A Speech Understanding System with Learning Capability

Abstract

A speech understanding system with learning capabilities is presented. Its relevant aspecis are: a) The spoken senience is represented concisely by a description that can be used to reconstruct the sentence and to verify whether its meaning was not degraded by the coding. b) Syllables or broader coarticulation segments are the smallest units. c) The evaluation of an hypothesis is based on the probability that a syllable. a word or the sentence can generate the spectrogram of the spoken message and that a syntactic structure can generate its witch contour. d) Coarticulation effects are described in terms of pattern grammers, generating all the possible formant irajectories for a given utterence. e) Spectral and prosodic features can be learned by inference of stochastic finite-state-automats. f) More formant choices are allowed for a single syllable segment and an algorithm is proved for assigning to each choice a probability.