This paper describes the introduction of an automatic segmentation algorithm as a pre-processing of a speaker-independent digit recognition system based on hidden Markov models ; so an acoustic analysis is performed over each segment. The observation sequence is composed of the sequence of the provided vectors of acoustic coefficients. The used Markov modelling reposes on a description in two levels each digit is translated in terms of basic units (allophone, phoneme, pseudo-diphone), and each basic unit is developped in terms of acoustic segments. The recognition performances obtained for the allophone based model with different acoustic analysis lead to prove that the segmental analysis doesn't lose essential information in respect with the sliding-block analysis; the comparision between the articulatory interpretation of the segmentation and the best path found by the Viterbi algorithm in the pseudo-diphone based model validates our qualitative analysis of the acoustic segments.
Bibliographic reference. Maire, V. Le / Andre-Obrecht, Régine / Jouvet, D. (1989): "An acoustic-phonetic decoder an automatic segmentation algorithm", In EUROSPEECH-1989, 2392-2395.