XXXIIe Journées d’Études sur la Parole最新文献

{"title":"Impact des techniques d'adaptation au locuteur dans l'espace des paramètres pour des modèles acoustiques purement neuronaux","authors":"Natalia A. Tomashenko, Y. Estève","doi":"10.21437/jep.2018-64","DOIUrl":"https://doi.org/10.21437/jep.2018-64","url":null,"abstract":"Cet article explore l’utilisation de techniques d’adaptation au locuteur pour des modèles acoustiques bidirectionnels de type long short term memory (BLSTM) entrainés avec la fonction objective dite de classification temporelle connectionniste (CTC). Les modèles acoustiques BLSTM-CTC prennent de plus en plus d’importance dans les systèmes de reconnaissance automatique de la parole, mais peu d’études ont été menées jusqu’ici pour y appliquer des techniques d’adaptation au locuteur. Dans cet article, nous explorons l’utilisation de trois techniques différentes : l’approche par feature space maximum likelihood linear regression (fMLLR), celle s’appuyant sur l’utilisation de i-vectors, et une approche exploitant la technique d’adaptation maximum a posteriori (MAP) appliquée sur des modèles gaussiens dont sont dérivés des paramètres fournis au modèles acoustiques neuronaux. Enfin, cette étude présente une comparaison du comportement des modèles BLSTM-CTC avec celui de modèles markoviens associés à un time-delay neural network (TDNN). of feature-space speaker adaptation for end-to-end acoustic models. This paper investigates speaker adaptation techniques for bidirectional long short term memory (BLSTM) recurrent neural network based acoustic models trained with the connectionist temporal classification (CTC) objective function. BLSTM-CTC AMs play an important role in end-to-end automatic speech recognition systems. However, there is a lack of research in speaker adaptation algorithms for these models. We explore three different feature-space adaptation approaches for CTC acoustic models : feature-space maximum linear regression, i-vector based adaptation, and maximum a posteriori adaptation using GMM-derived features. In addition, the adaptation behavior is compared for BLSTM-CTC models and time-delay neural network (TDNN) models trained with the cross-entropy criterion.","PeriodicalId":310294,"journal":{"name":"XXXIIe Journées d’Études sur la Parole","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"129593941","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Développement de la parole et de la mastication : Evolution de la durée des cycles oscillatoires mandibulaires observés entre 8 et 14 mois chez 4 enfants québécois","authors":"L. Lemarchand, Andrea Macleod, Mélanie Canault, Sophie Kern","doi":"10.21437/JEP.2018-17","DOIUrl":"https://doi.org/10.21437/JEP.2018-17","url":null,"abstract":"Speech and feeding development: A longitudinal study on 4 Quebecois French-speaking children between 8 and 14 months. Speech production and chewing are two motor activities characterized by rhythmic jaw oscillation, which appear throughout the first year of life. The aim of this study was to investigate the evolution of jaw temporal oscillation patterns during nutrition and speech between 8 and 14 months and to determine in which ways these patterns could interact. To address this issue, we monthly recorded 4 Canadian-French-speaking children during speech and meal activities at 8, 10, 12 and 14 months of age. Acoustic and video analyses indicate a significant difference between syllable duration and chewing cycle duration. These results could reflect the early mandible temporal rhythmic organization for chewing and speech. MOTS-CLES : babillage, mastication, acquisition, oscillations mandibulaires.","PeriodicalId":310294,"journal":{"name":"XXXIIe Journées d’Études sur la Parole","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130093591","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Segmentation et Regroupement en Locuteurs: comment évaluer les corrections humaines","authors":"Pierre-Alexandre Broux, David Doukhan, Simon Petitrenaud, Sylvain Meignier, Jean Carrive","doi":"10.21437/JEP.2018-11","DOIUrl":"https://doi.org/10.21437/JEP.2018-11","url":null,"abstract":"Dans cet article, nous presentons un simulateur dedie a l'evaluation des corrections humaines sur la tâche de Segmentation et Regroupement en Locuteurs (SRL). Nous proposons quatre actions elementaires afin de corriger une SRL et un automate pour simuler la sequence de corrections. Une mesure est proposee pour evaluer le cout de correction. Le simulateur est evalue en utilisant des emissions francaises d'information tirees du corpus REPERE. ABSTRACT Computer-assisted speaker diarization : how to evaluate human corrections In this paper, we present a framework to evaluate the human correction of a speaker diarization. We propose four elementary actions to correct the diarization and an automaton to simulate the correction sequence. A metric is described to evaluate the correction cost. The framework is evaluated using French broadcast news drawn from the REPERE corpus.","PeriodicalId":310294,"journal":{"name":"XXXIIe Journées d’Études sur la Parole","volume":"13 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130741036","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Efforts de production de parole chez les personnes qui bégaient","authors":"M. Garnier, Anaïs Dafonseca, C. Savariaux, Thibault Cattelain","doi":"10.21437/JEP.2018-10","DOIUrl":"https://doi.org/10.21437/JEP.2018-10","url":null,"abstract":"Cette etude vise a objective et quantifier les tensions orofaciales visibles et communement decrites chez des personnes qui begaient (PQB), en comparant les efforts areeodynamiques, larynges et articulatoires lors de la production de consonnes occlusives entre des PQB et des personnes normofluentes (PNF). Dans cette etude preliminaire, 4 adultes qui begaient et 4 adultes normofluents ont ete enregistres. Ils devaient produire des mots commencant par des consonnes occlusives labiales dans une tâche de lecture et dans une tâche semi-spontanee. Contrairement a nos attentes, les 4 PQB ont montre une force interlabiale (mesuree a partir d'un capteur de pression) et une pression intra-orale reduites par rapport aux 4 PNF. Aucune difference significative n'a ete observee au niveau de ces parametres entre les syllabes begayees v. perceptivement fluentes des PQB. Cependant, les differences de pression intra-orale observes entre les PQB et le PNF sont renforcees sur des syllabes complexes (commencant par un cluster consonantique) et pour la tâche semi-spontanee.","PeriodicalId":310294,"journal":{"name":"XXXIIe Journées d’Études sur la Parole","volume":"33 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114550885","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Étude exploratoire des événements articulatoires pendant la réalisation de pauses en parole spontanée","authors":"Ivana Didirkova, Camille Fauth, Sébastien Le Maguer","doi":"10.21437/JEP.2018-51","DOIUrl":"https://doi.org/10.21437/JEP.2018-51","url":null,"abstract":"Les disfluences en tant qu’événements arrêtant momentanément le déroulé du discours font partie intégrante de la production de la parole. De nombreuses études ont analysé leurs caractéristiques en se fondant principalement sur leurs propriétés acoustiques. L’objectif de cette étude est d’enrichir la description acoustique des disfluences produites par quatre locuteurs en parole spontanée en la corrélant à une description articulatoire, obtenue à partir d’un articulographe électromagnétique (EMA), synchronisée avec les données acoustiques. Pour ce faire, nous avons observé leur production à l’aide de deux différentes méthodes : une étude articulatoire effectuée de façon automatique à partir de la vitesse des mouvements des articulateurs et une observation experte du mouvement de ces mêmes articulateurs. Les résultats de cette étude exploratoire montrent l’existence de plusieurs configurations articulatoires pour un même type perçu de disfluence et inversement, plusieurs catégories acoustiques peuvent parfois être traduites par une même configuration articulatoire. De même, l’anticipation du geste articulatoire relatif au phone subséquent semble dépendre de l’activité articulatoire supra-glottique. An exploratory study of articulatory events during pause realization in spontaneous speech. Disfluencies, as a momentary stop event in a discourse, are considered as a constitutive part of speech production. Numerous studies analyzed their specificities by relying mainly on their acoustic properties. The goal of the presented study is to extend the acoustic analysis by including a articulatory description. This description is obtained using Electro-Magnetic Articulography (EMA) synchronized with the acoustic data. To achieve this study, we have observed the production of the disfluencies following two complementary methodologies: an automatic articulatory study based on the quantification of the velocity of the articulators and an expert subjective analysis of the movement of these articulators. The results of this exploratory study show the same perceived disfluency category can rely on multiple articulatory configurations. Multiple acoustic categories can also be realized following a same articulatory configuration. Furthermore, the anticipation of the subsequent gesture seems to be dependent on the supraglottic activity.","PeriodicalId":310294,"journal":{"name":"XXXIIe Journées d’Études sur la Parole","volume":"35 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123378035","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Entre Québec et France, qu'en est-il de l'antériorisation de /ɔ/ en français contemporain ?","authors":"Xavier St-Gelais, Christophe Coupé, F. Pellegrino, Vincent Arnaud","doi":"10.21437/JEP.2018-59","DOIUrl":"https://doi.org/10.21437/JEP.2018-59","url":null,"abstract":"Quebec province, an acoustic analysis of the second formant frequency (F 2 ) of 2837 vowels uttered in (C)VC monosyllabic words and non-words by university students from Saguenay (Quebec) and Lyon (France) was conducted. By fitting a linear mixed effects regression to the data, /ɔ/ is found to be significantly more fronted in Lyon than in Saguenay. In both dialects, the places of articulation of the initial and final consonant also influence the vowel’s acoustical structure. No matter their position, front consonants (ex. /t, d/) favour a higher F 2 , while labials (ex. /p, b/) are associated with the lowest F 2 .","PeriodicalId":310294,"journal":{"name":"XXXIIe Journées d’Études sur la Parole","volume":"90 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116890503","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Étude exploratoire des stratégies de production du ton 3 en chinois mandarin","authors":"Yizhi Huang, V. Delvaux, Kathy Huet, M. Piccaluga, Guoxian Zhang, B. Harmegnies","doi":"10.21437/JEP.2018-62","DOIUrl":"https://doi.org/10.21437/JEP.2018-62","url":null,"abstract":"Basing on a corpus consisted of monosyllabic words and non-words produced in citation form by 10 native speakers of Mandarin Chinese, the present study aims at describing the occurences of non-modal phonation in Mandarin Chinese, particularly in T3 production, and the characteristics of the strategies which are implemented in tone production with various acoustic measurements.","PeriodicalId":310294,"journal":{"name":"XXXIIe Journées d’Études sur la Parole","volume":"7 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121114084","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}