5th International Conference on Spoken Language Processing

Sydney, Australia
November 30 - December 4, 1998

Speaker Verification Using Fundamental Frequency

Yoik Cheng, Hong C. Leung

The Chinese University of Hong Kong, China

This paper describes the use of speech fundamental frequency (F0) for speaker verification. Both Chinese and English have been included in this study, with Chinese representing a tonal language and English representing a non-tonal language. A HMM-based speaker verification system has been developed, using features based on cepstral coefficients and the F0 contour. Four different techniques have been investigated in our experiments on the YOHO database and a Chinese speech database similar to YOHO. It has been found that the pitch information results in a reduction of the equal error rates by 40.5% and 33.9% in Cantonese and English, respectively, suggesting that the pitch information is important for speaker verification and that it is more important for tonal languages. We have also found that the pitch information is even more effective when it is represented in the log domain, resulting in an ERR of 2.28% for Cantonese. This ERR corresponds to a reduction of the ERR by 54%.

Full Paper

Bibliographic reference.  Cheng, Yoik / Leung, Hong C. (1998): "Speaker verification using fundamental frequency", In ICSLP-1998, paper 0228.