Perceptual contributions of the consonant-vowel boundary to sentence intelligibility

Perceptual contributions of the consonant-vowel boundary tosentence intelligibilitya)

Daniel Fogertyb� and Diane Kewley-PortDepartment of Speech and Hearing Sciences, Indiana University, Bloomington, Indiana 47405

�Received 22 September 2008; revised 3 June 2009; accepted 4 June 2009�

Although research has focused on the perceptual contribution of consonants to spoken syllable orword intelligibility, in sentences vowels have a distinct perceptual advantage over consonants indetermining intelligibility �Kewley-Port et al., J. Acoust. Soc. Am. 122, 2365–2375 �2007��. Thecurrent study used a noise replacement paradigm to investigate how perceptual contributions ofconsonants and vowels are mediated by transitional information at segmental boundaries. Thespeech signal preserved between replacements is defined as a glimpse window. In the firstexperiment, glimpse windows contained proportional amounts of transitional boundary informationthat was either added to consonants or deleted from vowels. Results replicated a two-to-one voweladvantage for intelligibility at the traditional consonant-vowel boundary and suggest that vowelcontributions remain robust against proportional deletions of the signal. The second experimentexamined the combined effect of random glimpse windows not locked to segments and thedistributions of durations measured from the consonant versus vowel glimpses observed inExperiment 1. Results demonstrated that, for random glimpses, the cumulative sentence durationglimpsed was an excellent predictor of performance. Comparisons across experiments confirmedthat higher proportions of vowel information within glimpses yielded the highest sentenceintelligibility. © 2009 Acoustical Society of America. �DOI: 10.1121/1.3159302�

PACS number�s�: 43.71.Es, 43.71.Gv �JES� Pages: 847–857

I. INTRODUCTION

All languages share a common feature: the use of bothconsonant and vowel sounds �Ladefoged, 2001�. Intelligibil-ity research has focused on these two fundamental groups ofspeech sounds, attempting to tease apart the relative percep-tual contributions of each. The traditional view of these con-tributions has been stated clearly, as “it has become almost acommonplace statement in intelligibility testing that most ofthe information in speech is carried by the consonantsounds” �Owens et al., 1968, p. 648�. More recently, thistraditional view has been questioned by Cole et al. �1996�and Kewley-Port et al. �2007�. Both studies demonstrated anoverwhelming advantage of vowels compared to consonantsin sentence intelligibility testing using a segment replace-ment paradigm, also described as “noise replacement.” Thisparadigm turns successive portions of the speech signal onand off, effectively alternating the speech signal with eithersilence or noise. The “on” portions of speech are defined asglimpses �Cooke, 2003�. Noise replacement has been usedpreviously to examine the specific perceptual contributionsof certain speech acoustics contained in consonant or vowelsegments �e.g., Kewley-Port et al., 2007�, as well as, theperception of interrupted speech �e.g., Miller and Licklider,1950�. While these studies were able to isolate the perceptualcontributions of acoustic events present during the segment

a�Portions of the data were presented at the 154th Meeting of the AcousticalSociety of America �J. Acoust. Soc. Am. 122, 2972 �2007�� and at the156th Meeting of the Acoustical Society of America �J. Acoust. Soc. Am.124, 2439 �2007��.

b�Author to whom correspondence should be addressed. Electronic mail:
[email protected]
J. Acoust. Soc. Am. 126 �2�, August 2009 0001-4966/2009/126�2

of interest, it is well known that acoustic cues for segmentsare distributed across consonant-vowel �C-V� boundaries �asshown for stop consonants, Liberman et al., 1967�. It is alsoclear from work with silent-center syllables that the dynamicportions of the speech signal at these C-V boundaries pro-vide important contributions �Strange, 1987; Strange et al.,1983�.

In the current study, Experiment 1 sought to replicateand extend past findings by investigating how the perceptualcontributions of consonants and vowels are affected by sys-tematically varying amounts of acoustic information at theC-V boundary. Experiment 2 was conducted as a control toisolate potential confounds of the noise replacement para-digm, in order to confirm that the perceptual pattern revealedin Experiment 1 was a result of spectro-temporal informationin segments, and not of particular methodological properties.In addition, Experiment 2 compared the perception of ran-domly ordered glimpses, the preserved portions of speechbetween replacements, to glimpses locked to specific seg-ments in Experiment 1.

Segmentation of the highly dynamic acoustic waveforminto discrete units corresponding to consonants and vowels isinherently ambiguous. Particularly problematic is the assign-ment of formant transitions to either the consonant or vowelsegment, as these transitions provide information about bothconsonants and vowels �Liberman et al., 1967�. Normalcoarticulation means that acoustic cues for consonants andvowels largely overlap; the perceptual effects of which maybe observed over multiple segments �see Kent and Minifie,1977�. Therefore, it may be that a discrete division between
consonants and vowels is largely a convenience �see Lade-
© 2009 Acoustical Society of America 847�/847/11/$25.00

foged, 2001�. Stevens and his colleagues �summarized byStevens, 2002� long advocated that prominent acoustic land-marks may be used to effectively mark different segmentswithin the ambiguous speech signal. Therefore, identifyingthe C-V boundary is a convenient methodological tool fordividing the acoustic signal into mostly vocalic and mostlyconsonantal portions. However, the perceptual cues presentwithin the acoustic signals associated with consonants andvowels interact with each other. Cooper et al. �1952� wereamong the first to explore the relation between perceptionand acoustic cues in stop consonants, such as the frequencyrelease of the burst and the direction of the second formanttransition. They described the perception of these specificcues as dependent on their positional relationship to theneighboring vowel. Furthermore, the perception of speechsounds is highly dependent on the acoustic-phonetic context�Miller, 1994�. Therefore, the perception of consonants andvowels and their contributions to sentence intelligibility ap-pear to be the result of distributed and overlapping cues.

The current study uses the fundamental grouping ofspeech sounds into consonants and vowels to globally ex-plore how the acoustic information contained within thesesegments contributes to sentence intelligibility. In addition,as the perceptual information of these segments is highlydistributed and overlapping, this study investigates how seg-mental contributions change with the addition or deletion oftransitional information between them. The current studywas not designed to redefine discrete segmental boundariesbut rather to investigate the relative perceptual contributionsof these two primary categories of speech sounds.

A. Previous research on the relative importance ofconsonants and vowels

There have been several approaches to understandingthe relative contributions of consonants and vowels to speechperception. Studies that have investigated the consonant-to-vowel ratio �i.e., the difference between consonant andvowel intensity levels� have found that amplifying the inten-sity of the consonant relative to the vowel in CVCs enhancesrecognition of voiced stops �Freyman and Nerbonne, 1989;Freyman et al., 1991�, which corroborates well with the im-portance of consonant acoustics for speech intelligibility.However, Sammeth et al. �1999� used a different approachthat controlled for the intensity of three stop consonants.They maintained the intensity level of the consonant in CVsyllables while decreasing the intensity level of the vowel by6 or 12 dB. Bringing the consonant-to-vowel ratio to 0 dBdid not improve identification performance of the conso-nants. However, in some conditions the consonants were ef-fectively presented in isolation, representing a maximum at-tenuation for the vowel. Reintroduction of the vowelproduced a large and significant improvement in identifica-tion, suggesting that the presence of the vowel was importantfor the successful identification of the consonants.

