same talker paradigm on the primary andccmpetingnmssage. Whereas Dirks and Bower observed very Similar results for both CMF and CMB conditions, the present effort produced different results. Although the factor of semantic content was not significant in isolation, subject performance under the forward-same and backward-same listen- ing condition was not equivalent between the ~12 and ~24 dB MCR groups. The backward-same condition always generated higher performance levels than did the opposing forward competing message condition. Dirks and Bower may have used a single ten item set to minimize variance attributable to test lists. They probably succeeded in this attempt, but it is possible that their procedure introduced a "testing effect" productive of a Type II bias. It is impossible to verify the possible presence 58 of such bias in the data of Dirks and Bower, but if such bias were present, it might explain some of the difference between what was reported by Dirks and Bower and what was found in the present study. Similarly, Dirks and Bower trained their subjects extensively in the single set of stimuli used. Although training Should serve to reduce in- ter-and intra-subject variance, it might also facilitate development of listening strategies that obscure the effects of semantic content by reducing the potency of stimuli as novel linguistic events. P~I functions generated by the present data were gen- erally linear, with the exception of the same-talker forward competing message condition (Figure 7). For the CMF function, the lepes were ~l.3%/dB (~12 to ~18 dB MCR) and +3.8%/dB (~18 to ~24 dB MCR). The slopes Of the CMB function reported here were +1.0%/dB and +1.5% dB (respectively) for the same MCR intervals. Generally, these slopes are less steep than those reported by Dirks and Bower (see Table 9). They found P~I function slopes of +4.0%/dB for the CMF function and +4.3%/dB for the CMB condition, each across ~l7 dB to ~28 dB MCR. These are probably attributable to methodological dif— ferences between the two studies. The use of a single stim- ulus set and repeated Observations of subjects across MCRS ought to produce steeper slopes. These P~I slope and methodo~ logical differences suggest that subject training and/or the use of one stimulus set contributes to an easier listen- ing task, whereas the training and/or the use of different Table 9. 59 Psychometric function slopes compared three studies. from data reported in Dirks and Bower (1969) Speaks, Karmen and Benitez (1967) Present Study Competing Message Condition CMF CMB ~30 -34 ~12 ~18 ~12 ~18 MCR Range to ~16 dB to ~19 dB to ~18 dB to ~24 dB to ~18 dB to ~24 dB Slope +4. %/dB +4.3%/dB +3.3%/dB ~l.3%/dB +3.8%/dB +1.S%/dB 60 stimuli contributes to a more difficult listening task. Speaks, Karmen and Benitez (1967) also reported P~I functions for synthetic sentence identification with a same-talker competing message, although their methods differed appreciably from those used here and by Dirks and Bower. Speaks, and his colleagues used a set of 20 synthetic sentences divided into primary and secondary subsets of 10 sentences each. Subjects viewed a list of the ten sentences from the primary subset. Half of the sentences presented to listeners were from the primary subset, and half were from the secondary subset. Subjects judged whether a given stimulus came from the primary subset, and if so, which sentence from among the ten alternatives had been presented. Subjects were told that half of the sentences would be selected from the primary sub- set and half from the secondary subset; it was assumed that subject judgments reflected these a priori probabilities. The interpolated slope for the linear portion of the P~I function was approximately +3.3%/dB. Given the extensive differences in methods among the three studies summarized in Table 9, it is surprising that the variation in slopes was not greater. Differences among P~I function slopes give some indica- tion about variances in subject performance on the synthetic sentence identification task. In the present study, the range of percent-correct scores was considerable, arising at least in part from the use of "naive" subjects. Such 61 variance tends to produce somewhat flatter P~I functions than would be obtained with trained listeners. Because Dirks and Bower (1969) and Speaks, Karmen and Benitez (1967) did not report dispersion results, it is impossible to assess the contributions of variance to P~I function slopes. The Effect of MCR-by-Talker Difference Interactions The main effect of MCR was statistically significant (F 19.04), as was the main effect of talker difference (F 6.10). However, the interaction between factors of MCR and talker difference was also significant (F-ratio of 3.46). As MCR became more negative, the differences in performance for the backward-same and backward-different2 listening con- ditions increased. The less negative MCRS produced no sig- nificant differences as a function of competing message source, probably because these conditions did not sufficiently tax the central processes responsible for separating the primary message from the competing message. The Effect of Semantic Content-by-Talker Difference Interactions Although the effect of semantic content was not (by itself) significant, the interaction between semantic content and talker difference was (F = 5.15). Additional analyses were. conducted to evaluate this interaction (see Table 10). Because the factor of MCR was significant alone and in com- bination with talker difference, the semantic content-talker difference means were evaluated separately for each Of the three MCRS. For each MCR, six means were compared: each of three levels of talker for each of two levels of semantic 62 .mcomwhmmsou unmuwwwcwwm monocwpm N Ham mm.mH >u w.m Nam He.wH m>u m.a A.H Nam m~.om 3.5 H.@ m.m e.H Ham mm.HN w>u H.5H m.w e.e o.m ma mm.NN >u e.eH H.@ 4.5 m.m w.o mm hawaev mmweev mmwmev awWOAV Aw.mAV fie.eev Emu: me EH-H am am aa aa ma mm mz6 e.o ma He.wa m>u R.AH R.HH Ham mN.oN e>u R.HH R.HH e.o Noe mm.am w>u N.eH N.eH m.~ m.~ 8am mm.NN so c.mH c.mH w.m m.m m.e mm flmwmev Am.wev fiewowv acmewv ammeU am.mwv .euz me NH-L mm mm ma ma mm mm mzoH one .Osam> Hmufiuflpo Oumwpaopamm ecu mcoouxo pa can: ucmuflmflcwwm mw memos 03H %:m coozuon mono -hOMme < .appmam memos xmm ow Human memos 03p anmv momcmh oHnwmmom Ham pom :o>wm Ohm mosfim> Hmuwufieo .mmuz oumemmom omega pow moupsom pew monk“ ommmmee mcwuoasou xwm pom muHSmow umou mCOmwmeEou-6wwwuomm .mazox-:mEZOz .OH manme 63 .mEOmmuMQEou ucmummwamfim monocoam mm.mH Hv.wH mN.om mm.HN mm.NN >0 >0 >0 >0 \DLDQ’MN o.m o.oa m.mH 0.0H «w.cm fle.em0 om o.m m.w c.HH «m.m~ an HeV am m.m 0.0 «w.o~ fie eev ma m.nH he H omv ma N.vH Am.mm0 we Nam mm mm mm mm mm I-INI—I hm.eev Hmuz me e~-0 mm mz