Neural Network Based Nonlinear Discriminant Analysis for Speech Recognition

Neural networks have been one of the most successful recognition models for automatic speech recognition systems because of their high discriminative power and adaptive learning. In many speech recognition tasks, especially for discrete speech classification, it has been shown that neural networks are very powerful for classifying short-time acoustic-phonetic units, such as individual phonemes. Moreover, neural networks have a strong ability for dimensionality reduction. In contrast to many linear dimensionality reduction techniques including Principal Components Analysis (PCA) and Linear Discriminant Analysis (LDA), neural network based nonlinear reduction approaches are able to form a dimensionally-reduced representation for complex data such as speech features, while preserving variability and discriminability of the original data. In this paper, a neural network is combined with Hidden Markov Models (HMMs) for a continuous phonetic speech recognition system, in which the neural network is trained with phonetic labeling information as a classifier to maximize discrimination among speech features for the speech recognition based on HMMs. Additionally, the dimensionality of speech features is reduced by the neural network with the goal of creating a compact set of highly discriminative features for accurate speech recognition. Experimental evaluation using the TIMIT database shows that the combination of neural networks and HMMs is quite effective for improving recognition accuracy.