xml version 1.0 encoding UTF-8 standalone no
record xmlns http:www.loc.govMARC21slim xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.loc.govstandardsmarcxmlschemaMARC21slim.xsd
leader nam Ka
controlfield tag 001 001967704
007 cr mnu|||uuuuu
008 081028s2007 flu sbm 000 0 eng d
datafield ind1 8 ind2 024
subfield code a E14-SFE0002137
Psychometric functions of clear and conversational speech for young normal hearing listeners in noise
h [electronic resource] /
by Jane Smart.
[Tampa, Fla.] :
b University of South Florida,
ABSTRACT: Clear speech is a form of communication that talkers naturally use when speaking in difficult listening conditions or with a person who has a hearing loss. Clear speech, on average, provides listeners with hearing impairments an intelligibility benefit of 17 percentage points (Picheny, Durlach, & Braida, 1985) over conversational speech. In addition, it provides increased intelligibility in various listening conditions (Krause & Braida, 2003, among others), with different stimuli (Bradlow & Bent, 2002; Gagne, Rochette, & Charest, 2002; Helfer, 1997, among others) and across listener populations (Bradlow, Kraus, & Hayes, 2003, among others). Recently, researchers have attempted to compare their findings with clear and conversational speech, at slow and normal rates, with results from other investigators' studies in an effort to determine the relative benefits of clear speech across populations and environments.^ However, relative intelligibility benefits are difficult to determine unless baseline performance levels can be equated, suggesting that listener psychometric functions with clear speech are needed. The purpose of this study was to determine how speech intelligibility, as measured by percentage key words correct in nonsense sentences by young adults, varies with changes in speaking condition, talker and signal-to-noise ratio (SNR). Forty young, normal hearing adults were presented with grammatically correct nonsense sentences at five SNRs. Each listener heard a total of 800 sentences in four speaking conditions: clear and conversational styles, at slow and normal rates (i.e., clear/slow, clear/normal, conversational/slow, and conversational/normal). Overall results indicate clear/slow and conversational/slow were the most intelligible conditions, followed by clear/normal and then conversational/normal conditions.^ Moreover, the average intelligibility benefit for clear/slow, clear/normal and conversational/slow conditions (relative to conversational/normal) was maintained across an SNR range of -4 to 0 dB in the middle, or linear, portion of the psychometric function. However, when results are examined by talker, differences are observed in the benefit provided by each condition and in how the benefit varies across noise levels. In order to counteract talker variability, research with a larger number of talkers is recommended for future studies.
Thesis (M.S.)--University of South Florida, 2007.
Includes bibliographical references.
Text (Electronic thesis) in PDF format.
System requirements: World Wide Web browser and PDF reader.
Mode of access: World Wide Web.
Title from PDF of title page.
Document formatted into pages; contains 73 pages.
Adviser: Jean C. Krause, Ph.D.
t USF Electronic Theses and Dissertations.
Psychometric Functions of Clear and Conversational Speech for Young Normal Hearing Listeners in Noise by Jane Smart A thesis submitted in partial fulfillment of the requirement s for the degree of Master of Science Department of Communication Sciences and Disorders College of Arts and Sciences University of South Florida Major Professor: Jean C. Krause, Ph.D. Gail Donaldson, Ph.D. Stefan Frisch, Ph.D. Date of Approval: July 10, 2007 Keywords: speech intelligibility, speakin g rates, talker variability, benefit size, intelligibility benefit, study comparisons Copyright 2007, Jane Smart
Acknowledgments I would like to thank my thesis commi ttee members, Dr. Jean Krause, Dr. Gail Donaldson and Dr. Stefan Frisch for their guidance and support. In particular, I would like to acknowledge my thesis advisor, Dr. Jean Krause, for her many hours of assistance, endle ss patience and constant encouragement. I would also like to thank all of the listeners for taking the time to participate in this study. Participant screenings and sessions could not have been completed without the help of Ann Siapno, an d I thank her for all of her cheerful assistance. Additionally, I would like to thank Ann and John Lum for their efforts with scoring listener responses. Finally, I would like to thank my family for their love and support in all my endeavors.
i Table of Contents List of Tables iii List of Figures iv Abstract v Chapter 1 Introduction 1 Chapter 2 Literature Review 5 Background 5 Acoustic Properties of Clear Speech 6 Role of Speaking Rate 6 The Clear Speech Benefit 8 Presentation Environments 9 Stimuli 11 Talkers 12 Populations 14 Summary 15 The Need for Psychometric Functions 16 Comparisons Across Studies 18 Statement of the Problem 20 Chapter 3 Methods 22 Participants 22
ii Materials 22 Presentation Sessions 24 Scoring Procedures 28 Data Analysis 28 Chapter 4 Results 30 Psychometric Functions 34 Effect of Talker 36 Psychometric Functions by Talker 40 Talker Variability with Condition 46 Chapter 5 Discussion 49 Reliability 50 Effect of Stimuli 51 Talker Variability 54 Listener Variability 55 Other Listener Issues 56 Future Work 58 Clinical Implications 60 References 62 Appendices 66 Appendix A: Participant Demographics 67 Appendix B: Key Word Scores by Talker 68 Appendix C: Pair wise Comparisons 72
iii List of Tables Table 1 Sentence Blocks 24 Table 2 Sentence Pres entation Schedule 26 Table 3 Between-subject Effects and Variables 31 Table 4 Individual talker intelligib ility benefits by condition and SNR 45 Table B1 T1 68 Table B2 T3 69 Table B3 T4 70 Table B4 T5 71 Table C1 Pairwise Comparisons by Condition 72 Table C2 Pairwise Comparisons by Talker 73
iv List of Figures Figure 1 Average intelligibility, in percent key words correct, for each test condition. 32 Figure 2 Average psychometric func tions, in percent key words correct, across signal-to-noise ratio. 35 Figure 3 Average intelligibility, in percent key words correct, for each talker in each condition. 37 Figure 4 Average intelligibility, in percent key words correct, for each talker across signal-to-noise ratio. 43 Figure 5 Intelligibility benefits by talker (T1 blue, T3 green, T4 aqua, T5 red) and SNR, rela tive to baseline performance, for each condition. Dotted line indicates maximum possible benefit for each baseli ne performance level. 47 Figure 6 Intelligibility benefits of clear /slow speech by talker for Liu et al. (a) versus the current st udy (b), (T1 blue, T3 green, T4 aqua, T5 red) and SNR, relative to baseline performance, for all listener s. Diagonal line indicates maximum possible benefit fo r each baseline performance level. 53
v Psychometric Functions of Clear and Conversational Speech for Young Normal Hearing Listeners in Noise Jane Smart ABSTRACT Clear speech is a form of communica tion that talkers naturally use when speaking in difficult listening conditions or with a person who has a hearing loss. Clear speech, on average, provides listeners with hearing impairments an intelligibility benefit of 17 percentage points (Picheny, Durlach, & Braida, 1985) over conversational speech. In addition, it provides increased intelligibility in various listening conditions (Krause & Braida, 2003, among others), with different stimuli (Bradlow & Ben t, 2002; Gagne, Rochette, & Charest, 2002; Helfer, 1997, among others) and across list ener populations (Bradlow, Kraus, & Hayes, 2003, among others). Recently, researchers have attempted to compare their findings with clear and conversational speech, at sl ow and normal rates, with results from other investigatorsÂ’ studies in an effort to determine the relative benefits of clear speech across populations and environments. However, relative intelligibility benefits are difficult to determine unle ss baseline performance levels can be equated, suggesting that listener psychomet ric functions with clear speech are needed. The purpose of this study was to determine how speech intelligibility, as measured by percentage key words corre ct in nonsense sentences by young
vi adults, varies with changes in speaking cond ition, talker and si gnal-to-noise ratio (SNR). Forty young, normal hearing adults were presented with grammatically correct nonsense sentences at five SNRs. Each listener heard a total of 800 sentences in four speaking conditions: clear and conversational styles, at slow and normal rates (i.e., cl ear/slow, clear/normal, conversational/slow, and conversational/normal). Overall results indicate clear/slow and conversational/slow were the most intelligible conditions, followed by clear/normal and then conversational/norma l conditions. Moreover, the average intelligibility benefit for clear/slow, clear/normal and conversational/slow conditions (relative to conversational /normal) was maintained across an SNR range of -4 to 0 dB in the middle, or linear, portion of the psychometric function. However, when results are examined by ta lker, differences are observed in the benefit provided by each condition and in how the benefit varies across noise levels. In order to counteract talker vari ability, research with a larger number of talkers is recommended for future studies.
1 Chapter 1 Introduction Clear speech is a verbal form of co mmunication that diff ers from typical Â“conversationalÂ” style speech. It is used naturally by speakers when they are trying to communicate in a noisy environment or with a person who is hearing impaired. One of the firs t research studies conducted with a Â“clearÂ” speaking style revealed a 17 percentage point intell igibility benefit of clear speech over conversational speech for hearing impair ed listeners (Picheny, Durlach, & Braida, 1985). This landmark study was the first of many research studies to show the substantial gains in intellig ibility from clear speech (B radlow & Bent, 2002; Helfer, 1997; Krause & Braida, 2002; Payton, Uc hanski, & Braida, 1994; Uchanski, Choi, Braida, Reed, & Durlach, 1996; among others). Picheny et al.Â’s findings led other researchers to further ex plore clear speech with listeners from different populations and in various listening env ironments. Researchers studied clear speech in quiet (e.g., Payton et al., 1994) and noise (e.g., Bradlow, Kraus, & Hayes, 2003), with normal (e.g., Kraus e & Braida, 2002), hearing impaired (e.g., Payton et al., 1994), native and non-native li steners (e.g., Bradlow & Bent, 2002), and children (Bradlow et al., 2003). All of these studies found an intelligibility benefit for clear versus conversational speech, demonstrating that the clear speech effect is sizeable and robust.
2 In addition to the perceptual benefits, clear speech is very easy for talkers to produce. Instructions are minimal and talkers generally become proficient with only 10 to 15 minutes of practice (Schum 1997). Talkers have been instructed to speak as if they were communicating in a noisy environment, with a listener who has a hearing loss (Helfer, 1997, 1998; Pi cheny et al., 1985; Schum, 1997), or with a listener whose native language di ffers from the speakerÂ’s language (Bradlow & Bent, 2002; Smiljanic & Bradl ow, 2005). Additional instructions may include specific directions for the speaker to articulate each word in a clear and precise manner (Schum, 1996). Beyond these instructions, neither the talker nor the listener requires any special training or abilities. Thus, clear speech provides an economical, practical and beneficial means of communication. As a result of the many recognize d benefits of clear speech, there are several useful applications. Early research focused on determining the clear speech effect and the acoustic differences that contribute to speech intelligibility in order to enhance signal-processing algo rithms for hearing aids (Picheny, Durlach, & Braida, 1985, 1986). By learni ng more about the properties of clear speech, the goal was to develop signal-pr ocessing schemes aimed at converting conversational speech to clear speec h, improving speech perception and resulting in better overall satisfaction by hearing aid users. Clear speech is also a valuable tool for use in clinical settings. The clinician who uses clear speech can be better assured that the client is receiving the communicative message accurately, with less need for repetiti ons. In aural rehabilitation settings, education about clear speech is routinely provided to the family members of
3 listeners with hearing loss (Schum, 1996). Th is practice provides an effective approach to communication for families a nd caregivers, resulting in fewer communication breakdowns and less frustrat ion for both speaker and listener. Moreover, clear speech can be useful in educational settings, such as with children with learning disabilities (B radlow et al., 2003), among others. The teacher who uses clear speech in the cl assroom can convey the dayÂ’s lessons with greater confidence that the children are receiving the message accurately. Overall, using clear speech in different settings can facilitate more effective communication. Given the widespread potential and many applications of clear speech, investigation of its benefits remains an active area of research. Recently, investigators have attempted to compare results between studies in order to synthesize the results of multiple inte lligibility experiment s with clear speech (Krause & Braida, 2002; Liu, Del Rio, Bradlow, & Zeng, 2004; Panagiotopoulos, 2005; among others). These com parisons are necessary to determine if the clear speech benefit received by different lis teners is the same across different environments or if some listeners rece ive a greater benefit in one or more environments. For example, clear speech intelligibility benefits found for older listeners (Panagiotopoulos, 2005) have been compared to benefits found for younger listeners (Krause & Braida, 2002) Unfortunately, comparisons between studies are generally difficult due to differences in presentation levels, environments, listeners and stimuli. W hen such differences exist, absolute performance levels are affected and confound comparisons of relative
4 performance. Consequently, studies reporti ng a clear speech benefit relative to conversational speech should not be com pared unless absolute performance is controlled. To facilitate such com parisons, psychometric functions, which characterize listener performance (plotted on the y-axis) as a function of stimulus input (plotted on the x-axis), are needed. Ho wever, there is currently only one clear speech study with psychometric func tions calculated for a limited number of normal hearing and cochlear impl ant listeners in noise (Liu et al., 2004), and it is unknown to what extent these functions w ould vary with different types of speech materials. Therefore, in order to assess the validity of comparisons and measurements of the clear speech benefit across populations and environments that have been made to date, and to conduct additional such comparisons, further research is needed to fully charac terize the psychometric functions of clear speech for normal hearing listeners.
