Stop Bursts Research Articles

Developing children's ASR system under low-resource conditions using end-to-end architecture

The work presented in this paper aims at enhancing the performance of end-to-end (E2E) speech recognition task for children's speech under low resource conditions. For majority of the languages, there is hardly any speech data from child speakers. Furthermore, even the available children's speech corpora are limited in terms of the number of hours of data. On the other hand, large amounts of adults' speech data are freely available for research as well as commercial purposes. As a consequence, developing an effective E2E automatic speech recognition (ASR) system for children becomes a very challenging task. One may develop an ASR system using adults' speech and then use it to transcribe children's data, but this leads to very poor recognition rates due to the stark differences in the acoustic attributes of adults' and children's speech. In order to overcome these hurdles and to develop a robust children's ASR system employing E2E architecture, we have resorted to several out-of-domain and in-domain data augmentation techniques. For out-of-domain data augmentation, we have explicitly modified adults' speech to render it acoustically similar to that of children's speech before pooling into training. On the other hand, in the case of in-domain data augmentation, we have slightly modified the pitch and duration of children's speech in order to create more data capturing greater diversity. Data augmentation approaches helps in mitigating the ill-effects resulting from the scarcity of data from child domain to a certain extent. This, in turn, reduces the error rates by a large margin. In addition to data augmentation, we have also studied the efficacy of Gamma-tone frequency cepstral coefficients (GFCC) and frequency domain linear prediction (FDLP) technique along with the most commonly used Mel-frequency cepstral coefficients (MFCC) for front-end speech parameterization. Both MFCC as well as GFCC capture and model the spectral envelope of speech. On the other hand, application of linear prediction on the frequency domain representation of speech signal helps to effectively capture the temporal envelope during front-end feature extraction. Employing FDLP features that model the temporal envelope provides important cues for the perception and understanding of stop bursts and, at times, complete phonemes. This motivated us to perform a comparative experimental study of the effectiveness of the three aforementioned front-end acoustic features. In our experimental explorations, the use of proposed data augmentation in combination of FDLP features has shown a relative improvement in character error rate by 67.6% over the baseline system. The combination of data augmentation with MFCC or GFCC features is observed to result in lower recognition performances.

Stop voicing contrast in Lio

Typologically, two-way or three-way voicing contrasts in stops can be reliably defined along the VOT continuum, that is, by measuring the gap between the start of voicing and the stop burst, which can be negative or positive. In the literature, three main categories are therefore derived: prevoiced, unaspirated, and aspirated stops (Keating, 1984; Lisker and Abramson, 1964; and Cho et al., 2019). For Lio, an understudied Austronesian language of Central Flores (Eastern Indonesia), it has been claimed that the stop voicing contrast is realized with prevoicing for voiced stops and aspiration for voiceless stops (Elias, 2018). However, it is unclear how long the aspiration is, and whether Lio voiceless stops can be characterized as unaspirated or aspirated; furthermore, this claim was not based on an acoustic analysis, and relied heavily on one speaker. This study offers a descriptive analysis on voicing contrasts in Lio stops. The acoustic analysis reveals that the contrast is one of ‘true voicing’, with the closure phase of voiced stops presenting voicing (prevoiced) and short-lag VOT voiceless stops (unaspirated). This study will offer a valuable contribution to voicing contrasts typology, especially compared to Cho et al. (2019), and serves as a building block for future Lio research.

Comparison of VOTs in Mandarin–English bilingual children and corresponding monolingual children and adults

Word-initial stops in Mandarin and English show a distinctive phonological categorization but a similar phonetic realization along the VOT (Voice Onset Time) continuum. Previous research reported that native Mandarin adults produce measurably longer long-lag VOTs than native English adults. The present study examined whether and how the difference between Mandarin and English VOTs is manifested in monolingual children and Mandarin–English bilingual children. The participants included 15 five- to six-year-old sequential bilingual children, 24 corresponding monolingual children (15 Mandarin, 9 English), and 22 monolingual adults (12 Mandarin, 10 English). The bilingual children were divided into two groups (Bi-low and Bi-high) based on the amount of experience in English. Each participant was recorded producing 18 Mandarin words and/or 18 English words containing six stops in each language. The VOT values were measured from the beginning of stop burst to the onset of the voicing. The results showed that the language difference in VOT in the monolingual children was manifested in a pattern similar to the monolingual adults. However, Mandarin and English VOTs showed less separable distributions in the two groups of bilingual children. Further analysis suggested that both groups of bilingual children tended to separate Mandarin and English short-lag VOTs but only the Bi-low children showed different long-lag VOTs between the two languages. These results suggested that due to the bilingual effects and L1–L2 (first language – second language) interactions, even though the bilingual children tried to separate the two VOT systems, they implemented the separation in a different manner than the monolingual speakers.