Much of the work that has championed the importanceof consonant acoustics over that of the vowel has investi-gated this issue at the level of the syllable structure. Owrenand Cardillo �2006� took this one step further by investigat-
ing multisyllabic words in isolation. They used a segment
848 J. Acoust. Soc. Am., Vol. 126, No. 2, August 2009 D. Fogerty

replacement paradigm to present consonant-only words orvowel-only words, where segments and formant transitionswere replaced by silence. Listeners heard pairs of words andwere asked to make two judgments: whether the two wordsshared the same meaning and whether they were spoken bythe same talker. Results demonstrated that consonant-onlywords yielded higher d� scores for judging the meaning;vowel-only words had higher scores for judging the talker.Owren and Cardillo �2006� concluded that consonants maybe more important for intelligibility, while vowels carrymore indexical information about the talker.

Research at the sentence level has revealed a much dif-ferent picture. In a preliminary report, Cole et al. �1996� useda segment replacement paradigm where segments were eitherreplaced by speech-shaped noise, a harmonic complex, orsilence. Words repeated correctly were scored. Their resultssuggested that vowel-only sentences maintained a two-to-one advantage over consonant-only sentences, regardless ofthe type of segmental replacement. Furthermore, this advan-tage remained, even after 10 ms was deleted from the onsetand offset of the vowels. This two-to-one advantage of vow-els was later replicated by Kewley-Port et al. �2007� whennormal-hearing listeners were presented sentences at 70 dBsound pressure level �SPL�, as well as for elderly hearing-impaired listeners at a higher presentation level of 95 dBSPL. For young normal-hearing listeners, the 95 dB SPLlevel resulted in an improvement for recognition ofconsonant-only sentences, but participants still maintainedsignificantly better performance for vowel-only sentences.This advantage for vowels is even more remarkable consid-ering that vowels actually comprise at least 10% less of theoverall sentence duration than consonants �Ramus et al.,1999�.

As perceptual cues are highly distributed throughout thespeech signal, it could be that certain subsegmental informa-tion plays an important role as well. Lee and Kewley-Port�2009� investigated four types of subsegmental informationthat presented equal proportions of consonants and vowels ineach condition. Segmental onset and offset, as well as tran-sitions and center portions, were investigated. These portionsof segmental information were based relative to traditionalsegmental boundaries. Performance based on words repeatedcorrectly was equal across all conditions, indicating no ad-vantage for any type of subsegmental information whenequal proportions of consonant and vowel information weremaintained. However, while transitional and center portionsof segments may make equal perceptual contributions tooverall sentence intelligibility, it remains to be seen howthese two different types of acoustic signals interact.

The purpose of the current study was to investigate howtransitional information at the C-V boundary as traditionallydefined facilitates the perceptual contributions made by con-sonants and vowels to sentence intelligibility. It has becomewidely accepted that transitional dynamics at the C-V bound-ary contains important information for perception �Libermanet al., 1967; Strange, 1987; Strange et al., 1983�, but it is lessclear how this information relates to the perceptual contribu-
tions of consonants and vowels. Our investigation explores
and D. Kewley-Port: Contributions of the consonant-vowel boundary

how certain types of speech information contained in conso-nants, vowel centers, and the transitions between them inter-act in sentential contexts.

B. Glimpsing speech during noise replacement

In studies that use noise replacement to isolate percep-tual contributions, it may be that preserved perceptual cuesin the signal are not the only performance factor, but alsohow those cues interact with a listener’s ability to integrateinformation across preserved portions of the speech signal.The current study examined these perceptual consequencesthat are inherent in interrupted speech produced during noisereplacement. The portions preserved between replacementsare described here as glimpses, taken from the literature oninterrupted speech in noise. Cooke �2003� defined a glimpseas “an arbitrary time-frequency region which contains a rea-sonably undistorted view of the target signal” �p. 579�. Ineveryday listening conditions, it may be most beneficial for alistener to focus on the glimpses of speech between back-ground noise fluctuations, thereby attending to the preservedspeech information rather than trying to extract cues from anoisy signal. The speech intelligibility index has recentlybeen modified to include consideration of these temporalfluctuations and now provides a good account for speechrecognition thresholds in interrupted speech �Rhebergen andVersfeld, 2005�. The redundancy of the speech signal makesthe use of such glimpses effective in speech perception�Cooke, 2006�.

In glimpse studies, in order to maintain the same amountof speech information presented across all conditions there isan acoustic trade-off between interruption rate �i.e., the fre-quency of available glimpses� and glimpse window width�i.e., the duration of the available speech information�. Millerand Licklider �1950� demonstrated that perceptual perfor-mance for monosyllabic words is affected nonmonotonicallyas the glimpse rate is increased and the glimpse windowwidth is decreased with the best performance at a glimpserate around 10 Hz. Therefore, properties of glimpses, both inrate and size, may effect perception in noise replacementstudies. This may also be related to the type of speech ma-terial used, as the perception of speech continuity is main-tained at shorter glimpse durations for discourse than forisolated words �Bashford and Warren, 1987� and is associ-ated with increased speech intelligibility when higher-ordercontextual cues are available �Bashford et al., 1992�.

In order to investigate the segmental contributions ofconsonants and vowels independently of glimpse contribu-tions, a second experiment was conducted. Experiment 2used the same distributions of glimpse window durationsmeasured from consonants and vowels, but randomly reor-dered them, thereby, creating glimpses with random portionsof both consonants and vowels. This allowed for the directcomparison of �1� any difference in performance due to thedifferent glimpse durations measured from consonants andvowels and �2� performance when windows were eitherlocked to specific segmental information or when they con-
tained random acoustic cues.
J. Acoust. Soc. Am., Vol. 126, No. 2, August 2009 D. Fogerty and D.

II. EXPERIMENT 1: SEGMENTAL CONTRIBUTIONS

This first experiment was designed to answer somequestions raised in previous studies conducted in our labora-tory �Kewley-Port et al., 2007; Lee and Kewley-Port, 2009�.Results showed that using traditional segmental rules, vow-els contributed more than consonants to the perception ofsentences �Kewley-Port et al., 2007�, and that there appearedto be no advantage of transitional versus quasi-steady-stateinformation when the relative proportion of consonant-to-vowel segment durations remained constant �Lee andKewley-Port, 2009�. However, given that perceptual infor-mation is distributed across the transitional area at the C-Vboundary, a major unanswered question concerns the domi-nance of vowel information obtained in previous studies inrelation to the allocation of transitional C-V boundary infor-mation. Thus, the present investigation systematically shiftedthe C-V boundary to examine how transitional informationcontributes to the perceptual roles of consonants and vowelsfor sentence intelligibility.