5 Chapter 2 Literature Review Background Picheny et al.Â’s (1985) landmark study was conducted to assess the intelligibility benefit of cl ear speech for hearing impaired listeners. In this study, three male talkers who had some experi ence with public speaking or knowledge of clear speech recorded nonsense sentences using both clear and conversational speaking styles. These s entences were syntactically comparable to simple English sentences, but semant ically anomalous. Five listeners with sensorineural hearing loss heard the sentences via headphones, with two frequency-gain characteristics and three pr esentation levels. The proportion of key words each listener identified in the clear and conversational conditions were compared. Results showed an average intelligibility benefit of 17 percentage points for the sentences spoken clearly ac ross the different presentation modes. The compelling results from Picheny et al .Â’s study prompted further research into the attributes of clear speech. Inspec tion of acoustic features revealed many acoustic differences between clear and c onversational speech (Picheny et al., 1986).
6 Acoustic Properties of Clear Speech Several acoustic characteristics that distinguish clear from conversational speech have been identified. Temporal envelope modulations, the slowly varying amplitude changes that naturally occu r in speech, are measured using a temporal modulation index. Clear s peech has been found to have a greater temporal modulation index than conversational speech (Krause & Braida, 2004; Liu et al., 2004) fo r frequencies less than 3-4 Hz. P honetic features identified with clear speech include higher average fundamental frequency, a wider frequency range and expanded vowel space (Bradlow et al., 2003; Ferguson & KewleyPort, 2002; Krause & Braida, 2004; Pic heny et al., 1986). In addition, clear speech has some distinguishing phonologi cal characteristics. These include fewer vowel modifications and stop burst e liminations, a decrease in the use of alveolar flaps and an increase in schwa in sertions (Bradlow, et al., 2003; Picheny et al., 1986). One of the mo re striking characteristics associated with clear speech, however, is its reduced rate. Pi cheny et al. (1986) noted that clear speech, with nonsense sentences, has a markedly slower rate of 91 to 101 words per minute (wpm), a striking contra st to conversational sentences which were 160 to 200 wpm. This noticeable difference led researchers to question the influence of rate on intelligibility. Role of speaking rate. In order to determine if ra te could be manipulated to improve intelligibilit y, two groups of researchers arti ficially altered the rates of clear and conversational speech. Whether the time-scale of sentences were adjusted uniformly (Picheny, Durlach, & Br aida, 1989) or nonuniformly (Uchanski
7 et al., 1996) by altering the duration of i ndividual phonemes, over all intelligibility decreased for both clear sentences that were sped up to match the conversational rate and conversational sentences slowed down to match the clear rate. While the nonuniform time-scaling did not degrade the intelligibility of speech as much as uniform time-scali ng, it nonetheless produced sped up clear speech that was less intelligible than conversational speech. Moreover, when sentences were reprocessed to their original rates, intelligibility levels returned to their previous levels, indicating that t he decrease in intelligibility for processed sentences was not due to signal processing artifacts. In a related study, however, Uchanski et al. reported that clear sentences with deleted pauses were more intelligible than both unprocessed conver sational sentences and conversational sentences with added pauses. The conver sational sentences with added pauses were the least intelligible, and the altered clear sentences were not as intelligible as the original sentences that were pr oduced clearly. So, while pause deletions decreased intelligibility for the clear sentences, the alterations did not completely remove the intelligibility gain, suggesting that the benefits of clear speech are not entirely attributable to a slower speaking rate. Unsuccessful attempts to achieve clear speech at normal rates through artificial manipulations of conversational speech led res earchers to investigate if clear speech could naturally be produced at a normal speaking rate. This was first attempted by Uchanski et al. ( 1996) with a professional speaker who was instructed to produce sentences as clearly as possible at rates up to 400 wpm. These sentences were presented to norma l hearing listeners in quiet and noise,
8 and to hearing impaired listeners. Alt hough the talker was able to produce speech at different rates, no intellig ibility benefits were observed. In a subsequent study, five talkers with ba ckgrounds in public speaking received training for producing clear speech at normal (clear/normal) and quick (clear/quick) rates (Krause & Braida, 2002). The talkers also produced clear speech at slow (clear/slow) rates, and conversational sentences at slow (conversational/slow), normal (c onversational/normal) and quick (conversational/quick) rates. The sentenc es were presented to normal hearing listeners in noise. Results showed that lis teners received compar able intelligibility benefits with clear/slow and clear/normal speech, but benefits were not found with clear/quick speech. Krause and Braida Â’s research demonstrated that a slower speaking rate is not necessary fo r listeners to receive the intelligibility benefits associated with clear/slow speech. Thus, acoustic factors other than rate must be responsible for its increased intelligibility. The Clear Speech Benefit Although the specific acoustic charac teristics responsible for increased intelligibility with clear speech have yet to be identified, its inte lligibility benefits are well established and continue to be extended to additional listening situations. Researchers have measured inte lligibility benefits with clear speech in a variety of listening environments, with diffe rent stimuli, talkers and for different listener populations. Though the amount of benefit received may vary from one study to another, all have shown some im provement in intelligibility with clear versus conversational speech.
9 Presentation env ironments. Of the many listening environments that have been used to assess the benefits of clear speech, several have focused on presentation in noise. By presenting stimuli with noise to normal hearing listeners, researchers attempt to simulate hearing loss in their listeners. As a result, presentation of clear speech in noise is the most frequently used test environment for normal hearing listeners. These environments include speechshaped (Krause & Braida, 2002; Liu et al ., 2004; Payton et al., 1994), broadband (Gagne, Rochette, & Charest, 2002) and Gau ssian (Uchanski et al., 1996) noise. Speech-shaped noise, or filter ed white noise, is created to match the average long-term spectra of the signal (e.g. Payt on et al., 1994). Signal-to-noise ratios have varied, from -14 dB (Gagne et al., 2002) to +20 dB (Liu et al., 2004), but intelligibility benefits for clear speec h have been consistently obtained. The benefits of clear speech in noise ar e generally sizeable, ranging from 4 rationalized arcsine transformation units (RAU) (Bradlow & Bent, 2002) to 25 percentage points (Liu et al., 2004). Other presentation environments fo r which a clear speech benefit has been reported include low and high pass f iltering. Krause and Braida (2003) designed a study to assess the intelligibilit y of two talkersÂ’ clear speech in a low pass environment (achieved by presenti ng the stimuli through 1/3 octave band filters with frequencies from 80 to 1000 Hz) and a high pass environment (achieved by presenting the stimuli thr ough a 1/3 octave band f ilters centered at 3150 Hz). While a benefit in both environm ents was reported for both talkersÂ’ clear/slow speech, each talker achiev ed a benefit in onl y one environment with
10 clear/normal speech. In the low pass environment, the female talkerÂ’s clear/normal speech was more intelli gible (15 percentage points) than her conversational/normal speech, while the male talkerÂ’s clear/normal speech had a greater benefit (19 percentage points) in the high pass environment. Thus, large intelligibility benefits from clear speech are possible in filt ered environments, although the benefit may be talker-dependent fo r clear speech at normal rates. Another environment in which the inte lligibility benefit of clear speech has been established is reverberation. Payt on et al. (1994) tested normal hearing listeners in quiet and varying levels of noise in three different reverberant environments. An anechoic (ANEC) environm ent with no reverberation time (RT), a Â“living roomÂ” (LIVR), with 0.18 sec ond RT and a Â“conference roomÂ” (CONF), with 0.60 second RT, were used. Results showed that as the noise, reverberation, and noise plus reverberat ion levels increased, clear speech intelligibility benefits increased. The benefit between clear and conversational speech ranged from 15 percentage points in the LIVR environment with no noise, to 34 percentage points in the CONF envir onment with a 0 dB SNR. Krause and Braida (2003) found similar results for clear/slow speech in the CONF environment as well as a 19 percentage point benefit for clear/normal speech presented in reverberation with their male talker. These results suggest a strong clear speech benefit in re verberant environments. Presentation environments using audio-visual and visual-only modalities have also shown large benefits with clear speech. In a study with normal hearing listeners in noise, for example, Gagne et al. (2002) found a significant benefit for
11 clear versus conversational speech when syllables were presented with a carrier phrase in audio-visual, visual-only and audi tory-only modalities. The clear speech benefit averaged 7, 10 and 13 percentage points, respectively. Additionally, Helfer compared perception of nonsense sentences presented in auditory-only and auditory-visual conditions for both young (1997) and older (1998) listeners. Older listeners received an average clear speech benefit of 15 percentage points in the auditory-only condition and 11 percentage points in the auditory-visual condition. Younger listeners received si milar benefits for clear speech, with averages of 14 and 18 points for auditory-only and auditory-visual modes. The total benefit between auditory-only conv ersational and auditory-visual clear speech for younger listeners was 32 perc entage points. This difference equaled the sum of the benefit for clear speec h in auditory-only and auditory-visual modes, suggesting that the visual present ation of clear speec h provided additive intelligibility benefits. Stimuli. Not only does clear speech provide an intelligibility improvement in a variety of presentation environment s, but its intelligibility benefits are not limited to specific speech materials. A variety of stimuli have been used in clear speech experiments. In the landmark study by Picheny et al. (1985), as well as studies by several other investigators, grammatically correct but semantically anomalous sentences consisting of thr ee to four key words were used with intelligibility benefits rangi ng from 15 to 34 percentage points (Helfer, 1997, 1998; Krause & Braida, 2002, 2003; Payton et al., 1994; Picheny et al., 1989; Smiljanic & Bradlow, 2005; Uchanski et al., 1996). Nonsense sentences were chosen to
12 allow the talkers to vary their prosody as in natural speech, but without the benefit of semantic context for the list ener. Meaningful sentences have also been used, from the Bamford-Ko wel-Bench sentences (Bradl ow & Bent, 2002; Bradlow et al., 2003, Liu et al., 2004) to the J ohns Hopkins Lipreading Corpus (Schum, 1996). These meaningful stimuli have shown a clear speech benefit ranging from 4 (Bradlow & Bent, 2002) to 22 RAU (S chum, 1996). And, Gagne et al. (2002) found clear speech intelligibility gains of 7 to 13 RAU when CV and VCV syllables were presented with a carrier phras e. In sum, regardle ss of the diversity of stimuli presented in multiple studies all listeners received an intelligibility benefit with clear speech. Talkers. Intelligibility benefit s with clear speech have also been obtained for a variety of talkers. Although some st udies used recordings from talkers with some knowledge of clear speech (Payton et al., 1994; Picheny et al., 1985; Uchanski et al., 1996) and specifically selected talkers with a background in public speaking for eliciting clear speech at normal and quick rates (Krause & Braida, 2002), special training or a special population of talkers is not required for talkers to achieve a form of clear speech t hat is beneficial to intelligibility (Helfer, 1997, 1998; Picheny et al., 1985; Schum, 1997). In a study conducted to assess the intelligibility of older and younger talk ersÂ’ clear and conversational speech, Schum (1996) discovered no statistically significant difference between the groupsÂ’ ability to produce clear speech. In general, it has also been reported t hat the benefit of clear speech is largely independent of talker, at least when the listener populat ion and listening
13 environment are held constant (Krause & Braida, 2002; Picheny et al., 1985). While the talker main effect was signific ant in Picheny et al.Â’s study, analysis of variance (ANOVA) revealed talker and talker x speaking mode accounted for only 7% and 2% of the variance, respecti vely. Similarly, Krause and Braida found the talker x speaking mode interacti on was responsible for only a small percentage of the inte lligibility variance, although the ta lker factor was statistically significant. Despite differences in talkersÂ’ overall intelligibility, a comparable clear speech benefit can be received. In some cases, however, variability bet ween different talkersÂ’ clear speech intelligibility must be c onsidered. Even with closely matched acoustic features, different intelligibility outcomes have been observed between talkers (Goy, Pichora-Fuller, van Lieshout, Singh, & Schneider, 2007). Goy et al. presented identical high and low context sentences, recorded by two male talkers, to younger and older listeners in noise. Signifi cant differences in intelligibility between the two talkers were noted for both Â“clearÂ” and Â“normalÂ” speech, despite the talkersÂ’ comparable average FO and ra te. Acoustic characteristics of the more intelligible talkerÂ’s target wo rds included longer duration, higher FO and slightly increased intensity. A further sour ce of variability is that talkers may use different strategies to produce clear s peech, particularly when constrained to normal speaking rates. For example, tw o talkers made opposing adjustments in voice onset time and stop releases for clear/normal speech, and while these individual strategies resulted in similar in telligibility benefits in noise (Krause & Braida, 2002), substantial differences in the amount of benefit were observed in
