Search

International audience; This paper details a study to evaluate the effects of modified speech on perceptual speaker identification (SID) performance by naive listeners. Speech recordings made by eight male, native-French speakers were selected from the PTSVox database. The pitch and speech tempo of the recordings were modified at the word-level. The first 75% of words spoken were modified, such that the percentage of modification began at 100% and gradually decayed to 0%. The direction of the modifications was also examined, such that pitch modifications began at ±600 cents and speech tempo modifications began at a ratio of either 1:2 or 3:2 (modified to normal speech tempo). Following a familiarization period, participants completed two rounds of 48 "go/no-go" task trials (balanced), where each round corresponded to a different speech modification type. The main results showed perceptual SID performance was significantly affected when participants were presented speech recordings that contained pitch modifications in comparison to speech tempo modifications. The findings revealed participants were able to overcome higher percentages of speech tempo modifications to make correct distinctions between speakers. Although modified pitch influenced in voice perception performance, high variability between participant responses were observed, which suggests listeners model speakers differently.

International audience; Human languages have the flexibility to be acoustically adapted to the context of communication, such as in shouting or whispering. Drummed forms of languages represent one of the most extreme natural expressions of such speech adaptability. A large amount of research has been conducted on drummed languages in anthropology or linguistics, particularly in West African societies. However, in spite of the clearly rhythmic nature of drumming, previous studies have largely neglected exploring systematically the role of speech rhythm. Here, we explore a unique corpus of the Bendré drummed speech form of the Mossi people, transcribed published in the 80’s by the anthropologist Kawada Junzo. The analysis of this large database in Mooré language reveals that the rhythmic units encoded in the length of pauses between drumbeats match more closely with vowel-to-vowel intervals than with syllable parsing. Meanwhile, we confirm for the first time a result found recently on the drummed speech tradition of the Bora Amazonian language. However, the complex acoustic structure of the Bendré skin drum required much more attention than the simple two pitch hollow log drum of the Bora. Thus, we also present here results on how drummed Bendré timbre encodes tones of Mooré language.

International audience; Predicting the acoustic and linguistic parameters of an upcoming conversational turn is important for dialogue systems aiming to include low-level adaptation with the user. It is known that during an interaction speakers could influence each other speech production. However, the precise dynamics of the phenomena is not well-established, especially in the context of natural conversations. We developed a model based on an RNN architecture that predicts speech variables (Energy, F0 range and Speech Rate) of the upcoming turn using a representation vector describing speech information of previous turns. We compare the prediction performances when using a dialogical history (from both participants) vs. monological history (from only upcoming turn's speaker). We found that the information contained in previous turns produced by both the speaker and his interlocutor reduce the error in predicting current acoustic target variable. In addition the error in prediction decreases as increases the number of previous turns taken into account.

International audience; Phonologization is a process whereby phonetic substance becomes phonological structure [1]. The process involves at least two steps: (i) a universal phonetic ('automatic') variation becomes a language-specific ('speaker-controlled') pattern, (ii) the language-specific pattern becomes a phonological ('structured') object. This paper will focus on the first step and ask the question of whether three universal phonetic variations of the laryngeal feature of word-final codas (final devoicing, voicelessness assimilation and voicing assimilation) are becoming language-specific patterns in two Romance languages, Romanian and French. Our results suggest that neutralization processes (final devoicing) might be beginning their phonologization process in both French and Romanian whereas assimilation processes (regressive assimilation of voicing and voicelessness) remain universal phonetic tendencies.

International audience; Languages tend to license segmental contrasts where they are maximally perceptible, i.e. where more perceptual cues to the contrast are available. For strident fricatives, the most salient cues to the presence of voicing are low-frequency energy concentrations and fricative duration, as voiced fricatives are systematically shorter than voiceless ones. Cross-linguistically, the voicing contrast is more frequently realized word-initially than word-finally, as for obstruents. We investigate the phonetic underpinnings of this asymmetric behavior at the word edges, focusing on the availability of durational cues to the contrast in the two positions. To assess segmental duration, listeners rely on temporal markers, i.e. jumps in acoustic energy which demarcate segmental boundaries, thereby facilitating duration discrimination. We conducted an acoustic analysis of wordinitial and word-final strident fricatives in American English. We found that temporal markers are sharper at the left edge of word-initial fricatives than at the right edge of word-final fricatives, in terms of absolute value of the intensity slope, in the high-frequency region. These findings allow us to make predictions about the availability of durational cues to the voicing contrast in the two positions.

