September 22-25, 1997
We are working on performing acoustic to articulatory inversion by using Maeda's model. The purpose of this work is to adapt the model to a new speaker. The adaptation quality is assessed by verifying that vowels uttered by the speaker lie inside the vocalic space defined by the model. It is with this aim in view that we realized a series of MR images for eleven oral French vowels (/i, e, , y, , oe, a, , , o, u/). The adaptation may include modifications of: scale factors for the pharynx and the mouth cavity, the wall of the vocal tract and coefficients for the calculation of the area function from a sagittal shape. The scale factors have been determined by superimposing Maeda's model on the MR images. The wall has been obtained by calculating the mean value of the exterior contours of the vocal tract in the image series. As some discrepancies between natural and synthetic vowels remained the wall contour has been iteratively optimized by means of formant sensibility functions calculated for each section in the vocal tract. The inversion is carried out by means of a table-lookup procedure constrained by the smoothness of articulatory trajectories.
Bibliographic reference. Mathieu, Bruno / Laprie, Yves (1997): "Adaptation of Maeda's model for acoustic to articulatory inversion", In EUROSPEECH-1997, 2015-2018.