14 other environments (Krause & Braida, 2003). In a low-pass environment, a female talkerÂ’s clear/normal speech was more intelligible, while a maleÂ’s was more intelligible in high-pass and re verberant environments. These findings demonstrate how benefits with clear s peech can be dependent on the strategies employed by the talker. Populations. Perhaps the most notable advantage of clear speech is that it benefits a variety of listener populations. These include normal hearing listeners in quiet and noise (Gagne et al., 2002; Kr ause & Braida, 2002; Liu et al., 2004; Payton et al., 1994; Uchanski et al., 1996) hearing impaired listeners (Payton et al., 1994; Picheny et al., 1985; Uchanski et al., 1996), older listeners (Helfer, 1998; Panagiotopoulos, 2005), an d listeners with cochlear implants (Liu et al., 2004). In addition, both children with l earning disabilitie s (LD) and normally developing children (Bradlow et al., 2003) show higher intelligibility scores with clear versus conversational speech. Ho wever, because children with LD have lower overall intelligibility scores and their speech perception is affected by noise more than children without LD, with cl ear speech, children with LD obtain intelligibility scores that are comparabl e to the average non-LD childrenÂ’s scores for conversational speech. Non-native English speak ers are another population shown to benefit from clear speech (Bradlow & Bent, 2002; Krause & Braida, 2003). Although both nonnative and native speakers of English re ceive intelligibility benefits with clear speech, only one study found the benefit to be smalle r for non-native listeners (Bradlow and Bent, 2002). This difference was attributed to the amount of time
15 that non-native listeners had been exposed to English, and to the possibility that native listeners may have received a c ontextual advantage from the meaningful sentence stimuli. Without the availabi lity of context, Krause and Braida found no difference in the amount of benefit between the two groups for nonsense sentences in noise. They attributed the discrepancy between the two studies to the differences in stimuli and length of time the listeners had been exposed to English. It is also worth noting that the clea r speech benefit is not restricted to English. One study compared perception of clear speech nonsense sentences in noise, produced in Croatian and English for their respective native listeners (Smiljanic & Bradlow, 2005). The average in telligibility benefit with clear speech was 16 RAU for English and 15 RAU for t he Croatian listeners. There was no significant effect of language, and the authors reported that talkers in both languages demonstrated a reduced rate of speech, expanded vowel space and an increase in pitch range. These findings suggest that applications of clear speech in noise or with a hearing impai red listener would be beneficial, regardless of specific language. Summary Clear speech provides benefits to listeners in many different environments. Even with a variety of stimuli, clear speech intelligibility is greater than intelligibility for conver sational speech. Talkers may vary the strategies they employ when producing clear speech, but the benefit remains. However, the extent to which the benefit varies across di verse listening situations has not been
16 well characterized. More research is needed to determine how the clear speech benefit for each listener population changes with environment, stimuli, and talker. The Need for Psychometric Functions Given that the clear speech benefit applie s to so many listening situations, questions have arisen regarding which populations and env ironments benefit most. To answer these questions, research ers have attempted to compare their findings, with clear and conversational s peech at slow and normal rates, with results in other investigatorsÂ’ studies for different populations or environments. These comparisons are complicated by inconsistent methods in the form of participants, stimuli and presentation env ironments across st udies. Therefore, comparisons are at times poorly justifi ed due to insufficient matching of listener performance at baseline (i.e., conv ersational speech intelligibility) and psychometric functions with clear speech, at normal and slow rates, are needed. Psychometric functions for speech perception are represented by a graphical plot of performance (e.g., percent key words correct), on the y-axis, in relation to stimulus level or SNR, on the x-axis. Psychometric functions are available for conversational speech in noise, and have been used to help predict performance for normal and hearing impaire d listeners (Wilson & Strouse, 1999). These same functions are needed for lis teners with clear speech to determine relative intelligibility benefits of clear s peech across various levels of degradation. Many factors contribute to the intellig ibility of speech and as a result, the psychometric functions for clear speech could be substantially different than those for conversational speech. Indi vidual speech sounds vary in length,
17 intensity and frequency, with the changing patterns of the sounds contributing to perception and recognition by the listener (Frenc h & Steinberg, 1947). One predictor of speech intelligibility is t he articulation index (AI), which determines intelligibility of speech sounds with a gi ven frequency, intensity and noise level (French & Steinberg, 1947). However, the AI has not been shown to be an accurate predictor of intelligibility for clear speech (Payton et al., 1994), suggesting that psychometric functions for clear speech are likely to be different from those of conversational speech. Even if the psychometric functions fo r clear and conversational speech are similar, the psychometric functions that have been obtained previously for conversational speech with word list s (Wilson & Strouse, 1999) and CV/VC syllables (Miller & Nicely, 1955) may not be appropriate for nonsense sentence materials frequently used in clear speech ex periments. One factor affecting slope of a psychometric function is the type of test materials (Wilson & Strouse, 1999). Average listener response for easier or more homogeneous stimuli, such as spondaic words, will have a steeper slope t han sentences, which present a more difficult perception task. In addition, large intersubject variability results in a flatter slope, small variability between subjects provides a steeper slope and a greater number of test subjects help to lessen the influence of intersubject variability. Previous studies may not have controlled for intersubject variability by using a small number of listeners and/ or stimuli. Therefore, complete psychometric functions for both conversational and clear speech, at normal and slow rates, are needed. The increased knowledge that could be gained from listener
18 psychometric functions with clear s peech would provide a foundation for researchers to make more legitimate comparisons between studies, taking baseline performance into account. Comparisons across studies Researchers currently tend to compar e the relative benefits of clear speech with results reported in other studi es, without regard to absolute levels of performance. The validity of such com parisons relies on the assumption that underlying psychometric functions have sim ilar slopes. Most comparisons that have been reported are for normal heari ng and hearing impaired listeners. Multiple studies with hearing impaired listeners have shown comparable clear speech intelligibility benefits, from 15 to 17 percentage points (Payton et al., 1994; Picheny et al., 1985; Uc hanski et al., 1996). This level of consistency is somewhat unexpected, since listener s with hearing impairments have greater variability with word recognition tasks (Wilson & Strouse, 1999) and baseline performance was not controlled, but t he studies were conducted with similar listeners, and used the same stimuli and presentation environments (Payton et al., 1994; Picheny et al., 1985; Uchanski et al., 1996). Comparing studies between normal hearing and hearing impaired listeners has also been undertaken, but results are less straightforw ard. In an effort to simulate hearing loss or difficult listening situations, normal hearing listeners have been presented different types of stimuli in varying environments. The number of listeners per condition has ranged from 5 (Uchanski et al., 1996) to 12 (Gagne et al., 2002). To accurately compare the results for normal hearing listeners to individuals with
19 hearing loss, basic psychometric functions for a large number of subjects with multiple signal-to-noise ratios and similar stimuli are needed. Research comparisons with clear speech have also been made between older and younger listeners. Panagiotopo ulos (2005) presented nonsense sentences at 0 dB SNR to older listener s to measure intelligibility of clear and conversational speech at different rate s. Results were compared to data from Krause and BraidaÂ’s (2002) study of young normal hearing listeners, which used the same sentence lists with presentati on at -2 dB SNR. Intelligibility for conversational/normal and clear/normal speech was the same in both studies, and clear/slow speech intelligibility was comparable. An unexpected finding in PanagiotopoulosÂ’ research was a 21 per centage point benefit for the older listeners with conversational/slow speech. This was larger than the clear/normal speech benefit and roughly equal to the benefit observed with clear/slow speech. The conversational/slow intelligibility benef it may be due to acoustic properties of this condition that are more beneficial to older listeners than younger listeners. Another explanation may be that there is a ceiling effect for older listeners with slow speech at 0 dB SNR, which could also occur with younger listeners. Due to the differences in signal-to-noise ratio between PanagiotopoulosÂ’ and Krause and BraidaÂ’s studies, direct comparisons of the results are difficult. Other research comparisons hav e been conducted between listeners within a study. For example, Bradlow and Bent (2002) compared native and nonnative listener performance. They repor ted a decrease in the clear speech benefit for non-native listeners and proposed that this was due to the listenerÂ’s
20 length of exposure to the language. In contrast, Krause and Braida (2003) found the same clear speech benefit for bot h native and non-native listener groups. Neither study attempted to control basel ine performance in order to ensure that the across-group comparisons were valid Moreover, differences in signal-tonoise ratio and stimuli (meaningful vs. nonsense sentences) complicate comparisons between the two studies. To truly compare intelligibility benefits for native and non-native listeners, similar stimuli and performance levels should be used. Statement of the Problem Comparisons of clear speech data obt ained in different studies must be interpreted with caution until basic psychometric functions are known. Researchers need to be able to compare re sults across studies, where superior or inferior performance between subjects has been controlled. The influence of intersubject variability can be reduced with a larger pool of participants and greater number of stimuli (Wilson & St rouse, 1999). Characterization of basic functions for young normal hearing lis teners will provide a foundation for comparing intelligibility with older list eners, hearing impaired listeners and other populations. These basic functions are needed to provide information about listener performance with clear speech at different rate s and noise levels. Finally, listener functions with nonsense sentences are necessary to control for any influence of semantic context (Picheny et al., 1985). There is some research providing p sychometric functions for clear speech. Liu et al., (2004) reported intelligibility of clear and conversational speech as a
21 function of SNR for normal hearing listener s and for cochlear implant users. In one of four experiments, five listeners heard clear and conversational speech recorded by a female talker, and six list eners heard a maleÂ’s speech. The stimuli consisted of meaningful sentences in signal-t o-noise ratios of -20 to +20 dB in 5 dB presentation steps. Each liste ner heard only eight clear and eight conversational sentences at each SNR. When psychometric functions were graphically plotted, the slope for clear s peech stimuli was steeper than the slope for conversational speech stimuli, with an average intelligibility benefit of 29 percentage points. Although these psychom etric functions contribute important information regarding the relative intelligi bility benefits of clear speech at various SNRs, to what extent these functions are applicable to other types of speech materials is not known. The 29 point benef it with clear speech, larger than the benefit found in previous studies with nonsense sentences, may be due to contextual cues received by the listener s or to the range of SNRs tested. In addition, intersubject variability may have skewed the results due to the small number of participants and limit ed stimuli presented. Since nonsense sentences are frequently used in clear speech research, the purpose of this study is to determine how speech intelligibility, as measured by percentage key words correct in nonsense sentences by young adults with normal hearing, varies with changes in s peaking condition, talker and varying signal-to-noise ratios.