A. Listeners

Twenty young �age 19–30, M =24� normal-hearing lis-teners �six male, fourteen female� were paid to participate inthe study. All participants were native speakers of AmericanEnglish and had pure-tone thresholds bilaterally at no greaterthan 20 dB HL at octave intervals from 250 to 8000 Hz�ANSI, 1996�.

B. Stimuli

1. Speech presented in sentences

The same 42 sentences used in previous studies�Kewley-Port et al., 2007; Lee and Kewley-Port, 2009� wereselected from the TIMIT database �Garofolo et al., 1990,www.ldc.upenn.edu� to be used in the current study. Thesesentences were spoken by 21 male and 21 female talkersfrom the North Midland dialect region that matched the dia-lect region of daily exposure for the participants in this study�i.e., Indianapolis, Indiana and further north�.

Sentences contained an average of 8 words �range 6–10words� and 33 phonemes �11 vowels, 22 consonants�. TheC-V boundaries are specified by the TIMIT database. Pho-nemes were identified by traditional segmental boundaries,although TIMIT first employed CASPAR �Leung and Zue,1984� automatic segmentation to identify the boundaries,which experienced phoneticians later confirmed �see Zue andSeneff, 1988�. The TIMIT segmentation methods includedrules such as �1� finding highly salient and abrupt acousticchanges to mark phoneme boundaries for stops and �2� di-viding formant transitions in half during slow periods ofchange for liquids �presumably because these transitions pro-vide information regarding both phonemes�. Kewley-Portet al. �2007� verified the segmental boundaries provided bythe TIMIT corpus and added three minor rules appropriatefor identifying the vowels and consonants in sentences.These rules were used in the present study and are noted
below using the TIMIT arpabet.
Kewley-Port: Contributions of the consonant-vowel boundary 849

�1� Stop closures �e.g., “bcl”� were combined with the stop�e.g., “b”� and treated as a single consonant.

�2� Syllable final V+ �r� as in “deer” were transcribed inTIMIT as a vowel plus the consonant �r�. These V+ �r�transcriptions were considered as a single rhotocizedvowel in this study.

�3� The glottal stop �q� was a separate consonant in TIMIT,generally realized by silence. When the glottal stop wastranscribed as occurring between two vowels, �VqV�, itwas treated as part of a vowel string.

For the purposes of the noise replacement paradigmused here and elsewhere �see Kewley-Port et al., 2007�, con-sonant strings were treated as a single unit for replacement,as were vowel strings. Therefore, the experimental manipu-lation of C-V boundaries was only between consonant andvowel units. The acoustic boundary was shifted by a propor-tion of the vowel duration �VP�, thereby, reassigning voweltransitions, because previous research had demonstrated rela-tively low performance for consonant-only stimuli �Kewley-Port et al., 2007�. It was expected that manipulating by aproportion of the vowel would yield the most interpretabledata by avoiding floor performance. This method allows di-rect investigation of how transitional information typicallyassigned to vowel segments influences the perceptual contri-butions of vowels and consonants.

Two processing strategies each created a different stimu-lus type. One strategy preserved the consonant informationand added some proportion of the vowel transitions �C+VP�, replacing the vowel centers with noise. The secondstrategy preserved only a proportion of the vowel center in-formation, replacing the consonants and remaining voweltransitions at the C-V boundary with noise �V�VP�. Theseprocessing strategies were manipulated by shifting the C-Vboundary by six vowel proportion amounts �2�6 design�,yielding a total of 12 conditions. Figure 1 shows a schematicof these 12 conditions. The schematic example shows thenoise replacement for a single CVC; however, this methodwas applied to all consonant and vowel units across the en-tire sentence. The conditions are labeled in two ways on the

FIG. 1. Schematic of noise replacement conditions depicted for a singleCVC within the sentence �not drawn to scale�. The C only and V only �forC+0 and V-0 conditions� uses the original TIMIT C-V boundaries. Solidbars indicate the speech portions presented. Stippled bars indicate noisereplacement.

figure, first according to how the C-V boundary was shifted


�e.g., C+5, meaning the original consonant segment plus 5%of the vowel�, and second, by an equation of the proportionof the segments preserved �e.g., C+.05V�. The rest of thismanuscript will use the first label type in referring to theseconditions.

2. Noise presented in sentences

Segments were replaced by a speech-shaped noise gen-erated in MATLAB that was based on a standard long-term-average speech spectrum �ANSI, 1969�. The noise was de-signed to be flat from 0 to 500 Hz and had a −9 dB/octaveroll-off above 500 Hz. The noise level was presented 16 dBbelow the average sentence level �21 dB below the vowelwith loudest rms�. This level was used in a previous study forconsonant noise replacement �Kewley-Port et al., 2007� toavoid any potential for temporal masking of consonant seg-ments. Every segment was replaced using a unique noise.Figure 2 displays the waveform of a single CVC, “mean,”excised from an experimental sentence under vowel-only andconsonant-only noise replacement conditions with the fullwaveform as reference.

C. Calibration

Similar procedures to Kewley-Port et al. �2007� and Leeand Kewley-Port �2009� were administered to verify signallevels. First, scripts were written in MATLAB to verify that all42 test sentences had similar average rms levels �i.e., within�2 dB�. Second, a MATLAB script was used to find the mostintense vowel across all sentences and then iterated thatvowel to produce a calibration vowel of 4 s. As these stimuliwere previously used to evaluate sentence intelligibility withlisteners who had hearing impairment, all sentences and thecalibration vowel were filtered by a low-pass FIR filter thatwas flat to 4000 Hz with a 3 dB cutoff at 4400 Hz, and a200 dB/octave slope in a Tucker-Davis Technologies pro-grammable filter. Therefore, this allows direct comparison ofthis study with previous ones �Kewley-Port et al., 2007; Leeand Kewley-Port, 2009�. The sound level for the calibrationvowel was set to 100 dB SPL through ER-3A insert ear-

3

FIG. 2. Waveforms demonstrating two noise replacement conditions andthe full waveform of the word ‘‘mean’’. V only � vowel-only;C only � consonant-only.

phones in a HA-2.2-cm coupler using a Larson Davis model


2800 sound level meter with linear weighting. Relative to theloudest calibration vowel, the mean of the distribution of theloudest vowels in the other sentence was 95 dB SPL. Allsentences and the calibration vowel were passed through aprogrammable attenuator to present the stimuli at 70 dBSPL, being the nominal level referenced in this study. Anadditional low-level background noise was continuously pre-sented during testing. The purpose of this noise was to re-duce transients between speech and the replacement noise.This background noise was generated by a Tucker-DavisTechnologies waveform generator, and was also low-pass fil-tered at 4400 Hz. The level of this noise was presented at50 dB less than the calibration vowel �100 dB SPL� mea-sured using the same equipment described above.

