Second European Conference on Speech Communication and Technology

Genova, Italy
September 24-26, 1991


A Parallel HMM Approach to Speech Recognition

F. Brugnara (1), Renato De Mori (2), D. Giuliani (1), M. Omologo (1)

(1) Istituto per la Ricerca Scientifica e Tecnologica, Povo di Trento, Italy
(2) School of Computer Science, McGill University, Montreal, Quebec, Canada

Stochastic signal models represent a powerful way to approach the problem of speech recognition. A particular stochastic modeling, the first order Hidden Markov Model (HMM), has become increasingly popular, because it has a solid theoretical basis and offers practical advantages. In this paper we will extend the standard HMM theory to Parallel Hidden Markov Model (PHMM). The parallel model consists of two statistically related HMMs. This configuration permits a more complete and accurate characterization of the speech signal. In this framework, an observation consists of a couple of acoustic parameter vectors, one for a standard HMM and the other for an HMM whose parameters are probabilistic functions of the state of the first model.

Full Paper

Bibliographic reference.  Brugnara, F. / Mori, Renato De / Giuliani, D. / Omologo, M. (1991): "A parallel HMM approach to speech recognition", In EUROSPEECH-1991, 1103-1106.