An investigation of dependencies between frequency components and speaker characteristics based on phoneme mean F-ratio contribution

This paper proposes a new speaker feature extraction method, which is based on the non-uniformly distributed speaker information in frequency bands. In order to emphasize individual differences effectively, in this study, we first examine the differences of the distribution of individual information in the frequency region when a speaker utters different phonemes. Then we adopt an improved F-ratio, a phoneme mean F-ratio, to measure the dependences between frequency components and individual characteristics. According to the result of the analysis, we adopt an adaptive frequency filter to extract more discriminative feature. The new feature was combined with GMM speaker models and applied to the speaker recognition database which includes 50 persons. The experiment shows that the error rate using the proposed feature is reduced by 28.5% compared with the F-ratio feature, and reduced by 68.02% compared with the MFCC feature.