D. Procedure

Listeners were seated in a sound attenuated booth andlistened to sentences presented monaurally to the right earvia ER-3A insert earphones. Test stimuli were controlled byTDT system II hardware connected to a PC computer run-ning a MATLAB stimulus presentation interface. Each listenerwas instructed regarding the task using verbal and writteninstructions and completed a familiarization task prior to ex-perimental testing. The familiarization task consisted of sixnon-experimental sentences. The first two sentences were un-processed. The last four processed sentences were presentedin a random order of vowel-only and consonant-only condi-tions. The processed version was presented first followed bythe full unprocessed version of the same sentence. The lis-tener was requested to repeat aloud all of the words theyheard for both versions of the sentence in order to ensurethey understood the experimental task. All participantsmoved on to the experiment following these familiarizationsentences regardless of performance.

During the experimental testing, each participant wasrandomly assigned to one of two condition subsets, each ofwhich contained 6 of the 12 experimental conditions. Eachparticipant heard 14 sentences per condition �114–115words; �462 phonemes�. The six experimental conditionsincluded both processing strategies �V�VP and C+VP� forthree vowel proportions. Sentences were presented in twoblocks with different fixed orders of three quasi-randomlyassigned conditions each, resulting in participants listeningto a total of 84 sentences. Past research demonstrated littlebenefit of hearing the sentence presented twice, even whenpresented sequentially �Kewley-Port et al., 2007�. Here, thesame sentence was never presented in the same block toparticipants. The possibility of a sentence repetition effect isexplored in the results.

The task for each subject was to repeat each sentencealoud. Listeners were encouraged to guess, without regard towhether their responses made logical sentences. No feedbackwas provided. The experimenter scored responses on paperduring the session and digitally recorded them for later re-analysis by an independent scorer �Inter-rater agreement on10% of responses: 98.8%�. Only words repeated exactly cor-rect were scored �e.g., no missing suffixes�. For example, a
response of “pool” would be judged as incorrect if the target

sentence had used the plural form “pools.” In addition, wordswere allowed to be repeated back in any order to be countedas correct repetitions.

E. Results and discussion

1. Descriptive analysis

In the first analysis, possible effects of learning wereinvestigated because each sentence was presented twice tolisteners, each time under a different condition. Five sen-tences that occurred within the first ten sentences presentedin block 1 and within the last ten sentences in block 2 wereselected for analysis. For each sentence, the average percentcorrect across each of the participants was calculated. Thisaverage score was transformed to a z-score for its particularsegmental condition. No significant difference in perfor-mance was found, p=0.98; normalized scores correspondedto a performance increase of 1.6% when the sentence waspresented a second time. Furthermore, when consideringz-scores for all of the sentences, there was no significantdifference between blocks. These results indicate that listen-ers’ performance remained consistent across trials and wereunaffected by hearing a sentence presented a second time ina different condition. Thus, reported results appear reliabledue to listeners having stable, repeatable performance andwere not affected by procedural learning across the twoblocks.

Consonants and vowels were manipulated in differentdirections to investigate the contributions of transitional in-formation. This specific experimental design required a se-ries of six paired samples t-tests to be conducted using Bon-ferroni corrections for multiple comparisons to compareperformance across each of the vowel proportions tested.Overall, results indicated that the vowel-only �V only� con-dition �M =64.7% words correct� had a significant �t�9�=−9.7, p�0.001�, two-to-one advantage over the consonant-only �C only� condition �M =30.5% words correct�, replicat-ing Cole et al. �1996� and Kewley-Port et al. �2007�. Theother five t-tests between consonant �C+VP� and vowel �V�VP� conditions for each VP pair were significantly differ-ent at VP=5% �t�9�=−5.6, p�0.001�, VP=10% �t�9�=−5.7, p�0.001�, VP=30% �t�9�=4.2, p�0.008�, and VP=45% �t�9�=14.9, p�0.001�. However, no significant dif-ference was found when VP=15% �t�9�=−1.5, p=0.18�, in-dicating that consonant �C+15� and vowel �V�15� condi-tions approached similar intelligibility. Figure 3 shows thedata points and trends for the consonant and vowel condi-tions drawn with equal fit �R2=0.98� using least squares re-gression. A linear function approximated the C+VP condi-tions, while a cubic function was required to provide thesame fit for the V�VP conditions. Thus, C+VP conditionsincreased linearly with the addition of proportional vowelinformation. In contrast, V�VP conditions, with a nearlyconstant function around 55% correct, appeared to remainrobust against proportional decreases in information until30% of the vowel was removed.

These results suggest that increasing transitional infor-mation at C-V boundaries linearly predicts perceptual accu-
racy when consonants are presented. However, for vowels,

this transitional information at the C-V boundary appears toprovide little added perceptual benefit, suggesting that theC-V boundary transitions provide information redundantwith the contributions from vowel centers, when measuringsentence intelligibility. Similar results were found withCVCs by Strange et al. �1983� who suggested that informa-tion for the vowel is distributed across the changing acous-tics of the syllable. In our study, the resilience of vowelcenters remained until 30% of the transitional vowel infor-mation at the boundaries was deleted. This corresponded towhen intelligibility was on average 56% correct for vowelsand 69% for consonants. Note that this 30% proportionalC-V shift maximized performance for both consonants andvowels.

2. Sentence level analysis

Figure 4 shows on average, how much of the total sen-tence was preserved across the 12 conditions. The total sen-tence duration on average was 2.48 s. Sentence-level mea-surements for the original TIMIT boundaries demonstratedthat consonants comprised 55% of the total sentence dura-tion, while vowels only comprised 45%. While individualconsonants have a shorter duration than vowels, more con-sonants �M =22� occurred per sentence than vowels �M=11�. On average, 74% of consonants occurred in consonantstrings, while only 12% of vowels occurred in vowel strings.When 30% of the vowel is removed in the V�30 conditionperformance is at 56% correct. Performance in the consonantconditions does not surpass this 56% performance level untila 30% shift in the boundary for the �C+30�, where perfor-

FIG. 3. Results for the 12 experimental conditions in Experiment 1 plottedaccording to the vowel proportion by which the C-V boundary was shifted.The original TIMIT C-V boundary is at 0% VP, marked by the dashedvertical line. Error bars display the standard error of the mean.

FIG. 4. The average amount of the sentence preserved across the 12 condi-tions with the rest of the sentence replaced by noise. The amount presentedwas distributed over the entire sentence according to the specific noise re-
placement condition. The length of the average sentence was 2.48 s.

mance reaches 69% correct. Note that for this shift, 2 /3 ofthe sentence is presented in this consonant condition, com-pared to only 1 /3 presented in the corresponding vowel con-dition just mentioned. Clearly, the cumulative amount of thesentence presented is not an underlying factor that providesthe vowel conditions an advantage. When considering per-ceptual performance, it is notable that vowels contributetwice as much to sentence intelligibility when the originalC-V boundary is used despite preserving less of the overallsentence duration �only 45% for vowels�. When consideringduration of the presented speech signal, consonants mustpresent twice as much of the total sentence duration to sur-pass performance of the vowel.

