基于矢量量化的实时发音人确认系统研究

Real-Time Text-Dependent Speaker Verification Based upon Vector Quantization

Based on high speed DSP TMS32020, this paper describes a real-time text-dependent speaker verification system on a IBM personal computer. A new methodfor speaker verification based upon vector quantization (VQ) is used. In thisapproach, a speaker-based VQ codebook is designed to characterize a particularspeaker saying a particular uttrance. Later, the same uttrance is spoken by aspeaker to te verified and linear prediction analysis is performed to obtainPARCOR coefficient vectors. The verification decision is made by comparingthese vectors' average quantization distortion with a prespecificd codebook-specificthreshold. In the preliminary tests, 96 percent verification accuracy with only6.8 percent false acceptance rate is achieved. With the application of DSPTMS32020, only a half socond verification time is needed. In this paper, anapproach to specify the verification threshold is described and a method to speedup the VQ codebook training on TMS32020 is proposed. Finally, some commentson this verification system are made and several ways to improved the perform-ance of this verification system are also discussed.