Speaker Independent Recognition of Chinese Number Speeches Based on Hidden Markov Model

Abstract

In this paper, we describe the speaker independent speech recognition of Chinese number speeches 0~9 based on HMM. 560 speech samples are recorded and processed. The results of inside and outside testing achieve 92.5% and 76.79%, respectively. Furthermore, to improve the performance, two important features of speech; MFCC and cluster number of vector quantification, are unified and evaluated with various values. The best performance achieve 96%and 81% on MFCC Number = 20 and VQ clustering number = 64.