Prosody And Speech Recognition: Experiments And Perspectives

This paper concerns the study of information derived from the melodic, temporal and intensity characteristics of the material to be recognized in a speech recognition system, in French. One classical method for automatic prosodic analysis consists of three steps : parametrization, normalisation of the raw data taking into account the identity of the segments, and perception, and the application of suprasegmental rules. Part I concerns the intrinsic and co-intrinsic variations. Are all these microfluctuations detectable by automatic techniques, and if they are, can they be used to improve an acoustic-phonetic decoding process ? Our study shows that only few phenomena can be significantly observed by means of these techniques on large isolated French words corpora. The reliable phenomena have been successfully integrated into a lexical access system. Part II concerns the applications of suprasegmental rules to speech recognition. An automatic correlative system enabling experts to perf..