Apport d'une segmentation statistique du signal de parole dans un système de décodage acoustico phonétique basé sur les connaissances

Abstract
In a speech labelling system, traditionnaly the first stage consists in the transcription of the continuous acoustic signal into a series of discrete elementaries units. This stage called Acoustic-Phonetic decoding unrolls three phasis : parametrisation, segmentation and identification. All the systems do not use the module of segmentation stricktly speaking. For those that use it, they differ from one another in their strategy of segmentation. So we distinguish two different approachs, one is based on an a posteriori segmentation and the other on an a priori one. This paper presents an automatic phoneme recognition system based on a preprocessing segmentation algorithm called the Divergence method. The study of these performances got by Divergence and those obtained by a knowledge based system (SED) where the segmentation is an a posteriori result, has brought us to propose an hybrid system named the SED-Divergence method, using the advantages of the two methods. We asser these systems by integrating them into a labelling process.