Feature extraction of articulatory motion based on the speech production model

Abstract

This paper deals with two problems that exist in the transformation from an acoustic signal to the articulatory motion. The first concerns the description of articulatory motion during the utterance of voiced sounds. The configuration of the articulatory mechanism is expressed using a rather small number of articulatory parameters, and assuming simple dynamics. It is shown that articulatory movements can be produced by the step inputs for the system, and, since the coarticulation effect is reduced in the stepwise movement of the fictitious target values for each articulatory organ, vowel discrimination can be performed more easily by using these values. The second problem concerns the production model for consonants. Based on this model, the precise vocal tract shape and the position of the noise sound sources can be estimated. The estimation of the vocal tract shape, sound source, and articulatory motion using an articulatory model is a kind of inverse problem of articulatory to acoustic transformation, and requires a nonlinear optimization techinque for its solution.