Acoustic and ultrasound investigation of word-initial gemination in Moroccan Arabic

Moroccan Arabic uses geminate/singleton contrasts in medial position but it is controversial whether it maintains them utterance-initially. To address this issue, we made simultaneous ultrasound and acoustic recordings of five native speakers producing target words containing /t/-/tt/ and /d/-/dd/ contrasts utterance-initially and -medially, 10 times each. The ultrasound data were analysed via a novel method of using pixel-derived principal components and linear discrimination to generate a time-varying articulatory “closure” signal directly from the ultrasound images, without the need for manual tracing. This articulatory signal was subsequently used to measure closure duration by determining gestural landmarks from its velocity function. The results provide clear articulatory evidence that speakers produce utterance-initial singletons versus geminates with significantly different closure durations, although the contrast provides no acoustic evidence of closure duration per se. Rather, our results reveal that speakers realize geminate/singleton contrasts via acoustic dimensions not related to consonant duration: The vowel is significantly longer following geminates than singletons, and the stop bursts have greater amplitude for geminates than singletons. Together these findings provide evidence that the gemination contrast is maintained in initial position in Moroccan Arabic, and challenge traditional assumptions that consonant length contrasts are primarily carried by acoustic closure duration differences.

Stress Effects on Stop Bursts in Five Languages

This study examines the effects of stress on the stop burst in five languages differing in number of places of articulation, as reflected in burst duration, spectral centre of gravity, and ­spectral standard deviation. The languages studied are English (three places of articulation /p t k/), the Indonesian language Makasar (four places /p t c k/), and the Central Australian languages ­Pitjantjatjara, Warlpiri (both five places /p t ʈ c k/), and Arrernte (six places /p t t ʈ c k/). We find that languages differ in how they manifest stress on the consonant, with Makasar not ­showing any effect of stress at all, and Warlpiri showing an effect on burst duration, but not on the ­spectral measures. For the other languages, the velar /k/ has a “darker” quality (i.e., lower spectral centre of gravity), and/or a less diffuse spectrum (i.e., lower standard deviation) under stress; while the alveolar /t/ has a “lighter” quality under stress. In addition, the dental /t/ has a more diffuse spectrum under stress. We suggest that this involves enhancement of the features [grave] and [diffuse] under stress, with velars being [+grave] and [–diffuse], alveolars being [–grave], and dentals being [+diffuse]. We discuss the various possible spectral effects of enhancement of these features. Finally, in the languages with five or six places of articulation, the stop burst is longer only for the palatal /c/ and the velar /k/, which have intrinsically long burst durations, and not for the anterior coronals /t t ʈ/, which have intrinsically short burst durations. We suggest that in these systems, [burst duration] is a feature that separates these two groups of consonants.

The effect of jaw position on perceptual and acoustic characteristics of speech

Bite blocks are used to stabilize the jaw and to isolate tongue and lip movements from that of the mandible during speech and nonspeech activities. Ten normally speaking young adults produced sentences with an unconstrained jaw and with unilateral placement of 2-mm and 5-mm bite blocks. Six listeners rated sentences spoken without either bite block as the most natural sounding. Spectral characteristics of /s/, /J/ and /t/ (sibilant frication and stop bursts) differed significantly with than without bite blocks, such that mean spectral energy decreased, and variation and skew of spectral energy increased. Spectral kurtosis did not change for the group, but 2 participants exhibited highly kurtotic /s/ spectra without a bite block that normalized with bite blocks. The second formant frequency for the high vowel /i/ was lower with bite blocks; there was no systematic difference in F2 slope for diphthongs. Segmental and suprasegmental timing of speech articulation was not affected significantly by these small bite blocks. This study provides support for using small bite blocks to isolate the tongue from the jaw without large effects on speech, but cautions that speech is likely to sound less natural than when produced with an unconstrained jaw.

The bursts of stops can convey dialectal information.