International audience

International audience; Sonority sequencing principles can account for phonotactic processes across languages, but the articulatory correlates of sonority are not well established. This study examines the relationship between gestural overlap and sonority shape for two-consonant onsets in Georgian via electromagnetic articulography. We cross three sonority shapes (rise, plateau, fall) with two orders of place of articulation (front-back, backfront), since the latter has been shown to affect overlap. The degree of plateau overlap and relative overlap for the consonant gestures are evaluated. Both sonority shape and order of place of articulation are found to affect overlap. Importantly, in order to understand how Georgian organizes onsets of different sonority shapes we must look at both measures of overlap. The plateau overlap measure reveals that all Georgian clusters present a lag between achieved constrictions; we argue that this languagespecific setting allows for a wider range of consonant sequences to be perceptible. In parallel, the overall overlap measure shows that sonority falls are more overlapped during the formation phase of the second consonant's constriction, which presumably compensates for the later plateau lag and ensures tautosyllabic parsing of the clusters' consonants in sonority falls. Finally, front-to-back clusters are less overlapped than back-to-front clusters across all sonority shapes.

International audience; The perception of intonational meaning is affected by multiple features of the f0 curves. It remains unclear however how these features interact and how they are integrated in the perception of intonational categories. In this study we investigate how the shape of the f0 curves and the lag between f0 targets and tone bearing syllables (i.e. tonal alignment) affect the perception of high pitch accents. To this aim we conceived a minimal dynamical model accounting for the perception of abstract tonal categories from the continuous f0 curve. In the model the internal representation of the f0 curve is continuously updated by a law of change that depends on features of the actual f0 curve but also on the current state of its internal representation. Through this model, we could successfully simulate the main findings observed so far in the literature concerning the interaction between f0 shape and tonal alignment in the perception of High tones. The success of the model demonstrates the role of perceptual dynamics in the processing of intonational categories. Moreover, it permits explaining the partial success of previous accounts of the same facts as well as their shortcomings.

State-of-the-art speech recognition systems rely on fixed, hand-crafted features such as mel-filterbanks to preprocess the waveform before the training pipeline. In this paper, we study end-to-end systems trained directly from the raw waveform, building on two alternatives for trainable replacements of mel-filterbanks that use a convolutional architecture. The first one is inspired by gammatone filterbanks (Hoshen et al., 2015; Sainath et al, 2015), and the second one by the scattering transform (Zeghidour et al., 2017). We propose two modifications to these architectures and systematically compare them to mel-filterbanks, on the Wall Street Journal dataset. The first modification is the addition of an instance normalization layer, which greatly improves on the gammatone-based trainable filterbanks and speeds up the training of the scattering-based filterbanks. The second one relates to the low-pass filter used in these approaches. These modifications consistently improve performances for both approaches, and remove the need for a careful initialization in scattering-based trainable filterbanks. In particular, we show a consistent improvement in word error rate of the trainable filterbanks relatively to comparable mel-filterbanks. It is the first time end-to-end models trained from the raw signal significantly outperform mel-filterbanks on a large vocabulary task under clean recording conditions., Accepted for presentation at Interspeech 2018

International audience; This paper investigates the use of prosodic information signalling sentence accent and the role of different acoustic features on sentence accent perception during native and non-native speech perception in the presence of background noise. A phoneme detection experiment was carried out in which English native listeners and French highly proficient non-native listeners of English were presented with target phonemes in English sentences. Sentences were presented in different levels of speech-shaped noise and in two prosodic contexts in which the target-bearing word was either deaccented or accented. Acoustic analyses of the two prosodic conditions showed that the target-bearing words in the accented condition carried more energy, had a higher F0, and more spectral tilt than those in the deaccented condition. Results of the behavioural data showed that the native listeners outperformed the French listeners in the clean condition but not in the noise conditions and that the effect of noise was smaller for the non-native compared to the native listeners. Possibly, the non-native listeners use more and different acoustic cues than the native listeners who primarily relied on more local cues for sentence accent detection.

