Adaptive Methods In Speech Recognition Research

The research on Automatic Speech Recognition has seen great achievement in the passed twenty years. Especially, theoretical systems on Digital Speech Signal Analysis and Statistic Model have been fully established. Although many speech recognition systems are available, their poor robustness prevents them from practical use. Hence, adaptive ability becomes a main issue in speech recognition field.This paper concentrates on adaptive method in speech recognition. First, main causes effecting robustness of speech recognition systems are analyzed. Major adaptive methods — Map-based adaptation and Bayesian adaptation — are introduced representatively For many adaptive methods are realized by cepstral vector processing. We call them Cepstral Normalization Adaptation and give a brief introduction. Secondly, we test some of these methods on an experimental system. It is showed that Bayesian adaptation is superior to methods. This method takes all the negative factors into account, but system parameters are optimized for individual model, while not the whole model system.So, we present a new kind of method, Discriminative Adaptation, enlightened by Discriminative Training. Error classification probability is characterized by discriminative function. Parameters are optimized by minimizing error classification probability. An algorithm for our experimental system is developed. It shows that, after adaptive processing, recognition rate gain about five percent for new speaker. Discriminative adaptation can also be used in various speech recognition systems.