Speech Signal Classification with Hybrid Systems

This paper gives a brief overview on two successful hybrid approaches that combine artificial neural networks and Hidden Markov Models for speech signal classification tasks. At first, a short description of traditional stochastic-based Hidden Markov Model speech recognizers with different kinds of emission probabilities are given. The first proposed hybrid approach uses a neural network that approximates arbitrary emission densities in a model-free way. The second hybrid system uses discrete models and a neural network that is trained to work as optimal vector quantizer. The paper compares both systems and integrates them in the traditional stochastic model framework. Speech recognition results are given for the speaker-independent continuous speech ARPA resource management database.