Efficient acoustic modeling of HMM speech recognizer by subvector quantization method

Abstract

In the last decade, many researches have investigated automatic speech recognition techniques. As a result of these efforts, the fast, robust and effective speech recognition systems have been developed in the desktop environment. Among several techniques, Hidden Markov Model (HMM) shows a highly efficient recognition capability and provides a possibility that a voice user interface can be put into practical use. However, it hard to directly reproduce the algorithms suitable for the desktop applications onto mobile devices. Though it is possible, most cases show an inefficient performance, which is unacceptable for practical use. To overcome these limitations, several approaches such as computation reduction, voltage modulation, fixed-point arithmetic, alternative training and decoding algorithm and low-memory consumption are proposed. Among those technologies, memory reduction is a crucial issue for resource constrained automatic speech recognition (ASR) because large vocabulary and continuous HMM based ASR systems occupy significant amount of memory to store the parameters. A simple yet effective way to reduce the required resources with little effect on the performance is to quantize parameters. Several techniques have been used to achieve this objective. Scalar quantization simply clusters the individual elements of parameter vectors and sub-vector clustering breaks up vectors into several sub-vectors, allowing the complexity of the search and the storage requirements to be reduced at the cost of an increase in distortion. In most cases, however, only easily recognized knowledge is used in the choice of sub-vectors such as the type of features and the most strongly correlated pairs or higher dimensional subvectors which have the same dimension. This dissertation aims at maintaining the performance of parameter quantized ASR system as good as that of the original HMM based speech recognition system using the subvector quantization method. There are ...