International audience; The present study deals with the productions of English/French bilingual speakers in a reading task and in semi-spontaneous speech. Mean fundamental frequency, F0 range and F0 standard-deviation were measured in both languages. Results show a significant effect of gender and language on all these parameters. Overall, mean F0 was higher in French while F0 modulation was stronger in English. Regardless of language, female speakers exhibited higher F0 than males. Moreover, the increase of mean F0 in French was larger in female speakers. On the other hand, the decrease of F0 modulation in French was stronger for male speakers. These data support the idea of language-and gender-specific vocal codes, to which bilingual speakers seem to adapt. MOTS-CLES : fréquence fondamentale, intonation, bilinguisme, voix et genre, variations inter-langues.; La présente étude porte sur les productions de locutrices et locuteurs bilingues anglais/français lors d'une tâche de lecture de phrases et en parole semi-spontanée. La fréquence fondamentale (F0) moyenne, la plage de variation de F0 et l'écart-type de F0 ont été mesurés dans les deux langues. Les résultats indiquent un effet significatif des facteurs langue et genre sur l'ensemble de ces paramètres. La F0 moyenne est globalement plus élevée en français ; la modulation de F0 est quant à elle globalement plus forte en anglais. Indépendamment de la langue, les locutrices présentent une F0 moyenne plus élevée. En outre, les locutrices ont plus fortement augmenté leur F0 moyenne lors de l'emploi du français et les locuteurs ont réduit plus fortement leur modulation dans cette langue. Ces données suggèrent l'existence de normes vocales relatives au genre différentes dans les deux langues étudiées.

Our goal is to show the interest of distinguishing several phases in the gesture underlying the production of labial consonants, including during the occlusion, where lips are essentially static, in order to better understand their control. Four adult speakers were recorded, while producing logatoms including the consonants /p/ and /b/, followed by 3 vowels (/a/, /i/ or /u/) and with increasing levels of articulatory effort. Five surface EMG signals targeting five orofacial muscles involved in several lip movements were recorded simultaneously with a video of the lips. The speech gestures were segmented into four phases, based on the interlabial distance. Similar muscle activation patterns were observed for the four speakers, regardless of the consonant (/p/ or /b/) and the vowel context. Finally, the activities of the lower lip depressor and the mentalis muscles in some movement phases appear to be reliable descriptors of articulatory effort., Notre objectif est de montrer l'intérêt d'une d'analyse temporelle détaillée des gestes de production des consonnes occlusives labiales, en distinguant différentes phases, y inclus durant l'occlusion où les lèvres sont immobiles. Quatre adultes ont été enregistrés pour cette étude. Il leur était demandé de produire des logatomes contenant les plosives /p/ et /b/, dans différents contextes vocaliques (/a/, /i/, /u/) et avec différents niveaux d'effort articulatoire. Les signaux EMG de surface de cinq muscles péri-oraux ont été acquis simultanément avec une vidéo des lèvres. Les mouvements de parole ont été segmentés en quatre phases à l'aide du signal cinématique de distance interlabiale. Des patrons d'activation musculaire semblables ont été observés chez les quatre locuteurs, quels que soient la consonne et le contexte vocalique. Enfin, les activités du muscle dépresseur de la lèvre inférieure et du mentalis dans certaines phases du mouvement semblent être de bons descripteurs de l'effort articulatoire.

International audience; Speech perception often involves multisensory processing. Although previous studies have demonstrated visual [1, 2] and somatosensory interactions [3, 4] with auditory processing, it is not clear whether somatosensory information can contribute to the processing of audiovisual speech perception. This study explored the neural consequence of somatosensory interactions in audiovisual speech processing. We assessed whether somatosensory orofacial stimulation influenced event-related potentials (ERPs) in response to an audiovisual speech illusion (the McGurk Effect [1]). 64 scalp sites of ERPs were recorded in response to audiovisual speech stimulation and somatosensory stimulation. In the audiovisual condition, an auditory stimulus /ba/ was synchronized with the video of congruent facial motion (the production of /ba/) or incongruent facial motion (the production of the /da/: McGurk condition). These two audiovisual stimulations were randomly presented with and without somatosensory stimulation associated with facial skin deformation. We found ERPs differences associated with the McGurk effect in the presence of the somatosensory conditions. ERPs for the McGurk effect reliably diverge around 280 ms after auditory onset. The results demonstrate a change of cortical potential of audiovisual processing due to somatosensory inputs and suggest that somatosensory information encoding facial motion also influences speech processing.

