Modeling tongue-palate contact patterns in the production of speech

Abstract

Previous studies on lingual movements in speech have led to the assumption that the tongue can be divided into a small number of functional blocks. This work was aimed at nding out whether such a subdivision into independent components may also be observed in patterns of tongue-palate contacts, as revealed by electropalatography (EPG). An attempt was made to reduce EPG data to a small number of articulatorily relevant parameters in an empirical way, and to model the con guration of the linguo-palatal contacts in speech as a combination of these parameters. Our corpus was composed of ten natural sentences recorded by two female native speakers of French. EPG data reduction was performed by means of a feedforward multilayer neural network on the one hand, and of a factor analysis on the other hand.The results showed that variations in the tongue-palate contact pattern mainly occurred along two dimensions, related to lingual contacts in 1) the alveolar region and 2) the palatal region, respectively. These results were consistent for both subjects, and provided evidence for the hypothesis that the tongue tip/blade and the tongue dorsum are two independently controllable articulators. In a second step, potential applications of this empirical model to studies of coarticulation and timing are illustrated, through a pilot investigation of /kl/ clusters. Finally, implications of our work for a comprehensive model of tongue movements are discussed. Nguyen et al.Modeling tongue-palate contact patterns 2

Dimensionality reduction of electropalatographic data using latent variable models

“…Adaptive dimensionality reduction for EPG data has been investigated using both linear systems and neural networks (Nguyen et al, 1996;Holst et al, 1995). Linear approaches have been based on a rotated principal components analysis (referred to as factor analysis in Nguyen et al, 1996).…”

“…Potentially the tongue has many degrees of freedom owing to its lack of skeleton (Stone, 1991); however, a number of studies suggest that tongue movements in speech may be appropriately modelled using a few elementary articulatory parameters (e.g., (Nguyen et al, 1994(Nguyen et al, , 1996 and references therein). Spatial redundancy arises due to the limited number of possible tongue con®gurations ± only a tiny fraction of the potential 2 62 % 4X6 Â 10 18 EPG patterns can be produced by the articulatory system.…”

Lingual kinematics and coordination in speech‐disordered children exhibiting differentiated versus undifferentiated lingual gestures

“…Classification of UGs requires the identification of a high amount of tongue-to-palate contact occurring simultaneously in both the anterior and posterior regions of the palate or extended across the entire palate during singleton lingual consonant productions (Gibbon 1999b). In normal, mature speech production, the tongue tip and tongue body are recognized as functioning as two quasi-independent articulators (Hardcastle 1976, Nguyen et al 1996, Farnetani 1997, Gibbon 1999a, which are capable of producing distinct, well-defined patterns on the palate. UGs, therefore, are 'interpreted as reflecting a speech motor constraint involving either delayed or deviant control of [the] functionally independent regions of the tongue' (Gibbon 1999b: 382).…”

Experimental evaluation of latent variable models for dimensionality reduction

“…This suggests that the intrinsic dimensionality of the EPG data may be substantially smaller than that suggested (5 to 10) by other studies (e.g. (Nguyen et al, 1996)).…”

“…A number of studies suggest that tongue movements in speech may be appropriately modelled using a few elementary articulatory parameters (e.g. (Nguyen et al, 1996)). In this paper, we consider dimensionality reduction at the spatial level only.…”