3. Segmental level analysis

Another way to analyze the results is by each individualsegment, regardless of whether it occurred as part of a stringof segments. This analysis corresponds to examining the ef-fects relative to the average segment duration rather thanaverage duration of each glimpse of the speech waveform.Figure 5 displays the percent of words correct relative to theduration of signal information present in the average speechsegment calculated separately for each condition. For the Vonly and C only conditions, this corresponds to the averagevowel or consonant duration based on the original TIMITboundaries. On average, the C only condition provided ashorter segment duration �M =62 ms� than the V only condi-tion �M =97 ms�. However, for the rest of the conditions, asegment is defined according to the noise replacement para-digm used in this study. For example, a segment in the C+45 condition is the original consonant plus 45% of thevowel duration, with an average duration per segment of105 ms.

When the amount of speech information was varied by aproportion of the vowel, intelligibility as a function of seg-ment duration for the consonant and vowel conditionscrossed. Specifically, at the segmental level, equal perceptualcontributions to sentence intelligibility occurred when thesignal duration on average for consonants and vowels was

FIG. 5. The average duration of an individual segment in each condition�see text for details�.

equal to 83 ms. Listener performance was matched at this


point for a word identification score of 55% correct. Whilethis absolute value of 83 ms is likely dependent on factorssuch as speech rate, this point of perceptual equivalence oc-curred nearest to our 15% VP shift in both segmental condi-tions. As previously discussed, there was no statistical differ-ence in the C+15 and V�15 conditions. Thus, while a 30%VP shift of the C-V boundary maximized both consonant andvowel perceptual contributions to sentence intelligibility, a15% VP shift most successfully �out of the conditions pre-sented� equalized such contributions.

4. Glimpse window analysis

The durations of glimpse windows locked to consonantand vowel segments were measured across all 42 sentencesto determine if these durations composed significantly differ-ent distributions for the two segmental types �see Fig. 6�,thereby providing an opportunity for different perceptualcontributions related to glimpse duration. While there was alarge overlap in the distributions for consonant �M =113 ms,SD=67.6 ms� and vowel �M =102 ms, SD=56.6 ms�glimpse windows, there was a significant difference betweenthe distributions �t�460�=2.642, p�0.01�. Therefore, whileresults for Experiment 1 clearly demonstrated that glimpsewindows locked to vowel information were essential for sen-tence intelligibility, it may be that the results were largelyrelated to glimpse window differences rather than to segmen-tal differences found in the acoustic content of the windows.Experiment 2 was designed to tease apart the independentcontributions of segmental acoustics and glimpse windows.

III. EXPERIMENT 2: GLIMPSE CONTRIBUTIONS

The better performance with vowels in Experiment 1may be due to cues inherent to segment glimpse windowdurations rather than specific segmental acoustic cues pro-vided in the glimpse �as vowel windows overall were signifi-cantly shorter than consonant windows�. This may have oc-curred because performance changes as a function ofcombined changes in glimpse rate and duration �Miller andLicklider, 1950�. Specifically, glimpses of different dura-tions, as noted between the different durations of consonantsand vowels, may affect performance in addition to the actual
acoustics presented. To investigate this issue, a second ex-

periment was designed to control for specific segmental con-tributions by ensuring that acoustics from both consonantsand vowels were equally likely to occur within a singleglimpse. Experiment 2 examined: �1� the differential contri-butions to sentence intelligibility of glimpse window dura-tions measured from consonants and vowels and �2� the ef-fect of randomly placed glimpse windows that areasynchronous with segmental information. In addition, theseglimpse windows were varied according to a subset of sixdurational proportions selected from among the 12 condi-tions of Experiment 1.

A. Listeners

Ten young �age 19–25, M =22� normal-hearing listeners�three male, seven female� were paid to participate in thestudy. These listeners were different from those who partici-pated in Experiment 1. All participants were native speakersof American English and had pure-tone thresholds bilaterallyat no greater than 20 dB HL at octave intervals from250 to 8000 Hz �ANSI, 1996�.

B. Stimuli

The same sentences were used as in Experiment 1 anddiffered only in processing strategy. The same glimpse win-dows were used within each sentence, preserving both thenumber and duration of glimpses as calculated from theTIMIT database. The order of these windows was then quasi-randomized such that glimpse boundaries no longer corre-sponded with segmental boundaries. The quasi-randomization followed three rules. First, the windowsalternated between windows measured from consonant unitsand those measured from vowel units, such that no two win-dows measured from the same segment type occurred in ad-jacent intervals; thereby avoiding two glimpse windowscombining into a larger one. Second, for the majority of sen-tences there was an unequal number of windows between thetwo segmental types. In order to ensure that glimpse bound-aries were maximally offset at the beginning of the sentence,either the shortest or longest window measured from the seg-mental type with the most windows in that sentence was

FIG. 6. Glimpse window distributionsmeasured from all segments which oc-curred across all 42 sentences for �a�vowel glimpse windows and �b� con-sonant glimpse windows.

randomly selected to be presented first. When there was an


equal number of consonant and vowel windows, this rule didnot apply and the first segment window was selected at ran-dom from the opposite segmental type presented first in thesentence �e.g., if the sentence began with a consonant, avowel window was selected at random�. Third, when therewere pause windows �i.e., silent intervals� present in the sen-tence, these windows were presented in the randomized or-der closest to their original location in the sentence in anattempt to preserve sentence prosody. Without this latter rule,the segmental window that fell within a silent period in thesentence would convey no acoustic information. For Experi-ment 2, it was important to preserve the same amount ofspeech acoustics as presented in Experiment 1; otherwisepoor performance could be related to reduced amounts ofspeech acoustics overall. This possibility was explicitlyavoided by this third rule.

As with Experiment 1, two different processing strate-gies were used: one that used consonant glimpse windowdurations and one that used vowel glimpse window dura-tions. The durations of these windows were manipulated bythree of the previous six proportions of the vowel �0%, 15%,30%� to yield a total of six conditions. This subset of condi-tions was selected to explore intelligibility at glimpse dura-tions that provided relatively stable performance for vowelconditions. The glimpse windows measured in these condi-tions corresponded directly to those used in Experiment 1under the identical vowel proportion condition. The only dif-ference here was that glimpse windows were no longer time-locked to specific segmental boundaries. Calibration for thisexperiment used procedures identical to those in Experiment1.