This study investigates the effects of the dialect of the speaker on the spectral properties of stop bursts. Forty-five female speakers-20 Standard Modern Greek and 25 Cypriot Greek speakers-participated in this study. The spectral properties of stop bursts were calculated from the burst spectra and analyzed using spectral moments. The findings show that besides linguistic information, i.e., the place of articulation and the stress, the speech signals of bursts can encode social information, i.e., the dialects. A classification model using decision trees showed that skewness and standard deviation have a major contribution for the classification of bursts across dialects.

Clicks, stop bursts, vocoids and the timing of articulatory gestures in Kinyarwanda

This paper shows that differences in timing and coordination of articulatory gestures in Kinyarwanda’s complex consonants trigger the emergence of epiphenomenal clicks. Acoustic and aerodynamic data show that click bursts of weak intensity appear in sequences of front (bilabial or alveolar) and velar nasals. The possible consequences of this phenomenon for sound change are briefly discussed. The emergence of short vocoids due a different timing of articulatory gestures allows discussing the status of syllabic constituents in complex nasal consonants. Keyword: Kinyarwanda, clicks, phonetics, vocoids, articulatory gestures

Stop bursts in Pitjantjatjara

Pitjantjatjara is an Australian language with five stop places of articulation /ptʈck/ in three vowel contexts /aiu/. We present word-medial stop burst data from nine speakers, examining duration, formant, spectral moment and spectral tilt measures. Our particular focus is on the apical contrast (alveolar /t/ vs. retroflex /ʈ/) and on the alveo-palatal /c/ vs. velar /k/ contrast. We observe differences between the palatal and the velar depending on vowel context, and we discuss the possible aerodynamic and acoustic sources for these differences. By contrast, we find that differences between the alveolar and the retroflex are minimal in all three vowel contexts. Unexpectedly, in the context of /i/, various spectral measures suggest that the articulatory release for the retroflex /ʈ/ is in fact more anterior than the release for the alveolar /t/ – we discuss this result in terms of possible articulatory overshoot of the target for /ʈ/ before /i/, and suggest that this result provides additional explanation for the cross-linguistic rarity of retroflexes in an /i/ vowel context.

Listener consistency in identifying speech mixed with particular “bubble” noise instances

Previous work has shown that the intelligibility of mixtures of the same exact speech token with different instances of “bubble” noise is highly dependent on the configuration of the random time-frequency glimpses it provides. In the current study, the consistency of these judgments was measured for such mixtures involving six consonants, /t/, /d/, /f/, /v/, /t ʃ/, /d ʒ/, in an /a/-consonant-/a/ context. Intra-subject consistency on ten repetitions each of 60 mixtures was found to be high, as was inter-subject consistency between five normal-hearing subjects on 1200 mixtures. In addition, the noise level at a small subset of time-frequency points was significantly correlated with the overall intelligibility of the mixtures of a given token, suggesting that these regions contribute more strongly to correct identification than others. The current study finds that these regions are quite consistent across the five subjects. For example, for the token /ada/, these points surround the stop burst and the reso...

Lexical stress and stop bursts in Pitjantjatjara: Feature enhancement of neutralized apicals and the coronal/velar contrast

This study investigates duration and spectral properties of the stop burst for /p t ʈc k/, the five oral stops of Pitjantjatjara, a language of Central Australia. We focus on stressed vs. unstressed syllables, with stressed syllables also being word-initial in Pitjantjatjara. Notably, the contrast between the apico-alveolar /t/ and the medial apico-postalveolar (retroflex) /ʈ/ is neutralized in initial position, which is the stressed position. We label this neutralized apical /T/, and show that the stop burst for this prosodically prominent apical has an even higher spectral center of gravity than either of its non-neutralized counterparts. By contrast, the velar /k/ shows a lower spectral center of gravity in stressed position. We suggest that this reflects an enhancement of the feature [grave], with the apical being “lighter” and the velar being “darker” under stress (the bilabial and the palatal show relatively few spectral effects, due to acoustic and coarticulatory considerations respectively). However, the stressed apical /T/ does not have a longer stop burst duration in initial position, a result that is contrary to the other stop bursts examined here, and which goes against expected patterns of prosodic strengthening. We suggest that this is because a phonetic feature, short [burst duration], is an important cue to apical consonant identity in a system with many places of articulation.

