This paper investigates some automatic adjustments of the structure of Markov models: the reduction of the model complexity which is achieved by merging similar gaussian functions, and the improvement of the acoustical modelization which relies on the splitting of some gaussian functions and the discarding of unreliable parameters. These modifications are tested on isolated word vocabularies recorded by more than 500 speakers through the telephone network. On a 36-word vocabulary, and with regular word models, a 40 % reduction of the number of gaussians functions is achieved while keeping a similar recognition performance. Furthermore, a detailed analysis of the training phase shows that the merging operator is useful for discarding unreliable parameters. Several dynamic expansion procedures are also described, which lead to a 30 % error rate reduction for two types of basic units. Keywords: Speech recognition, Markov modelling, Merging gaussian functions, Splitting gaussian functions.
Bibliographic reference. Jouvet, D. / Mauuary, L. / Monné, Jean (1991): "Automatic adjustments of the structure of Markov models for speech recognition applications", In EUROSPEECH-1991, 927-930.