Multiresolution Analysis for Speech Recognition
Résumé
In the purpose to deal with artifact on observations measurements resulting from usual speech processing, we propose to extend the representation of the speech signal by taking a sequence of sets of observations instead of a simple sequence of observations. A set of observations is computed from temporal Multi-Resolution (MR) analysis. This method is designed to be adapted to any usual mode and technique of analysis. Its originality is to take into account two main variations in the analysis, -the center of the frame and -the duration of the frame. In speech processing, multi-resolution analysis has many applications. MR analysis is a basic representation -to locate the stationary and non-stationary parts of speech from the inertia computation, -to select the best representative observation from centroid or generalized centroid. Preliminary experiments are presented. The first one consists in the MR analysis of pieces of the French and the English-American speech databases (i.e., TIMIT, BREF80) and on the inertia as a criterion of location of stationary and non-stationary parts of the speech signal. The second one is on the computation of the phoneme prototypes of the two speech databases. At last, some perspectives are discussed.