A continuous speech recognition system for data base consultation

The continuous speech recognition system, developped at E.N.S.E.R. Grenoble, for a semantically restricted natural language is described. This system uses a sentence verification strategy : a phonetic matching algorithm, near to Viterbi algorithm, computes the similarity between the input signal and several phonetic hypotheses. The grammar of the language is not embedded in the lexical decoding network. The recognition of a not finite state language is allowed. The grammar model is the T.N.F. model generalized for our purposes. The Itakura´s linear predictive residual is used as a speech sound similarity measure. The associated learning system is described. This system builds a spectral template dictionary and the lexical decoding network with the help of a dynamic programming procedure.