22 Chapter 3 Methods Participants A total of 40 participants (31 females, 9 males; ages 18 to 38 years) were selected for the study. Demographic info rmation for participants is listed in Appendix A. For inclusion in the study, participants were required to be between 18 and 40 years of age and to pass an audiological hearing screening. Eligible participants had pure tone thresholds of 20 dB HL or below at 250, 500, 1000, 2000 and 4000 Hz in at least one ear. Part icipants who were not native English speakers or who did not have a high sc hool diploma or it s equivalent were excluded from the study. Materials The sentence materials used for t he listening sessions consisted of grammatically correct, but semantica lly anomalous, sentences created by Picheny et al. (1985) and previously recorded by Krause and Braida (2002). These nonsense sentences allowed the lis teners to receive prosodic information without any contextual cues (Picheny et al., 1985). An example of a nonsense sentence is Â“His right cane could guard an edge.Â” The sentence lists were previously recorded by four talkers from Krause and BraidaÂ’s (2002) study. All of the talk ers had experience with public speaking.
23 Of the five talkers in that study, one ma le (T5) and three female (T1, T3, T4) talkers were selected based on their abili ty to manipulate both rate (slow vs. normal) and intelligibility (clear vs. conversational). A total of nine distinct 50-sentence lis ts were selected from the Krause and Braida (2002) database. O ne list was used for practi ce. The remaining eight lists were presented twice, once in each of the two speaking styles. Specifically, for each of the four ta lkers, one list was presented in clear/slow and conversational/slow styles, and one list in clear/normal and conversational/normal styles. Thus, 800 utterances (8 lists x 50 sentences x 2 recordings) were divided evenly between four different speaking styles, and 200 utterances (50 utterances per talker) were presented in each of the following modes: conversational at a normal rate (conversational/normal), conversational at a slow rate (conversational/slow), clear at a normal rate (clear/normal) and clear at a slow rate (clear/slow). For pres entation, the eight lists were sorted into blocks of four, as shown in Table 1.
24 Table 1. Sentence Blocks Block Talker/Sentence Mode List A T1 Â– Conversational/normal T1 Â– Clear/slow T4 Â– Clear/normal T4 Â– Conversational/slow List 1 List 2 List 3 List 4 B T3 Â– Clear/normal T3 Â– Conversational/slow T5 Â– Conversational/normal T5 Â– Clear/slow List 5 List 6 List 7 List 8 C T1 Â– Clear/normal T1 Â– Conversational/slow T4 Â– Conversational/normal T4 Â– Clear/slow List 1 List 2 List 3 List 4 D T3 Â– Conversational/normal T3 Â– Clear/slow T5 Â– Clear/normal T5 Â– Conversational slow List 5 List 6 List 7 List 8 Presentation Sessions During the presentation sessions, each participant was assigned a workstation with a keyboard, monitor and headphones in the group lab of the Communication Sciences and Disorders depar tment of the University of South Florida. Because this lab is designed to accommodate four participants at one time, there were a total of ten groups of participants, with four participants per group. Due to scheduling conflicts, however some participants were alone in the lab during their sessions. The tester moni tored all sessions from an adjacent lab that contained a window providing a view of the participant(s), as well as audio feedback via microphones plac ed in the group lab.
25 Each group or individual participant was scheduled to attend four weekly two-hour presentation sessions. In each session, participants heard one block of four 50-sentence lists, presented monaurally over Sennheiser HD265 headphones. After listening to the first and third lists, participants received a five minute break, and a ten minute break was gi ven after the second list. The default presentation level was set to 85 dB SPL, and each participant had the opportunity to select a preferred listeni ng level (which was then fixed for the duration of the experiment) within 6 dB of the default level, adjusted in 3 dB increments. Sentences were presented in the pr esence of speech-shaped noise taken from the Hearing In Noise Test (Nilsson, Soli, & Sullivan, 1994). As shown in Table 2, a total of five signal-to-noise ra tios were used (-4, -2, 0, +2, +4 dB), but due to limitations in the number of sentences, each listener group heard sentences presented at only four of the noise levels The signal-to-noise ratio was varied across listening sessions, with one noise level for each of the four sessions (i.e. sentence blocks). The order of presentation for signal-to-noise ratio, also shown in Table 2, was from the most difficult in session one to the easiest in session four. This order wa s chosen to minimize fatigue and to maintain the participantÂ’s interest in the study. Finally, Table 2 shows the presentation schedule for the sentence lists. In order to minimize learning effects wit hin a list during the experiment, listeners received a one-week break in between se ssions two and three. This break
26 provided a three week time period between repeat lists by the same talker, at slow or normal rates (i.e. Blocks A and C; Blocks B and D). Table 2. Sentence Presentation Schedule Session 1 Session 2 Session 3 Session 4 Group SNR Block SNR Block SNR Block SNR Block 1 -4 A -2 B 0 C 2 D 2 -2 A 0 B 2 C 4 D 3 -4 D 0 A 2 B 4 C 4 -4 C -2 D 2 A 4 B 5 -4 B -2 C 0 D 4 A 6 -4 C -2 D 0 A 2 B 7 -2 C 0 D 2 A 4 B 8 -4 B 0 C 2 D 4 A 9 -4 A -2 B 2 C 4 D 10 -4 D -2 A 0 B 4 C For each group session, the tester sat at a personal computer and presented one sentence at a time vi a a Matlab software program, and the participants typed their responses in an Excel spreadsheet. Each participant signaled when they were ready for the next sentence by turning a page on a flip chart at his/her workstation that was vi sible to the tester. The tester did not present the next sentence until all participants signaled they were ready to continue. Participants were also given t he option of verbally requesting a pause, asking a question or raising their hand to signal a request for the tester to pause at any time during the session. For single participant sessions, participants controlled the rate of presentation of s entences from their workstation. In all
27 sessions, each participantÂ’s responses were saved as a text file after each sentence list was completed, and a new spreadsheet was generated for the next sentence list. At the first session, participants rece ived instructions and listened to 10 to 15 practice sentences to become familiar with the task. They were informed that the sentences would not make any sens e and that the background noise might make it difficult to understand the words. They were encouraged to make their best attempt at typing a complete sent ence, guessing if necessary. They were informed that their responses provided valuable information, whether their answers were correct or not, as there was no Â“rightÂ” or Â“wrongÂ” response. For subsequent sessions, participants received br ief instructions reminding them to do their best to type what they heard. Participants typed their responses to t he practice sentences in the Excel spreadsheet and turned the pages on their f lip charts to become accustomed to the session procedures. The participants did not receive any feedback from the tester regarding the correct ness of their responses, but were reminded to make their Â“best guess.Â” The first fi ve practice sentences were presented in quiet. This format was intended to allow participa nts to hear the nonsense sentences without any competing noise to adjust to the unfamiliar sentence structure. The next group of sentences was pr esented with speech-shaped noise, and participants then provided feedback about th e presentation level. The tester adjusted the gain plus or minus 3 dB or 6 dB, per participant request, and recorded each adjustment for future sessi ons. After the parti cipants completed
28 the practice sentences and the tester addr essed any questions that were posed, the first sentence list was presented. Scoring Procedures Each sentence was scored for three to four key words, consisting of all nouns, verbs and adjectives in the sentence. Responses were scored up to two times, once with an autoscoring program in Matlab, which graded each sentence as Â“correctÂ” if the participantÂ’s response contained all of the key words with exact spelling. A second text file containing any sentence with an error was then generated from Matlab. This file was hand graded by one of three graders, and credit was given for homophones, nouns wit h a plural added or deleted, changes in verb (ed) past tense, or spelling erro rs. For example, if the target sentence was Â“Her blond shore grins at her mannerÂ” and the response was Â“The wand shores grinned at her manorÂ”, Matlab placed this sentence in an error file and scored the entire sentence as incorrect. The key words in this sentence are Â“blond,Â” Â“shore,Â” Â“grinsÂ” and Â“manner.Â” The tester credited the respondentÂ’s sentence with three correct key word s (Â“shores,Â” Â“grinnedÂ” and Â“manorÂ”). The number of key words correct were tota led and then divided by the total possible key words in the sentence list to obtain the percentage of key words correct for each list. Data Analysis The percent-correct scores from the listening sessions were then compiled to construct 16 psychometric functions for clear and conversational speech at five signal-to-noise ratios. Data from ei ght listeners comprised each data point on
29 each psychometric function. Due to the limited number of sentence lists, the same eight listeners did not make up t he data points in a particular function. However, based on previous studies, responses from normal hearing listeners have minimal variability and should be re latively interchangeable (Wilson & Strouse, 1999). There were four functions per talker: clear/slow, clear/normal, conversational/slow, conversational/no rmal. Functions were compared between the clear and conversational conditions of the same rate to see how the size of the clear speech benefit varied with talk er and/or signal-to-noise ratio.
30 Chapter 4 Results The purpose of this study was to examine how speech intelligibility, measured as percent key words correct, varied with condition, talker and signalto-noise ratio for young normal hearing lis teners. In addition to assessing the relative intelligibility of each conditi on on average, the data collected were examined for talker differences by constructing 16 graphs (4 talkers x 4 conditions) depicting psychometric functions, with intelligibility as a function of signal-to-noise ratio. A three-way analysis of variance (ANOVA) was also performed on key word scores after an arcsine transformation ( Ij/100) was applied to equalize the variances. Betw een-subject factors were condition (clear/slow, clear/normal, conversati onal/slow and conversational/normal), SNR (five levels) and talker (four levels). As shown in Table 3, all main effects and most interactions were statistically si gnificant (p < 0.01); each of these will be discussed in detail in the following sections.
31 Table 3. Between-subject Effects and Variables Source Type III Sum of Squaresdf Mean SquaresF Sig. Partial Eta Squared talker 4.1673 1.389 263.785 0.000* 0.586 SNR 18.9604 4.740 900.161 0.000* 0.865 condition 3.5003 1.167 221.554 0.000* 0.543 talker x SNR 0.22012 0.018 3.486 0.000* 0.070 talker x condition 2.4709 0.274 52.116 0.000* 0.456 SNR x condition 0.11912 0.010 1.877 0.034 0.039 talker x SNR x condition 0.41936 0.012 2.208 0.000* 0.124 p < 0.01 Tables B1-B4 in Appendix B list key word scores obtained by all listeners in all conditions for each of the four talkers. These result s are summarized in Figure 1, which shows a large effect (partial 2 = .543, p = 0.000) of condition. As expected, conversational/normal was the least intelligible condition (53%), since this speaking style represents Â“typicalÂ” speech (normal rate with no particularly emphasis on clarity). Therefore, this condition was considered the baseline for measuring intelligibility benef its for the other three c onditions. Post-hoc t-tests were conducted to evaluate pairwise comp arisons of all conditions (see Table C1, Appendix C). Slow conditions were the most intelligible, with benefits of 15 (p = 0.000) and 17 (p = 0.000) percentage points for conversational/slow and clear/slow, respectively. Clear/normal speech also provided an intelligibility benefit of 11 percentage points (p = 0.000).
