Acoustic-to-articulatory Inversion Using Dynamical and Phonological Constraints

A well-known difficulty in using the articulatory representation for applications in the areas of speech coding, synthesis and recognition is the poor accuracy in the estimation of the articulatory parameters from the acoustic signal of speech. The difficulty is especially serious for most classes of consonantal sounds. This paper presents a statistical method of estimating the articulatory trajectories from the speech signal based on training databases of articulatory-acoustic parameters obtained from continuous speech utterances. The estimation of articulatory trajectories uses the extended Kalman filtering (EKF) technique and is based on new linguistic constraints imposed to acoustic-to-articulatory inversion. These new constraints are mainly implemented by dividing the whole articulatory-acoustic function into a number of phonological sub-functions, each corresponding to a unit of speech defined as the patterns of the continuous transition between two consecutive phonemes. The articulatory-acoustic sub-function is a part of the state-space model that represents each phonological unit of speech. A method of segmenting the speech signal and recognizing the phonological units was developed based on likelihood computation from Kalman filtering with different models. The final estimation of articulatory trajectories is obtained from Kalman smoother using the parameters of the recognized models. Estimation results compared to articulographic and X-ray speech data are presented in this paper. Average RMS errors of about 2 mm have been obtained between estimated and actual articulatory trajectories.