Speaker recognition by projection to speaker complementary space

摘要

In speaker recognition, it is a problem that variation of speech features is caused by sentences and time difference. If phonetic information and speaker information included in speech data can be separated, robust speaker recognition will be realized. We have already proposed a speaker recognition method using a subspace method based on principal component analysis in order to extract only speaker information included in speech data. Speaker eigenspace constructed by the subspace method represents a speaker information, but does not always represent difference from other speakers. In this study, we propose a speaker verification method that trains speaker models by VQ or GMM in the complementary space which can represent difference from other speakers. We carried out comparative experiments between the proposed method and conventional VQ and GMM trained in an observation space. As a result, it is clarified that the proposed method is effective in a case of a few codebooks or mixture densities in speaker complementary space.