32 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% conv/normalclear/normalconv/slowclear/slowPercent Correct (%) Figure 1. Average intelligib ility, in percent key words correct, for each test condition. For clear/slow and clear/normal speech, in telligibility benefit s of this size are consistent with benefits r eported in previous studies of young, normal hearing listeners tested under similar conditions (nonsense sentences presented in noise). For example, the average intelligi bility benefit of clear/slow speech in the current study, 17 percentage points, is comparable to benefits reported for three previous studies that utilized nonsense sentences in noise. Helfer (1997) and Krause and Braida (2002) repor ted listener intelligibilit y benefits of 14 and 18 percentage points, respectively, and Smiljanic and BradlowÂ’s (2005) listeners received benefits of 16 RAUs. Similarly, with clear/normal speech, the current group of listeners received an 11 point benefi t, which is consistent with the 14
33 point benefit reported by Krause and Braida. For conversational/slow speech, however, the average benefit of 15 percentage points was much larger than the 6 point benefit reported for Krause and Brai daÂ’s listeners. This difference is particularly surprising given that the st imuli used in the current study were obtained from the same database used in Krause and BraidaÂ’s research. One explanation that mi ght account for this discrepancy is that the average intelligibility benefit across all five SNRs employed in this study is not the most valid metric for comparing to prev ious studies that employed only one SNR. Therefore, comparisons we re further narrowed to benef its obtained at the same SNR. Using this method, consistency of results with previous research remains strong for clear/slow speech: at -2 dB SNR, the present listeners received intelligibility benefits of 20 points, versus Krause and Brai daÂ’s (2002) 18 points; at 0 dB SNR, the benefit was 19 points, compared to Smiljanic and BradlowÂ’s (2005) listener benefit of 16 RAUs, and at +2 dB SNR, listeners in the current study received a 14 point benefit, the same benefit reported fo r HelferÂ’s (1997) listeners. Another strong comparison is t he clear/normal speech benefit, at -2 dB SNR, which was 14 percentage points for listeners in the both the present and Krause and BraidaÂ’s study. Yet, the inconsis tency for conversational/slow speech benefits remains between the current study (19 points) and Kr ause and BraidaÂ’s research (6 points), even when benefits are compared only at the SNR used in both studies, -2 dB. In fact, at 0 dB, t he 19 point intelligibility benefit listeners received in this study is closer to benefits received by older normal hearing listeners (23 points) at t he same presentation level (Panagiotopoulos, 2005).
34 Psychometric Functions Figure 2 shows the psychometric f unctions for the four conditions, averaged across talkers. As expected, a large effect of SNR was observed (partial 2 = .865, p = 0.000), wit h listener performance improving with SNR. However, the psychometric functions were constructed to assess intelligibility benefits of each condition as a function of signal-to-noise ratio, or the SNR x condition interaction. Again, intelligibility for conver sational/normal speech was used as the baseline for measuring benefits fo r the other three conditions at each SNR. The broad pattern of the overall resu lts, seen in Figure 1, was not affected by changing signal-to-noise ratio (Figur e 2). Measurements between conditions, as well as visual inspection, indicate a roughly constant benefit from -4 to 0 dB SNR. However, the interaction of SNR with condition approached significance (p = .034) because the size of intelligibilit y benefits for all conditions decreased for the higher SNRs (+2 and +4 dB). This change in benefit can be observed for the clear/slow condition. At +4 dB SNR, t here is an 11 percentage point intelligibility benefit with clear/slow speech that incr eases to 21 points at -4 dB SNR. Similarly, with clear/normal speech, the largest benefit (14 points) was observed at -4 dB and the smallest benefit (8 points) at +4 dB The reason for the decrease in benefits at +2 and +4 dB SNR may be due to a ceiling effect, where intelligibility is asymptotically approachi ng maximum levels. For the SNR range of -4 to 0 dB, on the other hand, the intelligibility of eac h condition increases linearly with SNR, suggesting that th is region represents the Â“mi ddleÂ” of the psychometric function.
35 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% -4-20+2+4SNRPercent Correct (%) Conv/slow Clear/slow Conv/normal Clear/normal Figure 2. Average psychomet ric functions, in percent key words correct, across signal-to-noise ratio. Defining the Â“middleÂ” or linear port ion of the psychometric function and determining whether the relative intelligi bility of each condition varies in this region is one of the central purposes of this study. The data in Figure 2 and individual talker data in Figure 4 suggest that the middle of the psychometric function for young, normal hearing list eners presented with nonsense sentences in noise corresponds to the region for which intelligibility for the conversational/normal condition ranges from 20% to 65%. Over this region, the psychometric functions for all conditions are approximately linear. Because the slopes for all conditions were relatively similar in this re gion, the relative
36 difference in intelligibility between the p sychometric functions for each condition remained fairly constant, with a change of 2 percentage points or less in the benefit size for all conditions. For in stance, the intelligibility benefit for conversational/slow speech was 19 points at 0 dB SNR, and 17 points at -4 dB SNR. Excluding the ceiling effect, t hese results demonstrate a small change in the average size of intellig ibility benefits across signal-tonoise ratio. This finding suggests that young, normal hearing lis teners can receive consistent and predictable benefits for a giv en range of noise levels. Effect of Talker Not surprisingly, some talkers were more intelligible overall than other talkers, and the ANOVA (Table 3) showed a large effect of talker (partial 2 = 0.586). Post-hoc t-tests (see Appendix C, Ta ble C2) confirmed that T3, a female, was the least intelligible talker at 53% (p = 0.000) on average and the most intelligible talker overall was T5, a male, at 73% (p = 0.000) on average across all conditions. For the purposes of this thesis, howev er, the goal was to compare how the intelligibility benefits of each condition vari ed across talkers. Therefore, the talker x condition interaction was of primary interest. This interaction was significant (p = 0.000) and showed a large effect size (partial 2 =0.456), nearly as large as that of talker alone. To assess this in teraction, Figure 3 shows the average key words scores for each talker in each condition. T1Â’s data generally followed the overall pattern obtained for all talkers, with clear/slow and conversational/slow speech as the most intelligible conditions followed by clear/normal speech, and
37 with conversational/normal speech as the least intelligible condition. Even so, the intelligibility benefit of T1 Â’s clear/normal speech was somewhat smaller (6 points) than the benefit of this condition on average across all talkers (11 points). Patterns for the other three talkers differ ed more substantially from the overall results. For example, neither T4 nor T5 produced conversational/slow speech that provided as large an intelligibility advantage as clear/slow speech. Moreover, T5Â’s clear/normal speech provided a subst antially larger benefit (13 points) than his conversational/slow speech (5 points) nearly as large a benefit as that provided by his clear/slow speech (17 points). 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% T1T3T4T5Percent Correct (%) conv/normal clear/normal conv/slow clear/slow Figure 3. Average inte lligibility, in percent key wo rds correct, for each talker in each condition.
38 On the whole, however, it was T3Â’s configuration that differed from the average more than any talker, with intell igibility benefits of 20 points in the clear/normal and 33 points in the conversa tional/slow conditi ons, compared to 11 points and 15 points for these conditions on average. Moreover, it is worth noting that T3 was the only talker who did not have the largest intelligibility benefit with the clear/slow condition. While any definit ive reasons for T3Â’s relatively poor clear/slow intelligibility are unknown at this time, there are some possible explanations. In Krause and BraidaÂ’s (2002) study, clear/slow was the first condition that talkers produced, with no tr aining, and the first condition recorded. It is possible that T3Â’s production of clear speech improved with subsequent tasks, resulting in greater intelligibilit y with her clear/normal condition, which involved intense training and listener feedback. Additionally, all of the talkers had public speaking experience, r anging from five to seven years, with the exception of T3, who had been a broadcasting student for two years and was the least experienced in the group. Despite these variations, some si milarities across talkers can be observed. Intelligibility for a ll talkers improved with speaking clearly, at a slow or normal rate, and/or with speaking slowly, in a conversational or clear mode. For three talkers, T1, T4 and T5, clear/slow was the most intelligible condition. And, for T3, T4 and T5, clear/normal was t he second most intelligible condition. Therefore, clear speech prov ided an intelligibility benefit for listeners at both slow and normal rates.
39 Across talkers, the benefit of co nversational/slow speech was the most variable of any condition (indicating that it was least correlated with a talkerÂ’s conversational/normal intellig ibility), with the size of t he benefit varying as much as 28 points between different talkers (T 3 33 points, T5 5 points). One explanation for this variability could be t hat this speaking styl e may represent an unnatural task for talkers. In the original study w here the conversational/slow stimuli were recorded, Krause and Braida ( 2002) allowed their talkers to produce this particular speaking condition without any specific instructions other than to use a Â“normalÂ” style for speaking slo wly without any particular emphasis on clarity. This method of e licitation may have resulted in an unreliable mode of communication because it may not be rea listic to expect talkers to have a Â“normalÂ” style at slow rates that does not involve increased cl arity. Thus, talkers may have found it difficult to slow their speech without adopting any of the acoustic characteristics of clear speech. Other talkers may have used atypical prosody that distorted the acoustic cues that listeners typically rely upon with speech. Note that with this degree of in ter-talker variability, experiments that base conclusions on outcomes for single talk ers are likely to obtain substantially different results, from each other and from studies using multiple talkers, regarding the benefits of conv ersational/slow speech. In the clear/normal condition, for which talkers received training (see Krause & Braida, 2002), ther e was less variation in intelligibility benefit across talkers. The size of the benefit varied 14 points between T3 (20 points), and T1 (6 points). Even though this variability is somewhat smaller than that observed for
40 the conversational/slow benefit, it is still sizeable and likely to be a factor in experiments investigating t he benefits of clear/normal sp eech that involve only a single talker. The least variable condition was clear/slow, with the size of the intelligibility benefit vary ing just 7 points between T3 (13 points) and T4 (20 points) and only 3 points across three of the four talkers (T1 19, T4 20, T5 -17 points). The reduction in inter-talker va riability for clear/slow speech is good news for researchers who employ only one or two talkers, as more reliable results can be expected. However, a 7 point difference in intelligibility is not trivial, and further work is needed to determine whether the variability introduced by T3 is typical of talkers, or whether she is an outlier and the 3 point variability is representative of most talkers. Psychometric functions by talker. Further analysis of performance for each talker was conducted as a psychometric function of SNR in order to evaluate the combined effects of talker and SNR on the rela tive intelligibility of each condition. A medium size effect (partial 2 = 0.124; p = 0.000) fo r the talker x SNR x condition interaction was observed (see Table 3). Figure 4 displays each talkerÂ’s results, all of which reflect variations from the average psychometric functions shown in Figure 2. As with the average results for talker x condition, T1 (Figure 4a) was the only talker who followed the over all pattern relatively closely. But, in the middle of the psychometric function ( described earlier as the linear region corresponding to conversational/normal intelli gibility of 20% to 65%), from -4 to 0 dB SNR, benefits for T1 in all conditions were smaller at -2 dB than at 0 dB, rather than remaining constant. This change in benefit size with SNR may be
41 explained by the conversational/normal psychometric function, which is not completely linear in this region. There appears to be a Â“dropÂ” in intelligibility for conversational/normal speech at 0 dB and then a local Â“peakÂ” at -2 dB, resulting in a larger benefit at 0 dB and a smaller ben efit at -2 dB for all three conditions. For example, with clear/normal speech, T1Â’s maximal benefit is 11 points at 0 dB SNR, which decreases to 3 points at -2 dB and then increases to 8 points at -4 dB. Similarly, T3Â’s clear/normal benefit (Figure 4b) varies as much as 8 points (from a 16 point benefit at +4 dB to a 24 point benefit at -2 dB ) in the middle of the psychometric function from -2 to +4 dB SNR (note that the linear region of T3Â’s psychometric function spans a diffe rent range of SNRs due to her lower overall intelligibility) In contrast, however, T3 shows a fairly constant benefit for the clear/slow (13 -17 point s) condition in this region but even less consistency (benefits ranging from 26 39 points) for conversational/slow speech. In addition, the relative benefits for all conditions decreas es at -4 dB SNR for T3, likely due to a floor effect, as her conv ersational/normal speech intelligibility is less than 20% at this presentation level. Although re sults for T4Â’s (Figure 4c) clear/slow condition are consistent with the av erage psychometric function, suggesting a relatively constant benefit from -4 to 0 dB SNR, this talkerÂ’s intelligibility varied considerably from the average pattern for both clear/normal and conversational/slow speech, which both showed larger benefits (12 Â– 13 points) at -2 dB, than at either 0 dB (6 points in both conditions) or -4 dB (3 and 5 points for clear/normal and conversational/slow speech, respectively).
42 Perhaps the most notable differenc es from the average psychometric functions were demonstrated by T5 (Figure 4d). In the middle of the psychometric function from -4 to 0 dB SNR, conversational/slow benefits decreased with worsening SNR while the inte lligibility benefits of clear/slow and clear/normal speech increased over the same range. Moreover, the change in benefit over this range was sizeable: t he clear/slow advantage increased from 19 to 33 points, and the clear/normal advanta ge increased from 16 to 23 points. So, while the benefits for each condition are preserved across SNR for talkers on average, this was not the case for individual talkers.