International audience; According to accounts of inter-speaker coordination based on internal predictive models, speakers tend to imitate each other each time they need to coordinate their behavior. According to accounts based on the notion of dynamical coupling, imitation should be observed only if it helps stabilizing the specific coordinative pattern produced by the interlocutors or if it is a direct consequence of inter-speaker coordination. To compare these accounts, we implemented an artificial agent designed to repeat a speech utterance while coordinating its behavior with that of a human speaker performing the same task. We asked 10 Italian speakers to repeat the utterance /topkop/ simultaneously with the agent during short time intervals. In some interactions, the agent was parameterized to cooperate with the speakers (by producing its syllables simultaneously with those of the human) while in others it was parameterized to compete with them (by producing its syllables in-between those of the human). A positive correlation between the stability of inter-speaker coordination and the degree of f0 imitation was observed only in cooperative interactions. However, in line with accounts based on prediction, speakers imitate the f0 of the agent regardless of whether this is parameterized to cooperate or to compete with them.

International audience; Motor equivalence is a key feature of speech motor control, since speakers must constantly adapt to various phonetic contexts and speaking conditions. The Uncontrolled Manifold (UCM) idea offers a theoretical framework for considering motor equivalence. In this framework coordination among motor control variables is separated into two subspaces, one in which changes in control variables modify the acoustic output and another one in which these changes do not influence the output. Our work develops this concept for speech production using a 2D biomechanical model of the tongue, coupled with a jaw and lip model, for vowel production. We first propose a representation of the linearized UCM based on orthogonal projection matrices. Next we characterize the UCMs of various vocal tract configurations of the 10 French oral vowels using their perturbation responses. We then investigate whether these UCMs describe phonetic classes like phonemes, front/back vowels, rounded/unrounded vowels, or whether they significantly vary across representatives of these different classes. We found they clearly differ between rounded and unrounded vowels, but are quite similar within each category. This suggests that similar motor equivalence strategies can be implemented within each of these classes and that UCMs provide a valid characterization of an equivalence strategy.

International audience; This paper proposes an approach, named phonetic context embedding , to model phonetic context effects for deep neural network hidden Markov model (DNN-HMM) phone recognition. Phonetic context embeddings can be regarded as continuous and distributed vector representations of context-dependent phonetic units (e.g., triphones). In this work they are computed using neural networks. First, all phone labels are mapped into vectors of binary distinctive features (DFs, e.g., nasal/not-nasal). Then for each speech frame the corresponding DF vector is concatenated with DF vectors of previous and next frames and fed into a neural network that is trained to estimate the acoustic coefficients (e.g., MFCCs) of that frame. The values of the first hidden layer represent the embedding of the input DF vectors. Finally, the resulting embeddings are used as secondary task targets in a multi-task learning (MTL) setting when training the DNN that computes phone state posteriors. The approach allows to easily encode a much larger context than alternative MTL-based approaches. Results on TIMIT with a fully connected DNN shows phone error rate (PER) reductions from 22.4% to 21.0% and from 21.3% to 19.8% on the test core and the validation set respectively and lower PER than an alternative strong MTL approach.

International audience

International audience; In English Question+Answer (Q+A) pairs, periodicity typically emerges across turn space, to a degree of precision matching standards of music perception. Interactionally-aligned Q+A pairs display such shared periodicity across the turn, while unaligned pairs do not. Periodicity is measured as temporal location of f0 maxima or minima, 'pikes', in successive accented syllables. This study asks whether periodicity of pikes across a turn is accompanied by systematic use of musical pitch intervals across the turn space. Recordings of 77 Q+A pairs from 8 pairs of native English speakers talking naturally. Ratios of f0 in the last pike of the Question and the first of the Answer fell more reliably into Western musical interval categories when the Q+A pair's turn transition was periodic (the Answer was aligned or preferred, re the Question) than when it was aperiodic (disaligned, dispreferred). Similar results were found for ratios of modal f0. Such pitch ratios are better described by musical interval categories of Western tuning systems than by those of three non-Western systems, and best of all by semitones, suggesting close connections between culturally-specific uses of pitch in conversation and in music. Judgments of arousal/valence suggest weak relations with specific pitch intervals. Theoretical implications are discussed.

