Gender identification from Chinese dialects speech based on semi-supervised vector quantization

This paper describe a novel gender identification system from Chinese dialects. In this system, speech data is quantized by using semi-supervised learning principle and gender codebook models of male and female with supervision information is formed. It can also improve the deficiency of low precision of codebook effectively. For speech data of five Chinese dialects, the recognition accuracy of telephone speech result as high as 95.8%, which raise the rate of correct identification about 6.5% compared with conventional system. While testing data is clean speech, the recognition rate is higher, which could come to 99.3% simultaneously. Experimental results show that the accuracy and stability of the new identification system based on supervision information, are effectively improved compared with the traditional VQ system.