43 a) T1 b) T3 0% 20% 40% 60% 80% 100% -4-20+2+4SNRPercent correct (%) Conv/slow Clear/slow Conv/normal Clear/normal 0% 20% 40% 60% 80% 100% -4-20+2+4SNRPercent correct (%) Conv/slow Clear/slow Conv/normal Clear/normal c) T4 d) T5 0% 20% 40% 60% 80% 100% -4-20+2+4SNRPercent correct (%) Conv/slow Clear/slow Conv/normal Clear/normal 0% 20% 40% 60% 80% 100% -4-20+2+4SNRPercent correct (%) Conv/slow Clear/slow Conv/normal Clear/normal Figure 4. Average intelligibility, in perc ent key words correct, for each talker across signal-to-noise ratio. Inspection of the psychometric func tions for individual talkers also revealed substantial differences in the effe ct of SNR on the size of intelligibility benefits for each condition, r epresented in Table 4. To illustrate, with clear/slow
44 speech, T5 had the greatest change in the si ze of benefit across SNRs. At -4 dB, his clear/slow speech provided a benefit of 33 points relative to the conversational/normal baseline of 35%, but it provided a benefit of only 4 points at 4 dB, where baseline intelligibility was 86%. Thus, the benefit of clear/slow speech varied up to 29 percentage points for T5, although his intelligibility at the easier listening conditions likel y represents a ceiling effect.
45 Table 4. Individual talker inte lligibility benefits by condition and SNR Talker SNR (dB) Conv/normal (baseline) intelligibility Clear/normal benefit Conv/slow benefit Clear/slow benefit T1 -4 25 8 24 23 -2 46 3 19 21 0 53 11 23 22 +2 73 0 10 12 +4 76 4 11 13 Benefit range* 8 5 2 T3 -4 11 23 33 5 -2 21 24 39 14 0 36 17 35 15 +2 49 21 34 17 +4 63 16 26 13 Benefit range 8 13 4 T4 -4 32 3 5 25 -2 45 13 12 22 0 59 6 6 20 +2 69 9 6 18 +4 77 6 3 11 Benefit range 10 7 5 T5 -4 35 23 7 33 -2 55 16 5 23 0 64 16 10 19 +2 81 5 2 6 +4 86 3 2 4 Benefit range 7 5 14 Benefit range is calculated over the middle of the psychometric function, where conversational/normal intelligibility ranges from 20% to 65%. Data points in the middle of the psychometric function are indicated in bold. More importantly, how the size of the intelligibility benefit for each condition changed in the middle of the psyc hometric function (i.e avoiding ceiling effects) for individual talkers was also measured. As shown in Table 4 (data points corresponding to the middle of t he psychometric functi on are indicated in bold), there were substantial changes in intelligibility benefits with SNR for
46 individual talkers. For 10 out of 12 talk er-condition combinations (4 talkers x 3 conditions), the benefit varied by 5 or more percentage points across SNR. The two talker x condition combinations that varied by less than 5 percentage points (T1 and T3 in the clear/slow condition), s uggest that for some talkers clear/slow speech can be somewhat less variable across a range of SNRs. Although the benefit of clear/slow speech for these ta lkers was only affected a small amount by SNR, such a result did not occur for all talkers. In fact, the largest range of benefits observed in any condition was with T5Â’s clear/slow speech (14 points). Conversational/slow speech had similar differences across SNR for individual talkers, but the SNR x condition difference between talkers fell within a somewhat smaller range of 8 points. Specif ically, the conversational/slow benefit was least affected by SNR for talkers T1 and T5, whose benefits each changed 5 points across SNRs, and most affected for T3 whose intelligibility benefit varied by 13 points across SNR. The clear/norma l condition x SNR had the smallest difference between talkers (3 points). Ho wever, the range of benefits across SNR still varied substantially, changing anywher e from 7 points across SNR for T5 to 10 points for T4. Talker variability within condition. A further analysis of variability across talkers within condition was conducted. For each of the three conditions, a scatter plot in Figure 5 represents each talk erÂ’s average intelligibility benefit at each SNR, in relation to the conversational/ normal (baseline) inte lligibility for that SNR. This depiction of t he data allowed for comparis ons between talkers and across SNRs while equating baseline intell igibility, so that the variability in
47 intelligibility benefits for each condition c an be visually examined. As can be seen in Figure 5c, there is little variation over all with clear/slow s peech benefits for all talkers and SNRs. In the baseline range of 20 to 65%, benefits are fairly consistent. There is more variability wit h the clear/normal condition (Figure 5a) and conversational/slow speech (Figure 5b) appears to be highly variable. Moreover, conversational/slow speech show s a strong effect of talker, with T3 obtaining larger benefits than other talkers, even when baseline intelligibility is controlled. Thus, while the changes in inte lligibility benefits fo r individual talkers across SNRs are considerable, benefits ac ross talkers are generally predictable for clear/slow speech, and somewhat pr edictable for clear/normal speech. a) b) c) Figure 5. Intelligibility benef its by talker (T1 blue, T3 green, T4 aqua, T5 red) and SNR, relative to baseline perfo rmance, for each c ondition. Dotted line indicates maximum possible benefit fo r each baseline performance level. 0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100 0 5 10 15 20 25 30 35 40 45 50 Conv/normal ( % correct ) Benefit (relative to conv/normal) -4 -2 0 +2 +4
48 However, even in the clear/slow condition, which was least variable, noticeable talker differences were obser ved. For example, with this condition, T3Â’s conversational/normal speech was 36% intelligible at 0 dB SNR and listeners benefited by 17 perc entage points. At this same baseline measure, but at a more difficult SNR of -4 dB, T5Â’s listeners received an average intelligibility benefit of 23 points. This difference sugges ts that some talkers can achieve a greater intelligibility impr ovement with clear/slow speec h than others, even when baseline performance is equated and all other experimental conditions (e.g. stimuli, listener population, etc.) are held constant.
49 Chapter 5 Discussion Results indicate that, in noise, young normal hearing listeners benefit (relative to conversational speech at normal rates) when talkers speak clearly and/or when talkers speak slowly. Mo re specifically, clear/slow and conversational/slow conditions provide the greatest intelligibility benefits on average, followed by clear/normal speech. Although intelligibility in all conditions generally decreases with worsening SNR, the average intelligibility benefit of the three conditions is maintained in the mi ddle of the psychomet ric function, across signal-to-noise ratios of -4 to 0 dB. Ho wever, when individual talker data within a condition are examined, differences in intelligibility benefits are observed between talkers, on average and across noise levels. When the changes in intelligibility benefits within condition are analyzed, the range of benefits across SNRs can be quite variable between different talkers, even in the middle of the psychom etric function. T5, for example, had the largest change in benefit size across S NRs (14 points, i.e. benefits ranged from 19 to 33 points) for clear/slow speech. T he clear/slow speech benefit of the other three talkers, however, was much less a ffected by SNR. In fact, without T5, the benefit change for each of the three talker s would have ranged from just 2 points (T1) to 5 points (T4) across SNRs. Simila rly, for conversational/slow speech, T3
50 had the largest benefit range across SNRs ( 13 points), but without T3, the effect of SNR on this condition would have been a relatively small change in benefits across talkers, from 5 points (T1, T5) to 7 points (T4) in the middle of the psychometric function. So, the effect of condition and SNR contributed to the variability of benefits within and across talk ers, suggesting that another group of talkers may produce very different results. Reliability Reliability for scoring of listener responses was examined to assess variability with sentence list scorers. As noted previously, a program in Matlab automatically scored each response lis t for key words correct, generating an error file for lists that c ontained one or more errors. Ea ch error file was then rescored and credit was given for misspel led words, homophones, and plural and tense errors. Three graders shared respons ibility for re-scoring these files, creating the possibility for inter-rater re liability issues. Although the grading was largely objective in nature, particular concerns pertaining to reliability included atypical typing and/or tense errors produced by some listeners because these errors required some judgment on the par t of the rater to determine if credit should be given for a key word. To r educe inconsistencies between raters, an Â“ambiguous responseÂ” word list was created, where raters indicated the listener, sentence list, target word, listenerÂ’s response and whether or not credit was received. For example, if the listener ty ped Â“trimbsÂ” for the key word Â“trimsÂ”, credit was given and this response was added to the ambiguous response list. The list was updated by all three raters and then referenced whenever a
51 listenerÂ’s response was questionable, improving both interand intra-rater reliability. As a rough estimate of inter-rater relia bility in scoring, four sentence lists were randomly selected for reliability anal ysis. Each list was examined to check for accuracy of scoring, and ambiguous responses were tabulated. Two lists contained a total of four ambiguous listener responses, one of which was judged to be scored incorrectly. Overall scori ng agreement ranged from 98.7% to 100%, indicating good reliability between raters. Effect of Stimuli The psychometric functions for cl ear/slow and conversational/normal speech (obtained with nonsense sentences) were compared to Liu et al.Â’s (2004) corresponding psychometric functions fo r meaningful sentences, to determine any effect of stimuli. For meaningfu l sentences, Liu at al.Â’s normal hearing listeners received an average intelligibility benefit of 29 percentage points (when baseline intelligibility wa s 50%), for clear/slow speec h presented by one talker. When compared to the current studyÂ’s average benefit (at the same baseline intelligibility) of roughl y 19 percentage points for four talkers, the 10 point difference is most likely attributable to the dissimilar stimuli, as contextual cues in meaningful sentences can im prove listener responses. For Liu et al.Â’s listeners, the advantage of semantic context is likely to have further enhanced the advantage provided by the acoustic properties that increase the intelligibility of clear speech. The increased clear speech benefit for meaningful stimuli appears to be greater for conditions that produce ve ry low (<15%) baseline intelligibility.
52 Figure 6a shows that Liu et al.Â’s listener s obtained benefits up to nearly 60 points in this region, whereas the benefits in t he current study (Figure 6b) were much smaller (0 -15 points). However, only one talker in the current study (T3) produced intelligibility scores that fell bel ow 15% for the more difficult SNRs. Given the level of talker variability observ ed in the current study, it is possible that different results would have been ac hieved with the other three talkers (if smaller SNRs had been used to decrease their baseline intelligibility to <15%).
53 a) Liu et al. (2004) b) Current study 0 20 40 60 80 100 10 0 10 20 30 40 50 60 70 80 90 100 Conv/normal (% correct)Benefit (relative to conv/normal) Clear/slow 4 dB 2 dB 0 dB +2 dB +4 dB Figure 6. Intelligibility benef its of clear/slow speech by talker for Liu et al. (a) versus the current study (b), (T1 bl ue, T3 green, T4 aqua, T5 red) and SNR, relative to baseline performance, for all listeners. Diagonal line indicates maximum possible benefit for ea ch baseline performance level.