Stabilizing the production of nonnative consonant clusters with acoustic variability

Previous research on the perception, recognition, and learning of sounds and words has identified diverse effects of phonetic variation. The present study examined how variation affects cross-language production of consonant clusters. American English speakers shadowed words beginning with nonnative clusters in low- and high-variability conditions. Shadowing responses in the low-variability condition were quite sensitive to fine-grained phonetic properties that were manipulated across the stimuli. Notably, longer stop bursts led to increased rates of epenthesis, lower burst amplitudes resulted in more feature change and deletion, and intense periods of voicing at cluster onset elicited prothetic responses. Sensitivity to the acoustic manipulations was substantially attenuated in the high-variability condition, which combined stimuli from the first condition with baseline productions of the same items from two additional talkers. Detailed analyses of the response patterns indicate that more stable production targets in the high-variability condition resulted from integration, or blending, of the multiple talker stimuli. Implications of these findings for language-specific speech processing and the role of phonetic variability in second language acquisition are discussed.

Usability of SpeechMark® landmark analysis system for teaching speech acoustics

Learning about the intersection of articulation and acoustics, and particularly acoustic measurement techniques, is challenging for students in Linguistics, Psychology and Communication Sciences and Disorders curricula. There is a steep learning curve before students can apply the material to an interesting research question; for those in more applied programs such as Communication Disorders or ESL, there is an additional challenge in envisioning how the knowledge can be applied in changing behavior. The availability of software tools such as Wavesurfer, Praat, Audacity, TF32, the University College of London software suite, among others, has made it possible for instructors to design laboratory experiences in visualization, manipulation, and measurement of speech acoustics. Many students have found them complex for their first exposure to taking scientific measurements. The SpeechMark® acoustic landmark analysis system has been developed to automate the detection of specific acoustic events important for speech, such as voicing offset and onset, stop bursts, fricative noise, and vowel midpoints, and to provide automated formant frequency measurement used for vowel space analysis. This paper describes a qualitative multiple case study in which seven teachers of speech acoustics were interviewed to explore whether such pre-analysis of the acoustic signal could be useful for teaching.

Cue-weighting in the perception of intervocalic stop voicing in European Portuguese.

This paper describes the perception of intervocalic stop voicing in European Portuguese (EP) stimuli without a stop burst, when varying three acoustic cues: Vowel duration, stop duration, and voicing maintenance during stop closure. Perceptual stimuli were generated using biomechanical modeling. First, a discrimination experiment was conducted to determine the listeners' perceptual sensitivity to the voicing maintenance cue. Second, an identification experiment was conducted to examine the effect and interaction of vowel duration, stop duration, and voicing maintenance during stop closure on the voiced/voiceless identification responses of EP listeners. The results of the discrimination test show that voicing maintenance differences have a significant effect as soon as they exceed a certain threshold. In the identification experiment, evidence was found that only the two factors vowel duration and voicing maintenance significantly influence the listeners' decisions, but not stop duration. The ratio between stop duration and vowel duration plays a major role in distinguishing stop voicing, but only for highly devoiced stimuli. It is shown that in stimuli without a stop burst, both voicing maintenance, as a major but not required cue, and vowel duration are important acoustic cues for stop voicing distinctions in EP.

Lifetime benefits of musical training.

Lifetime benefits of musical training.

Acoustic characteristics of (alveolo)palatal stop consonants, and velar softening

The paper investigates using data from Majorcan Catalan the acoustic characteristics, and the vowel context and positional conditions, that contribute to the identification of the unaspirated (alveolo)palatal allophone [c] of /k/ as the palatoalveolar affricate /tʃ/ by listeners, and therefore to the implementation of velar softening in the world's languages. Results from perception tests run on [cV] excerpts reveal that affricate percepts are more likely to occur when the (alveolo)palatal stop appears before /i/ than before /a/, which is in agreement with universal patterns of velar softening, and in word-initial and word-final intervocalic position than word-medial intervocalically and utterance initially. Utterance finally [c] is prone to be heard as the fricative [ç]. Affricate identification appears to be associated with context- and position-dependent acoustic cues: high frequency F2 vowel transition endpoints and stop burst spectra, and a long burst, before /i/ and word initially; long range F2 vowel transitions next to /a/, and an intense stop burst in this same vowel context and in intervocalic position. High /tʃ/ identification percentages for [c] in the sequence [uc#u], as well as differences among speakers in producing affricate-like realizations of [c] and among listeners in perceiving the stop as an affricate, are also discussed.

