Odyssey 2008: The Speaker and Language Recognition Workshop

Stellenbosch, South Africa
January 21-24, 2008

Support Vector Machines Based Text Dependent Speaker Verification Using HMM Supervectors

Chengyu Dong (1), Yuan Dong (1,2), Jing Li (2), Haila Wang (1)

(1) France Telecom R&D Center (Beijing), Beijing, China
(2) Beijing University of Posts and Telecommunications, Beijing, China

Conventional subword based hidden Markov models (HMMs) have proven to be an effective approach for text-dependent speaker verification. The standard training method works by modeling the MAP adapted means of subword HMMs. In this paper, we propose the use of HMM supervectors from the speaker models as features in support vector machines (SVMs) classifier. An HMM supervector is constructed by stacking means of adapted mixture components from all states within HMMs. We present two SVM kernels: linear kernel and dynamic time alignment kernel (DTAK) based on the KL divergence to evaluate the system. In addition, another effective method is proposed to normalize SVM output scores using speaker independent HMM supervectors. Experimental results show that the SVM system with HMM supervectors achieves lower performance than conventional HMM verification system, but their fusion can give a significant improvement.

Full Paper     Presentation (PPT)

Bibliographic reference.  Dong, Chengyu / Dong, Yuan / Li, Jing / Wang, Haila (2008): "Support vector machines based text dependent speaker verification using HMM supervectors", In Odyssey-2008, paper 031.