基于统计推断和矢量量化的非特定文本话者识别方法

AN APPROACH FOR TEXT-INDEPENDENT SPEAKER IDENTIFICATION BASED ON STATISTICAL INFERENCE AND VECTOR QUANTIZATION TECHNIQUE

本文提出了一个基于统计推断和矢量量化技术的非特定文本的话者识别方法，给出了基于矢量量化技术的话者识别方法的统计依据，分析了测试语音样本量对系统正确识别率的影响，并给出了定量计算特定语音待征向量判别能力的公式．介绍了利用以上方法所实现的TISI系统及其实验结果．实验结果表明，在50人的话者集及测试语音长度大于60g的情况下，该系统的正确识别率达99％.

This paper presents an approach for text--independent speaker identification based onstatistical inference and vector quantization technique, introduces the system TISI (Text--IndependentSpeaker Identification) in which proposed approach is implemented. The statistics basis of the TISIbased on vector quantization is discussed, the effect of the sample quantity of testing speech on thecorrect rate is analyzed, and the formula for calculating the discriminatory ability of each componentof feature vector is given. First, for each of the registered speakers, the speaker's acoustic featurevector is extracted from training speech by Perceptually based Linear Predictive (PLP) analysis,then, The LBG algorithm for the VQ is used to get the clusters in the feature set and the statisticalcharacteristics of each code word is calculated. Finally, the method of statistical inference is used todetect the speaker identity. The experiments have shown that the lower order parts of the PLPcoefficients have more speaker individual information compared with that of higher order parts. Foraround 60 seconds testing sessions within 50 speakers, the correct rate of speaker identificationcloses to 99 %.