Normalisation articulatoire du locuteur par méthodes de décomposition tri-linéaire basées sur des données IRM (Articulatory speaker normalisation based on MRI-data using three-way linear decomposition methods) [in French]
1808Authors
nan
Abstract
________________________________________________________________________________________________ Articulatory speaker normalisation based on MRI-data using three-way linear decomposition methods The aim of this study was to characterise, to model and to compare the different lingual articulatory strategies of a group of speakers. Individual principal component analysis (PCA) models and multi-linear decomposition methods have been applied to the tongue contours extracted from a magnetic resonance imaging (MRI) corpus of seven speakers articulating 63 French vowels and consonants. On the average over the seven speakers, using 4 components, the Root Mean Square prediction Error (RMSE) was 0.13 cm for the individual PCA models while the RMSE for the parallel factor model (PARAFAC) was 0.29 cm, accounting for a percentage of variance explanation of 91% and 62%, respectively. A multi-linear regression (MRL) model could predict, with 10 components, the tongue contour of a target subject from a given source subject, with about 65% of the variance explained and an RMSE of 0.38 cm. All the models have been assessed by a leave-one-out cross-validation procedure. MOTS-CLES: Modélisation articulatoire, normalisation du locuteur, analyse factorielle, IRM. KEYWORDS: Articulatory modelling, speaker normalisation, factor analysis, MRI. 529 1.
Keywords
Full Content
Actes de la conférence conjointe JEP-TALN-RECITAL 2012, volume 1: JEP, pages 529–536, Grenoble, 4 au 8 juin 2012. c ⃝2012 ATALA & AFCP Normalisation articulatoire du locuteur par méthodes de décomposition tri-linéaire basées sur des données IRM Julián Andrés Valdés Vargas1, Pierre Badin1, G. Ananthakrishnan2, Laurent Lamalle3 (1) GIPSA-lab (Département Parole & Cognition), UMR 5216 CNRS – Grenoble University (2) Centre for Speech Technology, KTH (Royal Institute of Technology), Stockholm, Sweden (3) SFR1 RMN Biomédicale et Neurosciences (Unité IRM Recherche 3 Tesla), INSERM, CHU de Grenoble {julian-andres.valdes-vargas,Pierre.Badin}@gipsa-lab.grenoble- inp.fr, agopal@kth.se,Llamalle@ujf-grenoble.fr RESUME __________________________________________________________________________________________________ Le but de cette étude était de caractériser, modéliser et comparer les différentes stratégies articulatoires linguales pour un groupe de locuteurs. Des modèles individuels par analyse en composantes principales (ACP) et des méthodes de décomposition multilinéaires ont été appliqués aux contours de langue extraits d‘un corpus d’imagerie par résonance magnétique (IRM) de sept locuteurs prononçant 63 voyelles et consonnes du français. En moyenne sur les sept locuteurs, en utilisant quatre composantes, l'erreur quadratique moyenne de prédiction (RMSE) était de 0,13 cm pour les modèles individuels ACP et de 0.29 cm pour le modèle ‘parallel factor’ (PARAFAC), avec des pourcentages de variance expliquée de 91% et 62%, respectivement. Un modèle de régression multilinéaire permet également de prédire avec 10 composantes les contours de langue d’un sujet cible à partir de ceux d'un sujet source avec approximativement 65% de la variance expliquée et une RMSE de 0.38 cm. Tous les modèles ont été évalués par une procédure de validation croisée. ABSTRACT ________________________________________________________________________________________________ Articulatory speaker normalisation based on MRI-data using three-way linear decomposition methods The aim of this study was to characterise, to model and to compare the different lingual articulatory strategies of a group of speakers. Individual principal component analysis (PCA) models and multi-linear decomposition methods have been applied to the tongue contours extracted from a magnetic resonance imaging (MRI) corpus of seven speakers articulating 63 French vowels and consonants. On the average over the seven speakers, using 4 components, the Root Mean Square prediction Error (RMSE) was 0.13 cm for the individual PCA models while the RMSE for the parallel factor model (PARAFAC) was 0.29 cm, accounting for a percentage of variance explanation of 91% and 62%, respectively. A multi-linear regression (MRL) model could predict, with 10 components, the tongue contour of a target subject from a given source subject, with about 65% of the variance explained and an RMSE of 0.38 cm. All the models have been assessed by a leave-one-out cross-validation procedure. MOTS-CLES: Modélisation articulatoire, normalisation du locuteur, analyse factorielle, IRM. KEYWORDS: Articulatory modelling, speaker normalisation, factor analysis, MRI. 529 1. Introduction The Speech & Cognition Department at GIPSA-lab has developed acoustic-to-articulatory inversion methods to provide speakers with a visual articulatory feedback (Ben Youssef et al., 2011), based on a fairly complete orofacial clone. This clone is made of a set of models of articulators (jaw, tongue, velum, lips, etc.) based on articulatory data acquired on a single speaker (Badin & Serrurier, 2006). Therefore, the clone represents faithfully …
Content truncated to 500 words
Labeled Text
________________________________________________________________________________________________ Articulatory speaker normalisation based on MRI-data using three-way linear decomposition methods The aim of this study was to characterise, to model and to compare the different lingual articulatory strategies of a group of speakers. Individual principal component analysis (PCA) models and multi-linear decomposition methods have been applied to the tongue contours extracted from a magnetic resonance imaging (MRI) corpus of seven speakers articulating 63 French vowels and consonants. On the average over the seven speakers, using 4 components, the Root Mean Square prediction Error (RMSE) was 0.13 cm for the individual PCA models while the RMSE for the parallel factor model (PARAFAC) was 0.29 cm, accounting for a percentage of variance explanation of 91% and 62%, respectively. A multi-linear regression (MRL) model could predict, with 10 components, the tongue contour of a target subject from a given source subject, with about 65% of the variance explained and an RMSE of 0.38 cm. All the models have been assessed by a leave-one-out cross-validation procedure. MOTS-CLES: Modélisation articulatoire, normalisation du locuteur, analyse factorielle, IRM. KEYWORDS: Articulatory modelling, speaker normalisation, factor analysis, MRI. 529 1.
Publication Details
- Paper ID:
- 1808
- Year:
- 2012
- Publication:
- TIPSTER
- Publication Type:
- conf
- DOI:
- nan