C. Procedure

Test procedures were identical to those of Experiment 1.Each participant completed a familiarization task identical toExperiment 1, but with randomized glimpse conditions. Dur-ing the experimental testing, participants again completedtwo different blocks of 42 sentences. Each block containedthree different conditions in a fixed random order. Sentenceswere presented once in each block under a different condi-tion with a different randomization of glimpse windows. Par-ticipants again repeated aloud each sentence without feed-back. The experimenter scored responses on paper during thesession and digitally recorded them for later reanalysis by anindependent scorer �Inter-rater agreement on 10% of re-sponses: 96.5%�.

D. Results and discussion

1. Glimpse window analysis

As with Experiment 1, the effect of practice resultingfrom procedural learning and sentence repetition was exam-ined to determine the reliability of participant responses.Analysis of five sentences that were presented within the firstten sentences of block 1 and the last ten sentences of block 2were compared. No significant difference in z-scores wasobserved �p=0.71�. The z-scores for these two presentations
corresponded to a 3.1% increase in scores for the five sen-

tences at the end of the second block. Thus, results for thissecond experiment were stable under this processing strat-egy.

Figure 7 displays the results of Experiment 2 accordingto how the distributions of glimpse duration were shifted byvowel proportion. All condition labels for Experiment 2 areappended with an initial “g” to indicate that they only pre-serve the duration of speech within a glimpse and do notcontain specific segmental information. Results were ana-lyzed by a series of Bonferroni-corrected t-tests conductedbetween glimpses measured from consonants and vowels ateach of the three vowel proportions tested. The analysis in-dicated a significant difference between glimpse windows ofgC+VP and gV�VP conditions at 15% �t�9�=9.4, p�0.001� and 30% �t�9�=21.8, p�0.001� VP pairs, but not atthe original C-V boundary �0% VP�, �t�9�=1.7, p=1.88�.Thus, the two-to-one advantage obtained with segmentalwindows was not observed between randomly presentedwindows measured from consonants versus those from vow-els. These results suggest that the large difference betweenglimpses locked to segmental information in Experiment 1was a result of perceptual cues contained in the segmentalacoustics, not characteristics inherent in the distributions ofglimpse windows. Thus, the initial conclusions of Experi-ment 1 are validated: the perceptual cues contained withinvowel segments do contribute more to sentence intelligibilitythan those within consonant segments.

Figure 8 displays the six random glimpse conditions ofExperiment 2 �black� plotted according to the cumulativeduration of the sentence presented. As a reference, resultsfrom the corresponding conditions of Experiment 1 �gray�are also plotted. A linear trend �solid line� was fit to the sixglimpse conditions tested �R2=0.93�. In fact, when glimpseswere not locked to segmental information, the percent of thesentence presented closely predicted performance, particu-larly for the four longer values. That is, given the percent ofthe sentence presented at 45%, 55%, 62%, and 69%, thepercent of the words identified was predicted by those values

FIG. 7. Results for the six experimental conditions in Experiment 2 plottedaccording to the vowel proportion by which the C-V boundary was shifted.Glimpse window durations based on the original TIMIT C-V boundary areat 0% VP, marked by the dashed vertical line. Glimpse windows measuredfrom the consonant �black diamonds� and vowel �gray squares� conditionsof Experiment 1 are coded similarly to previous figures but now with theprefix “g” to indicate random glimpses. Error bars display the standard errorof the mean.

within 1%–5%. While such close predictions did not hold for


the two shorter durations �gV�15 and gV�30�, overall, thepredictability of word accuracy from only the duration of thesentence presented was remarkably high.

2. Comparison of results with Experiment 1

An analysis was conducted to examine the difference inexperimental results across the two experiments. Bonferroni-adjusted t-tests between Experiments 1 and 2 for the sixpaired conditions �see Fig. 8� demonstrated significant differ-ences for four conditions tested, but not for the two longestglimpse window conditions, C+15 �t�9�=2.0, p=0.08� andC+30 �t�9�=1.3, p=0.24�. Of the four conditions that weresignificantly different, three conditions �V�30, t�9�=−12.8,p�0.001; V�15, t�9�=−8.5, p�0.001; and V only, t�9�=−7.8, p�0.001� yielded higher performance in Experiment 1with segmentally locked glimpses that contained only vowelinformation. Performance was poorer with randomly placedglimpses that only contained partial vowel information. Thebenefit of vowel segmental information is clearly seen in allthree comparisons �V only, V�15, and V�30�, but espe-cially at the shortest glimpse duration, V�30. At the V�30duration, providing glimpses of only vowel information inthe segmental condition of Experiment 1 yields a perceptualadvantage over the random glimpses of Experiment 2 by afactor of 4. The fourth significant condition, C only �t�9�=12.6, p�0.001�, is a special case where segmentallylocked glimpses in Experiment 1 contained no vowel infor-mation. For this condition, performance was significantlybetter with randomized glimpses, presumably because of theinclusion of partial vowel information. Therefore, our inter-pretation of the results for the four conditions is as follows:significant differences between segmental and randomglimpses from the two experiments can be explained by in-creased amounts of preserved vowel information. Consider-ing the similar performance of the two longest glimpse win-dows, these conditions �C+15 and C+30� contained themost segmental information distributed over both consonantsand vowels. For these last two conditions, random glimpsesapparently provided the same amount of perceptual informa-tion as is provided when larger amounts of vowel informa-tion were added to C only information in Experiment 1.Therefore, these glimpse windows are long enough for the

FIG. 8. The percent of words correct is displayed for the six conditions ofExperiment 2 �black� and the comparison conditions of Experiment 1 �gray�as a function of the cumulative duration of the sentence preserved over theentire sentence. The black regression line is only across the six randomglimpse conditions. Asterisks � *� indicate a significant difference betweenExperiments 1 and 2, p�0.001.

combined acoustics of both consonants and vowels to con-


tribute to the perception of locked or randomized glimpsewindows. In summary, the presence of more vowel informa-tion in glimpses yields higher perceptual performance. Vowelinformation is particularly important when less of the sen-tence is presented.

IV. GENERAL DISCUSSION

A. Segmental contributions to sentence intelligibility

The current study investigated the perceptual contribu-tions of vowels, consonants, and the transitions betweenthem, using noise replacement with locked or randomglimpses. Results demonstrated the importance of vowel in-formation to sentence intelligibility as described by Coleet al. �1996� and Kewley-Port et al. �2007�. Vowels as tradi-tionally defined carry important perceptual cues that are notfound in consonants. Furthermore, even truncated portions ofvowels contribute strongly to sentence intelligibility despiteproviding much less of the overall sentence duration thanconsonants. Perceptual cues are distributed across the entirevowel, such that transitions into the vowel provide little ad-ditional perceptual benefit above what is provided by thevowel center alone, when the vowel center contains 70% ofthe total vowel. However, these dynamic transitions do pro-vide additional perceptual information to consonants whichsignificantly improve sentence intelligibility. Such perceptualinformation in the “vowel” transitions is not found in tradi-tionally defined consonants, as indicated by the relativelylow consonant-only sentence scores. Our results correspondwell with the literature that has identified C-V transitions�Liberman et al., 1967� and vowel context �Cooper et al.,1952� as important cues for consonant recognition, and withevidence from silent-center vowels that demonstrated the im-portance of dynamic information for vowel recognition�Strange, 1987; Strange et al., 1983�.

