Vector Quantization Based Classification and Maximum Likelihood Decoding for Speaker Recognition†

Abstract

In a VQ-based speaker recognition system, upon selecting a set of feature parameters that are useful for speaker characterization (such as short-term spectral parameters, pitch, gain, etc.), a vector quantizer is designed for each speaker. Each vector quantizer will then be used to separately encode the sequence of feature vectors generated by an unknown speaker. The speaker whose associate codebook results in the smallest cumulative distortion will then be selected. Previous studies [l]-[4] have shown that not all speech segments are equally effective for the task of speaker recognition. Broad phonetic classes (such as vowels and fricatives) and explicit phoneme classification schemes were proposed in [3], [4]. In these schemes, speech segments were classified into various phonetic categories. A weighted distortion measure based on these phonetic categories was used to identify the unknown speaker. In Section 72.2, we describe a classification scheme based on VQ.