Multi-feature combination for speaker recognition

Combination of different features has been proved to be a good method for improving performance in speech recognition. In speaker recognition (SRE), various features have also been developed to reflect complementary aspects of speaker´s characteristics. This paper proposed an effective multi-feature combination in speaker recognition. In order to avoid the “dimensionality disaster” and to delimit the redundant information, linear discriminant analysis (LDA) is used to reduce the high dimensionality of combined feature to be lower. Then feature-domain channel compensation is applied to improve the performance. In experiments, we use the popular short-term spectral Mel-frequency cepstral coefficients (MFCC) and novel spectro-temporal time-frequency cepstrum (TFC) to do feature combination followed by LDA and feature-domain latent factor analysis (fLFA) for channel compensation respectively. The experimental results on the NIST SRE2008 short2 telephone-short3 telephone test set show that the proposed multi-feature combination is an effective method to outperform both raw features.