54 Talker variability Talker variability merits further consi deration. Of the four talkers in this study, T3 was most different from the other talkers in at least two respects. Her overall intelligibility was the lowest of any talker in the study, and she was the only talker for whom clear/slow was not t he most intelligible condition. If results for T3 were omitted from the analysis, different outcomes would be observed. Across talkers, there woul d be smaller differences in average intelligibility benefits for all conditions. Wi th four talkers, conversa tional/slow was the most variable condition, with a difference of 28 percentage points between talkers with the smallest and largest intelligibility benef its. Without T3, this difference would reduce to 13 points. For clear/normal s peech, the variability in benefits across talkers would decrease from 14 to 7 point s. And, for the clear/slow condition, there would be only a 3 percentage point di fference between intelligibility benefits for T1, T4 and T5 (as compared to a 7 poi nt difference when T3 is included). Therefore, the variations in the size of benefits across talkers for all conditions would have been much smaller, with greater overall predictability, without T3. A possible explanation fo r the variability between talkers involves training. Most talkers in previous clear speech studies were not professionals (Schum, 1996; among others) and were only instruct ed to carefully enunciate their words and/or speak as if they are talking to a person with a hearing loss or in a noisy environment. Minimal training and effo rt on behalf of the talker has been proclaimed as one of the major benefits of clear speech (Schu m). Yet, to further counteract variability across talkers and avoid effects of a talker like T3,
55 additional instruction and/or listener f eedback may be required for more reliable intelligibility results. If additional training does reduce variability, more talker training may be warranted in all fu ture studies of clear speech. Regardless, talker variability was a si gnificant factor in this study, which suggests that it may also have been a fact or in previous clear speech studies. Among similar studies that presented nons ense sentences in noise to young normal hearing listeners, t he number of talkers ranges from one (Helfer, 1997, among others) to five (Kraus e & Braida, 2002; Smiljanic & Bradlow, 2005). If any of these studies had used a single talker with similar intelligibility to T3, their findings would have been considerably diffe rent. If Krause and Braida had only used T3 in their research, for example, they would have f ound greater average intelligibility benefits with clear/normal (20 points) and conversational/slow conditions (33 points) than with clear/s low speech (13 points). In order to counteract any effects of talk er variability, a much larger number of talkers should be used in future clear speech research. Only then can it be determined if T3 is an exception to the average talker, or if researchers should expect one out of every four talkers to exhibit t he characteristics seen with T3. Listener Variability Using Figure 6b, listener variability in the current study can be estimated. If listener variability is small, any two dat a points from the same talker (i.e. the same color) which have comparable bas eline intelligibility should have comparable benefit from clear/slow speec h. While there are many examples where this is the case, there are at least as many examples where listener
56 performance differed substantially. For example, when T5 (red) produced roughly 62% baseline conversational speech intell igibility, one listener obtained a benefit from clear/slow speech of less than 10 poi nts while another listener had a benefit of more than 20 points. Listener variability with the current study can also be compared to Liu et al.Â’s (2004) listeners (Figure 6a). Liu et al noted that listener intelligibility gains resembled an inverted Â“UÂ” shaped curve and maximum benefits were received in the middle of the psychometric functi on. This same phenomenon is visible in Figure 6b for the present studyÂ’s listeners, with the largest intelligibility benefits obtained in the linear portion of t he psychometric function, where conversational/normal speech intelligib ility is between 20 and 65%. While talker variability appears to increase as baselin e intelligibility decreases, listener variability is relatively constant (b etween 0 and roughly 10 points) across all baseline intelligibility levels. This s uggests that listener variability is more Â“random,Â” while talker variabi lity is more likely associated with individual talker characteristics. Other listener issues. One factor that may affect listener variability is the method of scoring used in this study. Res ponse lists were scored on a key word basis, so listeners may still have receiv ed some benefit of context. Listeners were informed that the sentences were grammatically correct, containing true words. When listeners heard only a portion of a word or sentence (e. g. Â“His right cane could ard an edgeÂ”), they may have tried to Â“fill inÂ” the missing word (Â“guardÂ”). With the knowle dge that the word must be a verb, many listeners may
57 have guessed correctly. However, other listenersÂ’ guesses (e.g., Â“cardÂ”) would be scored as completely incorrect, even when their guesses included some of the correct phonemes. To reduce variabilit y between listeners, responses could be scored on a phoneme-by-phoneme basis to increase consistency of the results in these cases. Another factor that could have affected listener variability was the effects of group dynamics on listener responses, an unexpected difficulty encountered in collecting the data. Most listeners complet ed the study as part of a group of four participants. During administration of t he sentence lists, some indications of competitiveness in typing of responses were noted between participants. This Â“rushÂ” to complete a sentence may hav e compromised typing accuracy and/or created poor listening habits. Additionally, t here were participants who attempted to start typing before a s entence presentation was comp leted, possibly impairing other listenersÂ’ ability to hear the stimulus. These issues were addressed by reminding the listeners to take as much time as needed to type their responses. In addition, the listeners were instructed that the sentences had to be separated by a minimum interstimulus interval, so that hurrying would not translate to finishing the list or the session any more quickly. To enforce this notion, when a group appeared to be rushing, the delay between sentence presentations was increased by the tester, who manually cont rolled the rate of presentation of the sentences. For the participants who stat ed that they needed to start typing as soon as each sentence began in order to re member what to type, they were
58 urged to type quietly in order to be cons iderate of the other participants in the room. Several recommendations are offered to eliminate or reduce the effects of individual listener or group issues for futu re studies. First, more instructions and reminders regarding sentence presentati on limitations and typing guidelines should be provided for each group session. As previously noted, there were some listeners who did not participate as part of a group. These listeners controlled the presentation of the sent ences, rather than t he tester, and all expressed a high level of comfort and satisfaction with this testing arrangement. There was no pressure to keep up or work ahead of the group and the listeners could type during the sentence without interfering with other participantsÂ’ listening. Preferably, exper iments should be conducted with individual listeners, or with multiple listeners in separate r ooms, eliminating the problems observed with group sessions. Future Work Future research involving the cu rrent data will include more detailed examination of the slopes of the psychometric functi ons for each talker. Although it appears that slopes are fairly const ant within talkers, there are some exceptions. For example, T5 appears to have one slope for conversational conditions, and a different slope for clear conditions. Statistical analysis of the slopes will help predict if this phenomenon ho lds at more difficult SNRs. And, comparing slopes between talkers will determine whether one talker (e.g. T4) has a flatter slope than the other talkers. This will help to establish if talkerÂ’s slopes
59 are truly different or if they represent different places on the psychometric function. It is also possible that slopes may differ between population s. Therefore it is important to characterize the psyc hometric functions for a given population. Once these psychometric functions ar e established, comparisons can be conducted to determine if clear speech in telligibility levels for a particular population have the same slope as for young normal hearing listeners. For example, it cannot be assumed that childre n with learning disabilities receive the same intelligibility benefits with cl ear speech as young adults, even when baseline (conversational/normal) le vels are equated. The slope of the psychometric function for children may be fl atter or steeper than for young adults, so performance at a given pr esentation level will vary. For these reasons, future clear speech research should, ideally, use full psychometric functions for a given populat ion. Studies should be conducted with multiple talkers and similar stimuli to mini mize the effects of talker and stimuli. However, conducting research with a large num ber of talkers to reduce variability is not always a realistic process due to limited time and financial resources. A possible solution is for researchers to s hare their talker databases, as well as stimuli and type of degradation used (i.e., speech-shaped noise, filtering). If researchers use identical stimuli in the same environments, comparable psychometric functions can be estab lished for multiple populations and differences in intelligibility benefits can be attributed to differences in those
60 populations. By controlling their methods, researchers can make valid comparisons across studies. Furthermore, if researchers conduct pilot experiments to determine the stimulus condition (e.g., SNR) producing approximately 50% in telligibility for a given population, a practica l range of signal-to-noise ratios can be established. This will reduce floor and ceiling effects and concentrate resources on the linear portion of the psychometric function where listeners receive maximum and consistent benefits. This would reduce or possibly eliminate the need for a full psychometric function while still providing baseline performance levels for comparisons. Clinical Implications Clinicians should keep in mind that in telligibility benefits with clear speech are dependent on the listening environment and the talker. This point needs to be emphasized when educating family mem bers and clients about the benefits of clear speech. Enhanced training and prac tice, preferably in different listening situations, should be conducted. Clients s hould be informed that there are some noise situations where acoustic cues of cl ear speech will not ben efit the listener. And, that the benefits of clear speech are dependent on the talkerÂ’s ability to speak clearly as well as the listenerÂ’s hearing acuity. These variability factors should be fully explained to clients and their families to minimize frustration in the event that clear speech does not appear to improve communication. Despite the variability issues identifi ed in this thesis, the many clinical implications of clear speech as an effe ctive and practical m ode of communication
61 remain. By using clear speech with all populations, clinicians can be assured their communicative message is enhanced and that clients will benefit. In a clinical setting, clients are frequently presented with information in a technical format or using medical jargon. Presenti ng that information in a clear, more intelligible form may enable the client to focus more on the communicative meaning by lightening their Â“cognitive l oad.Â” Additionally, clear speech requires minimal effort but offers immeasurabl e rewards through improved communication and relationships with clients.
62 References Bradlow, A. R., & Bent, T. (2002). The clear speec h effect for non-native listeners. Journal of the Acoustical Society of America, 112 272-284. Bradlow, A. R., Kraus, N., & Hayes, E. (2003). Speaking clearly for children with disabilities: Sentence perception in noise. Journal of Speech, Language and Hearing Research 46, 80-97. Ferguson, S. H., & Kewley-Port, D. ( 2002). Vowel intelligibility in clear and conversational speech for normal-hear ing and hearing-impaired listeners. Journal of the Acoustical Society of America, 112, 259-271. French, N. R., & Steinberg, J. C. (1947). Factors governi ng the intelligibility of speech sounds. The Journal of the Acoustica l Society of America, 19 90119. Gagne, J., Rochette, A., & Charest, M. (2002). Auditory, visual and audiovisual clear speech. Speech Communication, 37 213-230. Goy, H., Pichora-Fuller, K., van Lieshou t, P., Singh, G., & Schneider, B. (2007, August). Effect of withinand between-talker variability on word identification in noise by younger and older adults. Paper accepted for presentation at the meet ing of Interspeech 2007, Antwerp, Belgium.
63 Helfer, K. S. (1997). Auditory and auditory-visual perception of clear and conversational speech. Journal of Speech, Language and Hearing Research, 40 432-443. Helfer, K. S. (1998). Auditory and audi tory-visual recognition of clear and conversational speech by older adults. Journal of the American Academy of Audiology, 9 234-242. Krause, J. C., & Braida, L. D. (2002). Investigating alternative forms of clear speech: The effects of speaking rate and speaking mode on intelligibility. Journal of the Acoustical Society of America, 112, 2165-2172. Krause, J. C., & Braida, L. D. (20 03). Effects of list ening environment on intelligibility of clear speec h at normal speaking rates. Iranian Audiology, 2 39-47 Krause, J. C., & Braida, L. D. (2004). Acoustic properties of naturally produced clear speech at normal speaking rates. Journal of the Acoustical Society of America, 115, 362-378. Liu, S., Del Rio, E., Bradlow, A. R., & Z eng, F. (2004). Clear s peech perception in acoustic and electric hearing. Journal of the Acoustical Society of America, 116, 2374-2383. Miller, G. A., & Nicely, P. E. (1955). An analysis of perceptual confusion among some English consonants. The Journal of the Acoustical Society of America, 2, 338-352.
64 Nillson, M., Soli, S. D., & Sullivan, J. A. (1994). Deve lopment of the hearing in noise test for the measurement of s peech reception thresholds in quiet and in noise. Journal of the Acoustical Society of America, 95, 1085-1099. Panagiotopoulos, A. P. (2005). The benefits of clear speech at normal rates for older listeners with normal hearing Unpublished masterÂ’s thesis, University of South Florida, Tampa. Payton, K. L., Uchanski, R. M., & Brai da, L. D. (1994). Intelligibility of conversational and clear speech in noi se and reverberation for listeners with normal and impaired hearing. Journal of the Acoustical Society of America, 95, 1581-1592. Picheny, M. A., Durlach, N. I., & Braida, L. D. (1985). Speaking clearly for the hard of hearing I: Intelligibilit y differences between clear and conversational speech. Journal of Speech and Hearing Research, 28, 96103. Picheny, M. A., Durlach, N. I., & Braida, L. D. (1986). Speaking clearly for the hard of hearing II: Acoustic characte ristics of clear and conversational speech. Journal of Speech and Hearing Research, 29, 434-446. Picheny, M. A., Durlach, N. I., & Braida, L. D. (1989). Speaking clearly for the hard of hearing III: An attempt to determine the contribution of speaking rate to differences in intelligibility between clear and conversational speech. Journal of Speech and Hearing Research, 32, 600-603.
65 Schum, D. J. (1996). Intelli gibility of clear and conversational speech of young and elderly talkers. Journal of the American Ac ademy of Audiology, 7 212-218. Schum, D. J. (1997). Bey ond hearing aids: Clear speech training as an intervention strategy. The Hearing Journal, 50 (10), 36-40. Smiljanic, R., & Bradlow, A. R. (2005). Production and per ception of clear speech in Croatian and English. Journal of the Acoustical Society of America, 118, 1677-1688. Uchanski, R. M., Choi, S. S., Braida, L. D. Reed, C. M., & Durl ach, N. I. (1996). Speaking clearly for the hard of hearing IV: Further studie s of the role of speaking rate. Journal of Speech and Hearing Research, 39 494-509. Wilson, R. H., & Strouse, A. L. (1999). A uditory measures with speech signals. In F. E. Musiek & W. F. Rintlemann (Eds.), Contemporary perspectives in hearing assessment (pp. 21-66). Needham Height s, MA: Allyn & Bacon.
