The Morphic Generator Grammatical Inference Methodology and Multilayer Perceptrons: A Hybrid Approach to Acoustic Modeling

A new hybrid approach to acoustic modeling based on the use of a grammatical inference technique to infer the structure of the models of the sublexical units and an artificial neural network as a means to estimate the emission probabilities of such models is presented. The chosen grammatical inference technique is the so-called “morphic generator grammatical inference” methodology and the connectionist model is the multilayer perceptron. The results on a continuous speech recognition task are better than those obtained by other systems such as semi-continuous hidden Markov models and semi-continuous stochastic grammars. The system even performs slightly better than other hybrid approaches.