International audience; While the use of sentence-final discourse particles (SFPs) is typically linked to specific interactional or social functions, their realization is also associated with particular positional and intonational requirements. This raises the question of whether the use of SFPs may be partly driven by the phonological characteristics of sentence-final contexts. In [1], we showed that Singapore English lah is overrepresented in contexts involving sentence-final stress and underrepresented elsewhere. This is surprising if the use of lah is motivated by purely pragmatic considerations, but can be explained if (i) lah is recruited where it can relieve tonal crowding, or (ii) lah is avoided when it would result in a long sequence of non-prominent syllables. Such behavior is expected to be more prevalent for SFPs (like lah) whose pragmatic function is very general, but less prevalent for SFPs with a more specialized function. In this study, we consider the distributions of a wider range of Singapore English SFPs, including leh, lor, ah and hor. Overall, these particles were more evenly distributed across prosodic contexts compared to lah, suggesting that prosodic context conditions the use of SFPs, but only when this does not interfere with the speaker's intended message.

International audience; The growth of the head and neck and its components, including that of the vocal tract, is not homothetic but appears rather as an anamorphosis. The growth of various structures presents a phenomenon of heterochrony. Another important issue in vocal tract growth is sexual dimorphism. It was first claimed that sexual dimorphism appears at puberty, but a recent study has suggested that some prepubertal differences exist. To study these two phenomena, we used longitudinal radiographic data of sixty-eight typical subjects (966 radiographs, taken from 1 month to 25 years) and twelve fetuses (anatomical sections). In this study, we analyzed the growth curves and growth types of the hard and soft palate, the pharyngeal cavity and the estimated length of the whole vocal tract using non-linear mixed-effect models, in order to take advantage of our unique longitudinal dataset. Results indicate that most of the structures follow a neural/somatic growth type, while the pharyngeal cavity follows a more somatic growth type. As concerns sexual dimor-phism, no prepubertal differences were found, suggesting that the sexual dimorphism is likely to begin at puberty. These results have implications for the acoustics of speech production during development and should lead to improvements in vocal tract growth modeling.

International audience; Investigations of changes in speech rate and speaking style have reported that locus equation parameters are sensible to temporal variation though they usually do not evidence any ma-jor impact on place of articulation classification. The aim of the present study was to investigate the role of vowel length con-trasts on locus equation parameters in Jordanian Arabic (JA) speech production in order to study the influence of intrinsic temporal changes (vowel length contrasts) on locus equation parameters. Statistical analyses of slopes and intercepts show that the locus equation parameters of a consonant produced with long vowels are significantly different from those of the same consonant produced with short vowels. Though this obseva-tion confirms the impact of time variations on locus equations, further analyses will need to address issues related to the po-tential impact of duration-related spectral modifications and the role these changes may play on the perceptual categorization of consonantal place of articulation.

International audience; We present a study where we examined the influence of a regional accent in the perception of voice and/or speech disorders. These aspects are most of the time overshadowed in clinical context. This protocol, involving multiple sources of speech variations, is also interesting in perception theories. For the experiment, speakers with or without a Southern French accent and with or without speech/voice disorders were recorded on reading a text. The samples were then randomly played back to two groups of listeners (familiar vs non familiar with the regional accent), specialists in speech therapy. The task was the perceptual evaluation of voice quality, articulation disorders and dysprosody. We focused in this paper on the voice dimension. The main results on this part concern the weak influence of regional accent on the perception of moderate or severe dysphonia, where the speech signal is strongly disturbed by the disorder. By contrast, the effect of regional accent is important on normal voices perception: listeners non familiar with the regional accent judge speakers with accent without voice disorder as slightly dysphonic. This last result can be interpreted as a form of perceptual interference between different dimensions of speech variations around a central position.

International audience; Talkers adopt different speech styles in response to factors such as the perceived needs of the interlocutor, environmental noise and explicit instruction. Some styles have been shown to be beneficial for listeners but many aspects of the relationship between speech modifications and intelligibility remain unclear, particularly for prosodic changes. The current study measures the relative intelligibility in noise of speech spoken in 5 speech styles-plain, infant-, computer-and foreigner-directed, and shouted-and relates listener scores to acous-tic/prosodic parameters and quantitative estimates of energetic masking. Intelligibility changes over plain speech correlated well with durational modifications, which included elongations of all segments as well as increases in the number of unfilled pauses. Both mean fundamental frequency and its range displayed great variation across styles but with no clear intelligibility benefits. Energetic masking per unit time was similar in each style but the total amount of speech which escaped masking was a good predictor of word identification rate. These findings suggest that much of the prosody-related intelligibility gain is derived from durational increases.

