Geometric articulatory model adapted to the production of consonants

Résumé En

This works deals with the construction of articulatory models which can be easily adapted to a new speaker and enable a better approximation of tongue contours corresponding to consonants. Data used are three corpora of X-ray films. The first corpus was used to construct an articulatory model and design an adaptation procedure. The evaluation carried out on the third corpus shows that this adaptation performs well. Geometric fitting provided by the first model was often insufficient in the region of the consonantal places of articulation of the second corpus. Tongue contours delineated from X-ray images were thus corrected by considering virtual articulatory targets and the weight of consonants was increased in the Principal Component Analysis (PCA). Furthermore, the coefficients of the linear components are not calculated by projecting contours onto the PCA base vectors, but with an optimization procedure so as to guarantee a good approximation in the constriction region of consonants.