67 Appendix A: Participant Demographics Participant GenderAgeParticipantGender Age L1 F 23 L21 F 23 L2 F 19 L22 F 25 L3 F 38 L23 F 22 L4 F 21 L24 F 29 L5 F 22 L25 F 23 L6 M 21 L26 M 37 L7 F 22 L27 F 25 L8 M 24 L28 M 21 L9 F 20 L29 F 18 L10 F 21 L30 F 20 L11 F 27 L31 M 28 L12 F 26 L32 F 37 L13 M 18 L33 F 18 L14 F 20 L34 F 19 L15 F 21 L35 F 21 L16 F 33 L36 M 18 L17 M 19 L37 F 19 L18 F 33 L38 M 19 L19 F 21 L39 F 34 L20 F 19 L40 F 24
68 Appendix B: Key Word Scores by Talker Table B1. T1 % Correct SNR Listener Conv/normal Clear/ normal Conv/slow Clear/slow -4 L1 29.1 34.1 49.7 53.8 -4 L2 21.2 43.9 56.1 49.7 -4 L3 22.3 25.4 40.4 45.6 -4 L4 20.1 30.6 45.6 48.0 -4 L33 21.8 32.4 49.1 49.1 -4 L34 29.6 32.4 51.5 47.4 -4 L35 40.2 38.7 57.9 59.1 -4 L36 16.2 25.4 40.9 34.5 -2 L5 33.0 34.1 52.6 58.5 -2 L6 41.3 49.7 64.3 57.9 -2 L7 35.8 49.7 73.1 59.6 -2 L8 31.8 57.8 71.3 62.0 -2 L37 56.4 44.5 62.6 75.4 -2 L38 57.5 56.1 61.4 68.4 -2 L39 44.7 49.1 70.8 74.3 -2 L40 65.4 52.6 64.9 79.5 0 L9 58.7 74.6 81.9 75.4 0 L10 48.0 60.1 80.7 72.5 0 L11 58.7 63.0 77.8 76.0 0 L12 52.5 63.6 80.1 80.7 0 L21 58.7 61.8 71.9 78.9 0 L22 46.4 53.8 64.9 71.9 0 L23 60.9 71.7 77.8 75.4 0 L24 41.3 64.7 74.3 70.8 2 L13 72.6 76.9 85.4 84.8 2 L14 79.9 79.8 86.5 90.6 2 L15 60.3 70.5 80.1 80.7 2 L16 66.5 71.1 87.1 85.4 2 L25 69.3 68.2 77.8 76.6 2 L26 80.4 77.5 78.9 86.0 2 L27 71.5 81.5 88.9 89.5 2 L28 69.8 61.3 79.5 88.3 4 L17 62.6 86.1 90.1 81.9 4 L18 71.5 72.8 80.7 88.9 4 L19 82.1 82.1 89.5 91.8 4 L20 83.2 82.7 92.4 92.4 4 L29 80.4 85.5 85.4 84.2 4 L30 70.9 75.7 84.8 86.0 4 L31 79.3 72.3 86.0 93.0 4 L32 79.3 84.4 90.6 91.2
69 Appendix B (continued) Table B2. T3 % Correct SNR Listener Conv/normal Clear/ normal Conv/slow Clear/slow -4 L9 12.1 20.8 33.1 18.3 -4 L10 13.9 32.9 42.3 12.6 -4 L11 14.5 34.1 41.7 16.0 -4 L12 11.0 43.9 54.3 20.0 -4 L37 8.7 43.4 50.3 15.4 -4 L38 12.7 24.3 34.9 16.6 -4 L39 8.1 37.6 46.3 8.0 -4 L40 8.7 34.1 47.4 19.4 -2 L13 20.2 54.3 58.9 35.4 -2 L14 28.3 43.4 58.9 45.7 -2 L15 17.9 41.6 66.9 28.6 -2 L16 13.9 51.4 57.7 34.3 -2 L21 21.4 35.8 55.4 30.9 -2 L22 19.1 45.1 55.4 38.3 -2 L23 27.7 54.9 70.3 40.6 -2 L24 17.3 29.5 59.4 26.3 0 L17 28.3 52.0 76.0 39.4 0 L18 43.4 58.4 73.7 48.0 0 L19 37.6 47.4 70.9 57.1 0 L20 50.9 48.6 66.3 62.9 0 L25 28.9 56.6 73.1 52.0 0 L26 31.8 53.8 65.7 46.3 0 L27 32.9 54.3 69.1 52.0 0 L28 35.8 56.6 76.0 53.1 2 L1 52.0 74.0 85.7 62.3 2 L2 49.1 69.9 75.4 66.3 2 L3 43.9 70.5 86.9 72.0 2 L4 53.2 71.1 84.6 67.4 2 L29 47.4 63.6 82.9 66.3 2 L30 45.1 72.8 83.4 56.6 2 L31 48.6 80.9 85.1 65.7 2 L32 56.6 60.7 77.7 74.3 4 L5 65.9 79.2 84.6 74.3 4 L6 69.4 82.1 94.9 78.3 4 L7 58.4 75.7 87.4 74.9 4 L8 54.3 73.4 89.1 73.1 4 L33 50.9 74.6 82.9 71.4 4 L34 64.7 78.6 89.7 73.1 4 L35 74.6 80.9 93.1 82.3 4 L36 63.6 85.5 90.3 82.9
70 Appendix B (continued) Table B3. T4 % Correct SNR Listener Conv/normal Clear/ normal Conv/slow Clear/slow -4 L13 31.2 40.0 40.5 57.8 -4 L14 40.6 28.8 31.8 74.0 -4 L15 25.3 34.1 39.3 52.0 -4 L16 35.3 29.4 37.0 52.0 -4 L21 32.9 29.4 22.0 47.4 -4 L22 30.0 38.8 42.2 57.8 -4 L23 35.9 52.4 52.0 60.7 -4 L24 24.1 30.0 30.1 52.0 -2 L17 31.2 54.7 57.2 55.5 -2 L18 47.6 56.5 54.9 67.1 -2 L19 44.1 45.9 39.3 72.8 -2 L20 55.3 53.5 53.8 71.7 -2 L25 44.7 66.5 60.1 59.0 -2 L26 43.5 52.4 62.4 66.5 -2 L27 49.4 58.8 55.5 74.6 -2 L28 44.1 75.3 71.7 69.4 0 L1 65.3 68.8 74.0 83.2 0 L2 64.1 51.2 52.6 81.5 0 L3 60.6 68.8 65.9 86.1 0 L4 65.3 68.2 67.6 76.9 0 L29 55.9 66.5 66.5 77.5 0 L30 41.2 62.9 67.6 70.5 0 L31 62.4 74.1 72.8 80.9 0 L32 57.6 58.2 54.3 75.7 2 L5 69.4 76.5 77.5 82.7 2 L6 71.8 90.6 83.2 88.4 2 L7 67.1 70.0 61.8 86.7 2 L8 66.5 68.8 72.3 85.0 2 L33 54.7 74.1 73.4 87.9 2 L34 73.5 78.8 78.0 88.4 2 L35 82.9 85.3 74.6 92.5 2 L36 62.9 82.4 76.9 85.0 4 L9 68.8 77.1 71.7 92.5 4 L10 73.5 83.5 82.1 76.9 4 L11 80.6 82.9 83.8 94.2 4 L12 82.9 85.3 83.2 93.6 4 L37 84.7 83.5 80.3 87.9 4 L38 68.8 75.3 71.1 89.0 4 L39 75.9 85.3 81.5 90.8 4 L40 82.9 87.1 85.0 79.8
71 Appendix B (continued) Table B4. T5 % Correct SNR Listener Conv/normal Clear/ normal Conv/slow Clear/slow -4 L17 27.459.442.3 57.7 -4 L18 33.154.936.0 64.0 -4 L19 40.664.040.0 71.4 -4 L20 46.956.049.1 76.0 -4 L29 37.763.440.6 70.9 -4 L30 24.050.930.3 60.0 -4 L31 34.954.938.9 70.9 -4 L32 38.961.756.6 71.4 -2 L1 62.973.160.0 78.3 -2 L2 54.380.072.6 70.3 -2 L3 52.670.346.3 80.0 -2 L4 60.072.062.3 80.0 -2 L33 41.766.956.6 76.0 -2 L34 56.070.958.3 82.3 -2 L35 64.077.770.9 84.0 -2 L36 50.960.656.6 72.0 0 L5 64.064.662.3 78.9 0 L6 64.677.773.1 86.3 0 L7 60.077.777.1 81.1 0 L8 69.184.081.1 86.9 0 L37 62.378.964.6 86.3 0 L38 62.388.677.7 72.6 0 L39 57.787.473.7 82.3 0 L40 71.482.378.3 86.3 2 L9 84.684.684.6 87.4 2 L10 76.684.685.1 80.0 2 L11 79.486.988.6 89.7 2 L12 86.391.490.3 89.1 2 L21 85.187.482.9 89.7 2 L22 78.977.161.1 86.9 2 L23 82.985.784.0 87.4 2 L24 76.091.484.0 88.6 4 L13 82.989.186.3 89.1 4 L14 89.793.190.9 93.7 4 L15 74.982.985.7 81.1 4 L16 78.385.186.9 90.9 4 L25 84.686.384.0 92.6 4 L26 91.492.690.3 89.1 4 L27 93.793.796.0 96.0 4 L28 90.3 86.3 85.7 89.7
72 Appendix C: Pair wise Comparisons Table C1. Pairwise Comparisons by Condition Dependent Variable: RAU Tukey HSD 95% Confidence Interval (I) Condition (J) Condition Mean Difference (I J) Std. Error Sig. Lower Bound Upper Bound Clear/normal Clear/slow Conv/normal Conv/slow -.062312* .126107* -.046514 .0081131 .0081131 .0081131 .000 .000 .000 -.083218 .105201 -.067420 -.041407 .147012 -.025609 Clear/slow Clear/normal Conv/normal Conv/slow .062312* .188419* .015798 .0081131 .0081131 .0081131 .000 .000 .210 .041407 .167514 -.005107 .083218 .209325 .036704 Conv/normal Clear/normal Clear/slow Conv.slow -.126107* -.188419* -.172621* .0081131 .0081131 .0081131 .000 .000 .000 -.147012 -.209325 -.193527 -.105201 -.167514 -.151716 Conv/slow Clear/normal Clear/slow Conv/normal .046514* -.015798 .172621* .0081131 .0081131 .0081131 .000 .210 .000 .025609 -.036704 .151716 .067420 .005107 .193527 Based on observed means. *The mean difference is signifi cant at the 0.05 level.
73 Appendix C (c ontinued) Table C2. Pairwise Comparisons by Talker Dependent Variable: RAU Tukey HSD 95% Confidence Interval (I) Talker (J) Talker Mean Difference (I J) Std. Error Sig. Lower Bound Upper Bound T1 T3 T4 T5 .133969* .002895 -.092616* .0081131 .0081131 .0081131 .000 .984 .000 .113063 -.018010 -.113522 .154874 .023801 -.071711 T3 T1 T4 T5 -.133969* -.131074* -.226585* .0081131 .0081131 .0081131 .000 .000 .000 -.154874 -.151979 -.247490 -.113063 -.110168 -.205679 T4 T1 T3 T5 -.002895 .131074* -.095511* .0081131 .0081131 .0081131 .984 .000 .000 -.023801 .110168 -.116417 .018010 .151979 -.074606 T5 T1 T3 T4 .092616* .226585* .095511* .0081131 .0081131 .0081131 .000 .000 .000 .071711 .205679 .074606 .113522 .247490 .116417 Based on observed means. *The mean difference is signifi cant at the 0.05 level.