Gender-dependent feature extraction for speaker recognition

Gender information is believed helpful for speaker recognition. In GMM-UBM based speaker recognition, the use of gender information is mostly on a basis of constructing gender-dependent (GD) UBMs instead of extracting GD features. However, theoretical analysis and experimental observations show that females and males differ quite a lot in the feature domain of speech signal, including F0, formant, spectrum and cepstrum. In this paper, further analysis and experiments have been done to explore the differences between females and males. Afterwards, a GD MFCC feature extraction is proposed. In this method, the frame length of MFCC extraction is gender dependent, in other words the resolution for both the DFT analysis and hence the MFCC feature extraction is gender dependent. Experimental results demonstrate that compared with the gender-independent (GI) feature extraction, the GD feature extraction can achieve relative EER reductions of 21.7% and 12.2% for female and male speakers evaluations, respectively.