Owren and Cardillo �2006�, using a similar replacementtechnique with silence rather than noise, found that conso-nants are more important to the intelligibility of isolatedwords. However, it may be that vowels gain greater percep-tual importance from an increased amount of contextual orstylistic factors present in sentential contexts. In sentences,speech acoustics are characterized by shorter individual seg-ments and greater coarticulation. Thus, individual consonantsegments may become less distinct, requiring more transi-tional dynamics for recognition. This blurring of segmentalboundaries results in words, originally intelligible in sen-tences, becoming largely unintelligible when excised �Craigand Kim, 1990�. Clearly, listeners take advantage of coar-ticulatory and prosodic cues in the perception of fluent sen-tences. In fact, some speech recognition methods explicitlydefine coarticulatory events between words in order toachieve higher recognition performance for continuous sen-tences �Giachin et al., 1991�.

The most sonorant segments in speech, namely the vow-els, are most capable of carrying prosodic information suchas pitch contours or stress. Prosody spans linguistic levels�Kent and Read, 1992�, providing an opportunity for mul-tiple perceptual and cognitive levels involved in sentence
recognition to interact. For example, prosody may provide

information about syntactic structure through cues such asphrase-final lengthening. This particular prosodic cue, tied tothe last stressed syllable in a major syntactic phrase, is per-ceptually available for listeners to use in recognizing thestructure of speech �Read and Schreiber, 1982�. Vowels maybe particularly useful in conveying these cues for syntacticstructure, as stress patterns temporally align with the onset ofvowel energy �Tajima and Port, 2003�. Indeed, Toro et al.�2008� demonstrated that participants were able to extractsyntactic cues from vowels, but not from consonants. Fur-thermore, they support a functional dissociation of conso-nants and vowels, with consonants cuing the lexicon andvowels cuing syntactic structure. It may be that the syntacticcues that vowels uniquely carry facilitate top-down processesand constrain lexical activation.

Therefore, there are many types of information thatvowels may carry to provide them with an advantage overconsonants during the perception of compromised sentencescontaining only partial information. These vowel-associatedcues encompass acoustic, coarticulatory cues important forconsonant recognition, but may also involve perceptual cuesthat provide higher-order mechanisms to predict information,even though many explicit acoustic cues for segment recog-nition may be missing. Future studies will have to tease apartpotential contributions from the complex acoustic, percep-tual, and cognitive events that occur in sentences, which weas listeners navigate with relative ease in every day conver-sational contexts.

B. Glimpse contributions to sentence intelligibility

Even though there was a significant difference betweenthe durational distributions of consonant and vowelglimpses, no differential performance was found for percep-tual intelligibility of sentences when using these glimpsewindows in a randomly ordered replacement paradigm. In-stead, the amount of the speech signal presented, distributedover the length of the entire sentence, predicted perceptualperformance. This result is consistent with Iyer et al. �2007�who used periodic glimpses to investigate the effect of noiseand speech maskers. Above an 8 Hz interruption rate, thetrade-off between the glimpse window duration and numberof glimpses does not have much of an effect, as long as thecumulative sentence duration glimpsed �i.e., the total dura-tion of preserved speech information� remains constant. Alsoconsistent with the current study, Li and Loizou �2007� foundthat the cumulative duration of speech glimpsed predicts per-formance rather than glimpse window duration. By compar-ing performance across several glimpse conditions that pre-served different frequency regions of speech, Li and Loizou�2007� also identified the frequency region of 0–3 kHz asthe most important frequency region for glimpsing. This fre-quency region corresponds to the F1 and F2 frequency re-gions most important for vowel perception. Therefore, thisevidence provides further support for the important contribu-tion of vowel information to sentence intelligibility.

Experiment 2 investigated the combined perceptual ef-fect of the distributional properties and random �i.e., aperi-
odic� placement of glimpse windows. In a pilot study, we

separately assessed the contributions of overall sentence du-ration presented while avoiding differences in the distribu-tions of consonants versus vowels. The pilot used periodicreplacement based on the duty cycle that was the average ofthe glimpse windows �215 ms�. For the consonant glimpsecondition, the on phase was 113 ms, the mean glimpse win-dow size for consonants, while the off phase was 102 ms, themean glimpse size for vowels. For the vowel glimpse condi-tion, the on and off phases were reversed �e.g., the on phasewas the average vowel window duration of 102 ms�. Theseon/off phases were manipulated by a proportion of the vowelas in the experiments presented here to provide a direct com-parison. The pilot for all the periodic replacement conditionsyielded very similar results to those obtained in Experiment2 with aperiodic replacement, further suggesting that glimpseperception is predicted by the cumulative duration of thesentence presented. These results are consistent with Millerand Licklider �1950� who found that perception of monosyl-labic words during random, aperiodic interruption was effec-tively the same as during periodic interruption.

Overall, there is a significant difference between the seg-mentally locked glimpses in Experiment 1 and the random,segmentally asynchronous glimpses used in Experiment 2.There is a significant advantage for glimpses that capturespecific spectro-temporal information contained in vowelsegments.

V. CONCLUSIONS

This study used a fundamental division in speech soundcategories, specifically between consonants and vowels, toinvestigate contributions to sentence intelligibility. Thepresent study replicated results from Cole et al. �1996� andKewley-Port et al. �2007� that demonstrated a two-to-oneadvantage for sentences containing only vowels compared tosentences that contained only consonants. Furthermore, thisstudy investigated how these contributions were mediated bytransitional information into the vowel. This transitional in-formation appears to be redundant with information at thevowel center, suggesting a similar result to Lee and Kewley-Port �2009� who found no difference in transitional versusquasi-steady-state information. However, increased transi-tional information did provide a linear benefit relative toconsonant-only sentences, demonstrating that critical percep-tual information is present in the transition. Finally, the cur-rent study demonstrated that glimpse window distributionsdid not differentiate performance between consonant-onlyand vowel-only sentences. Instead, the cumulative amount ofthe sentence present and the amount of vowel informationpreserved served as predictors for listener performance. Re-sults confirm the importance of vowels to sentence intelligi-bility, as well as how dynamic transitional information at theC-V boundary significantly improves the perceptual contri-butions of consonants. Simply put, vowel information is es-sential for sentence intelligibility, particularly when less of
the overall sentence duration is presented.

ACKNOWLEDGMENTS

This work was supported by NIH-NIDCD TrainingGrant No. T32-DC00012 and R01-02229 to Indiana Univer-sity.

ANSI �1969�. “American National Standard Methods for Calculation of theArticulation Index,” ANSI S3.5-1969, American National Standards Insti-tute, New York.