(accepted); International audience; Pronunciation training based on speech production techniques illustrating tongue movements is gaining popularity. However, there is not sufficient evidence that learners can imitate some tongue animation. In this paper, we investigated human awareness of controlling their tongue body gestures. In a first experiment, participants were asked to perform some tongue movements. This task was evaluated by observing ultrasound imaging of the tongue recorded during the experiment. In a second experiment, a short session of training was added where participants can observe ultrasound imaging in realtime of their own tongue movements. The goal was to increase their awareness of their tongue gestures. A pre-test and post-test were carried out without any feedback. The results suggest that it is not easy to finely control tongue body gestures; and that we gain in performance after a short training session which suggests that providing visual feedback improves tongue gesture awareness.

International audience; Speakers of a given language vary with respect to their acoustics, articulation, and motor commands. This variation is driven by a variety of influences, such as emotional states, communicative interaction, and individual properties of the vocal tract. In this work we focus on the latter. First, we build speaker-specific biomechanical tongue models. Second, we discuss the impact of the relative position of the bending in the vocal tract on the basis of extensive simulations with two different models. We focus on /i,a,u/ by defining target regions in the acoustic space, and discuss the corresponding speaker-specific articulatory and motor command variability observed.

International audience; This study investigates the effect of “adverse” contexts, especially that of the consonant /ʃ/, on labial parameters for French /i,y/. Five parameters were analysed: the height, width and area of lip opening, the distance between the corners of the mouth, as well as lip protrusion. Ten speakers uttered a corpus made up of isolated vowels, syllables and logatoms. A special procedure has been designed to evaluate lip opening contours. Results showed that the carry-over effect of the consonant /ʃ/ can impede the opposition between /i/ and /y/ in the protrusion dimension, depending upon speakers.

International audience; This study investigates jaw-finger coordination in a task consisting in pointing to a target while naming it with a /pata/ or a /tapa/ utterance stressed either on the first ('CVCV) or on the second (CV'CV) syllable. Optotrack measurements of jaw and finger displacements show that for 'CVCV names, the moment at with the finger reaches the target alignment is synchronized with the maximum of the first jaw opening motion. For CV'CV names, the synchronization occurs between the moment at which the finger leaves the target-alignment position and the maximum of the jaw opening motion for the second vowel. This pattern of synchronization does not depend on the target position or on the consonants order. These results add some support to theories involving the coordination of orofacial and brachiomanual gestures in the development and phylogeny of human languages. They call for more investigations on the link between speech and brachiomanual gestures in face-to-face communication.

International audience; In a companion paper [1], we showed that CVCV utterances with a labial consonant followed by a coronal one (LC sequences) are more stable than reverse CL sequences in speeded reiterant speech. We proposed that this could explain why human languages select LC sequences more often than CL ones (the “LC effect”). We provide here articulatory data explaining where the greater LC stability could come from, by investigating inter-articulator coordination during LC and CL utterances at an increasing rate. Rate increase leads variegated CVCV (e.g. /pata/) to be produced in a single jaw cycle but this is not the case for duplicated CVCV (e.g. /papa/. Furthermore, LC and CL sequences both evolve towards the same cycle with a progressive phasing of lips and tongue close together in the jaw cycle. Taken together, these results provide new elements to argue for motor control constraints shaping phonological patterns from economy principles.

International audience; This paper presents a perceptual experiment on stimuli synthesized by means of a vocal tract area function model. The purpose was to compare the contribution of dynamic against static information to the identity of a coarticulated vowel. Three sources of information were perceptually analyzed: (i) the vowel nucleus; (ii) the acoustical contrast between the vowel nucleus and the ``stationary'' parts of its immediate context; (iii) and the transitions linking the stable parts of the speech signal. The results show that the vocoïds were better identified by dynamic information. This backs up the perceptual overshoot model proposed Lindblom. However, this conclusion must be confirmed by further experiments.