Perceptual and prosodic factors in cluster timing: Manner, order, and syllable position effects in Polish consonant clusters

We investigate timing in Polish tautosyllabic C1C2 clusters differing in manner, consonant order, and syllable position. Hoole et al. (2013) reported for German that perceptual constraints may condition timing differences in /kn/ and /kl/ clusters due to the nasal but not the lateral obscuring the preceding stop burst. Using articulography, we test this hypothesis for a variety of Polish onset clusters (C1={m, p, k}, C2={n, l, r}). Results from three speakers confirm a significant influence of both C1 and C2 on timing patterns. A C1 nasal shows more overlap than a stop. For C2, /l/ shows more overlap than /n/, consistent with the German results. However, the relative difference between C2=/n/ and C2=/l/ holds independently of whether C1 is a nasal or a stop, contra the perception hypothesis. Further, it is known from several languages that onset clusters overlap less than coda clusters, yet this observation has been confounded by the sonority conditioned change in consonant order in onset/coda. Polish has several clusters which do not change order as a function of syllable position, allowing us to tease these two factors apart. If consonant order is kept constant, there is no significant syllable position effect on C-C timing.

Presence of stop bursts and multiple bursts in individuals with Parkinson disease

Studies have reported that individuals with Parkinson Disease (PD) have imprecise articulation of stop consonants due to either slowness of articulators or decreased closure strength. The moment of release for stop consonants, called the burst, has been previously studied in individuals with PD and in other disorders. Multiple bursts (MBs) on the same stop consonant have been reported previously in some motor speech disorders, but no studies are known to have examined MBs in individuals with PD. The current study looked at the occurrences of bursts and MBs in initial stop consonants produced by nine individuals with PD (ON and OFF medication) and nine control speakers. Individuals with PD produced fewer overall bursts compared to control participants. In terms of place of articulation, individuals with PD primarily had loss of bursts in bilabial stops. In addition, individuals with PD had more MBs than control speakers, primarily in alveolars. Finally, no dopamine-related medication effects were found for occurrences of bursts or MBs in individuals with PD. Overall, the study provided evidence for loss of bursts and presence of MBs for stop consonants produced by individuals with PD.

Variability attenuates sensitivity to acoustic detail in cross-language speech production

In the production of non-native consonant clusters, speakers’ systematic errors have been attributed to the influence of native-language phonotactics [Dupoux et al. (1999)]. However, recent models of non-native speech production suggest that speakers are also sensitive to acoustic details [Wilson et al. (2012)]. We examine whether speakers’ sensitivity to phonetic detail is modulated by variability in the speech signal, and whether they abstract away from subphonemic detail given sufficient variability. This was tested by presenting English speakers with ill-formed clusters (e.g., bdafa, tmape, zgade) containing systematically manipulated sub-phonemic acoustic properties: stop burst duration and amplitude for stop-initial clusters, and the presence/absence of pre-obstruent voicing (POV) for voiced clusters. In experiment 1, which presented stimuli produced by a single Russian talker, significant effects were found for the duration manipulations on the rates of epenthesis, the amplitude manipulation on consonant change/deletion errors, and the POV manipulation on the rate of prothesis. In experiment 2, which contained stimuli produced by three talkers, there was a substantial attenuation of the influence of the acoustic manipulations on speakers’ productions. These results suggest that an account of non-native speech production that models the relative contribution of phonotactics and phonetic detail must incorporate information about variability in the environment.

Presence of Stop Bursts and Multiple Bursts in Younger and Older Adults

AbstractThe presence of more than one stop burst on the same stop consonant is termed a multiple burst (MB), and few studies are known to have quantitatively examined the presence of MBs in younger or older adults. Given that there is no known standard metholodology for identifying MBs, the current study describes a methodological approach to identify MBs. The study examines occurrences of stop bursts and MBs in initial stop consonants produced by younger adult (YA) and older adult (OA) speakers. A statistically significant loss of stop bursts was found in OA speakers for bilabials as compared to YA speakers. Although MBs were present across all places of stop articulation, they were predominant for the velar stop consonants. There were no age-related differences in MB occurrence, as they occurred in similar levels in both age groups. Previous research has found that velar stop consonants are more difficult to accurately produce for individuals with dysarthrias, and the current study extends this finding ...