(A) hybrid HMM/RBFN system for automatic speech recognition

Abstract

Research in automatic speech recognition (ASR) has been investigated for almost four decades. Especially, the methodology of hidden Markov models (HMMs) proposed in early 1980s, and artificial neural networks (ANNs) introduced in the late 1980s, have become major tools for speech recognition. The HMMs are widely used due to a natural and highly reliable way of recognizing speech for a wide range of applications. However the models have some limitations from the strong constraints: 1) the non-discriminative training/decoding criterion such as the maximum likelihood measure, 2) arbitrary assumptions on the parametric form of probability distribution, and 3) high sensitivity to environmental conditions due to the lack of generalization. On the other hand, connectionist architectures have recently been recognized as an alternative tool for pattern classification and speech recognition problems. Their main properties are their discriminant power and capability to learn and represent implicit knowledge. However, neural networks to date are still not well-suited for dealing with time-varying input patterns and segmentation of sequential inputs. In this sense, an idea of combining HMM and ANNs within a single, novel model, broadly known as hybrid HMM/ANN, has been suggested. This thesis suggest a work in this direction, are new method of combining HMM and radial basis function network (RBFN), one of regression models in ANNs. The suggested model is trained to estimate the posterior emission probabilities of HMM states for the given an acoustic observation sequence. For the training of this model, a two-phase discriminative training technique is used to optimize the RBF parameters. In the first phase of training, the hybrid system is trained by the RLS algorithm, optimizing a mean squared error (MSE) function for the desired and actual posterior emission probabilities. This results in an approximation of posterior emission probabilities of HMM states. In the second phas...