Phonetic knowledge embedded in a context sensitive MLP for French speaker-independent speech recognition

Abstract

The stop /p,t,k/ recognition part of a speaker-independent speech recognition system is described in this paper. This work is based on the conclusions of several perceptual experiments and on the results of an acoustic investigation with stop consonants. These experiments allowed us to evaluate the discrimination power of the burst regarding the stop place of articulation, and how the vocalic information may help stop identification, which could not be done efficiently without taking into account the nature of the following vowel. Thus, a novel system architecture is proposed which is made up of two stages: first, an automatic detector of reliable cues regarding stop and vowel features, and, then, a context sensitive multilayered perceptron (ODWE) fed by the previous acoustic cues. The training and the test have been done on two different corpora including male and female speakers. The results show a recognition rate of 90% over the stop consonants.