Noise robustness of speaker adapted recognition system

摘要

This paper deals with a new isolated word recognition system based on a speaker adaptation method. This system is made up of HMMs and an LVQ with information feedback. It has been found that a subword-unit-based isolated word recognizer for identifying words is capable of attaining high recognition accuracies in speaker-dependent applications. The performance of the recognizer, however, deteriorates considerably in speaker-independent applications or in noisy environments. So an LVQ-HMM isolated word recognizer with information feedback has been developed and its ability in speaker-independent applications and in noisy environments has been investigated by isolated word recognition experiments using a number of Japanese words uttered by several native male speakers. The results of the experiments showed that the performance of the new recognizer in quasi-periodic noise is improved.