Second International Conference on Spoken Language Processing (ICSLP'92)

Banff, Alberta, Canada
October 13-16, 1992

Hybrid Neural Network/Hidden Markov Model Continuous-Speech Recognition

Michael Cohen (1), Horacio Franco (1), Nelson Morgan (2), David Rumelhart (3), Victor Abrash (1)

(1) Speech Research Program, SRI International, Menlo Park, CA, USA
(2) Intl. Computer Science Inst, Berkeley, CA, USA
(3) Stanford University, Dept. of Psychology, Stanford, CA, USA

In this paper we present a hybrid multilayer perceptron (MLP)/hidden Markov model (HMM) speaker-independent continuous-speech recognition system, in which the advantages of both approaches are combined by using MLPs to estimate the state-dependent observation probabilities of an HMM. New MLP architectures and training procedures are presented which allow the modeling of multiple distributions for phonetic classes and context-dependent phonetic classes. Comparisons with a pure HMM system illustrate advantages of the hybrid approach both in terms of recognition accuracy and number of parameters required.

Full Paper

Bibliographic reference.  Cohen, Michael / Franco, Horacio / Morgan, Nelson / Rumelhart, David / Abrash, Victor (1992): "Hybrid neural network/hidden Markov model continuous-speech recognition", In ICSLP-1992, 915-918.