Combining Neural Networks and Hidden Markov Models for Continuous Speech Recognition

Abstract

We present a speaker-independent, continuous-speech recognition system based on a hybrid multilayer perceptron (MLP)/hidden Markov model (HMM). The system combines the advantages of both approaches by using MLPs to estimate the state-dependent observation probabilities of an HMM. New MLP architectures and training procedures are presented that 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 recognition accuracy and in number of parameters required.