ANSI �1996�. “Specifications for audiometers,” ANSI S3.6-1996, AmericanNational Standards Institute, New York.

Bashford, J. A., Jr., and Warren, R. M. �1987�. “Multiple phonemic restora-tions follow the rules for auditory induction,” Percept. Psychophys. 42,114–121.

Bashford, J. A., Jr., Reiner, K. R., and Warren, R. M. �1992�. “Increasing theintelligibility of speech through multiple phonemic restorations,” Percept.Psychophys. 51, 211–217.

Cole, R., Yan, Y., Mak, B., Fanty, M., and Bailey, T. �1996�. “The contri-bution of consonants versus vowels to word recognition in fluent speech,”in Proceedings of the ICASSP’96, pp. 853–856.

Cooke, M. �2003�. “Glimpsing speech,” J. Phonetics 31, 579–584.Cooke, M. �2006�. “A glimpsing model of speech perception in noise,” J.

Acoust. Soc. Am. 119, 1562–1573.Cooper, F., Delattre, P., Liberman, A., Borst, J., and Gerstman, L. �1952�.

“Some experiments on the perception of synthetic speech sounds,” J.Acoust. Soc. Am. 24, 597–606.

Craig, C. H., and Kim, B. W. �1990�. “Effects of time gating and wordlength on isolated word recognition performance,” J. Speech Hear. Res.33, 808–815.

Freyman, R. L., and Nerbonne, G. P. �1989�. “The importance of consonant-vowel intensity ratio in the intelligibility of voiceless consonants,” J.Speech Hear. Res. 32, 524–535.

Freyman, R. L., Nerbonne, G. P., and Cote, H. A. �1991�. “Effect ofconsonant-vowel ratio modification on amplitude envelope cues for con-sonant recognition,” J. Speech Hear. Res. 34, 415–426.

Garofolo, J., Lamel, L., Fisher, W., Fiscus, J., Pallett, D., and Dahlgren, N.�1990�. “DARPA TIMIT acoustic-phonetic continuous speech corpusCDROM,” National Institute of Standards and Technology, NTIS OrderNo. PB91-505065.

Giachin, E. P., Rosenberg, A. E., and Lee, C.-H. �1991�. “Word juncturemodeling using phonological rules for HMM-based continuous speechrecognition,” Comput. Speech Lang. 5, 155–168.

Iyer, N., Brungart, D., and Simpson, B. �2007�. “Effects of periodic maskerinterruption on the intelligibility of interrupted speech,” J. Acoust. Soc.Am. 122, 1693–1701.

Kent, R., and Minifie, F. �1977�. “Coarticulation in recent speech productionmodels,” J. Phonetics 5, 115–133.

Kent, R. D., and Read, C. �1992�. The Acoustic Analysis of Speech �Singular,San Diego�.

Kewley-Port, D., Burkle, T. Z., and Lee, J. H. �2007�. “Contribution ofconsonant versus vowel information to sentence intelligibility for young


normal-hearing and elderly hearing-impaired listeners,” J. Acoust. Soc.Am. 122, 2365–2375.

Ladefoged, P. �2001�. Vowels and Consonants: An Introduction to theSounds of Languages �Blackwell, Oxford�.

Lee, J. H., and Kewley-Port, D. �2009�. “Intelligibility of interrupted sen-tences at subsegmental levels in young normal-hearing and elderlyhearing-impaired listeners,” J. Acoust. Soc. Am. 125, 1153–1163.

Leung, H. C., and Zue, V. W. �1984�. “A procedure for automatic alignmentof phonetic transcriptions with continuous speech,” in Proceedings of IC-ASSP’84, pp. 2.7.1–2.7.4.

Li, N., and Loizou, P. �2007�. “Factors influencing glimpsing of speech innoise,” J. Acoust. Soc. Am. 122, 1165–1172.

Liberman, A. M., Cooper, F. S., Shankweiler, D. P., and Studdert-Kennedy,M. �1967�. “Perception of the speech code,” Psychol. Rev. 74, 431–461.

Miller, J. L. �1994�. “On the internal structure of phonetic categories: Aprogress report,” Cognition 50, 271–285.

Miller, G., and Licklider, J. �1950�. “The intelligibility of interruptedspeech,” J. Acoust. Soc. Am. 22, 167–173.

Owens, E., Talbot, C. B., and Schubert, E. D. �1968�. “Vowel discriminationof hearing-impaired listeners,” J. Speech Hear. Res. 11, 648–655.

Owren, M. J., and Cardillo, G. C. �2006�. “The relative roles of vowels andconsonants in discriminating talker identity versus word meaning,” J.Acoust. Soc. Am. 119, 1727–1739.

Ramus, F., Nespor, M., and Mehler, J. �1999�. “Correlates of linguisticrhythm in the speech signal,” Cognition 75, AD3–AD30.

Read, C. and Schreiber, P. A. �1982�. “Why short subjects are harder to findthan long ones,” in Language Acquisition: The State of the Art, edited byE. Wanner and L. Gleitman �Cambridge University Press, Cambridge�.

Rhebergen, K. S., and Versfeld, N. J. �2005�. “A speech intelligibility index-based approach to predict the speech reception threshold for sentences influctuating noise for normal-hearing listeners,” J. Acoust. Soc. Am. 117,2181–2192.

Sammeth, C. A., Dorman, M. F., and Stearns, C. J. �1999�. “The role ofconsonant-vowel intensity ratio in the recognition of voiceless stop con-sonants by listeners with hearing impairment,” J. Speech Lang. Hear. Res.42, 42–55.

Stevens, K. N. �2002�. “Toward a model for lexical access based on acousticlandmarks and distinctive features,” J. Acoust. Soc. Am. 111, 1872–1891.

Strange, W. �1987�. “Information for vowels in formant transitions,” J.Acoust. Soc. Am. 26, 550–557.

Strange, W., Jenkins, J. J., and Johnson, T. L. �1983�. “Dynamic specifica-tion of coarticulated vowels,” J. Acoust. Soc. Am. 74, 695–705.

Tajima, K., and Port, R. �2003�. “Speech rhythm in English and Japanese,”Papers in Laboratory Phonology VI, edited by J. Local, R. Ogden, and R.Temple �Cambridge University Press, Cambridge�, pp. 317–334.

Toro, J. M., Nespor, M., Mehler, J., and Bonatti, L. L. �2008�. “Findingwords and rules in a speech stream: Functional differences between vow-els and consonants,” Psychol. Sci. 19, 137–144.

Zue, V. W., and Seneff, S. �1988�. “Transcription and alignment of theTIMIT database,” in Proceedings of the Second Meeting on AdvancedMan-Machine Interface Through Spoken Language, pp. 11.1–11.10.


Perceptual contributions of the consonant-vowel boundary to sentence intelligibility

Documents

Transcript of Perceptual contributions of the consonant-vowel boundary to sentence intelligibility