Short duration voice data speaker recognition system using novel fuzzy vector quantization algorithm

The performance of any speaker recognition system depends on the duration of the speech samples. The higher the number of feature vectors is, the better is the efficiency. A major contribution of this paper is in enhancing the identification accuracy of the speaker recognition system through minimization of the objective function and associated distortions. With nonlinear mapping, the sectional set fuzzy vector quantization with novel norm is utilized here as usual to form speaker's model in the high-dimensional feature space. However, during feature extraction, the traditional triangular shaped bins have been replaced by Gaussian shaped filter (GF) and Tukey filter (TF) for calculating the mel frequency cepstral coefficients (MFCCs). The paper presents experimental evaluation of three modeling techniques, viz. fuzzy c-means, fuzzy vector quantization (FVQ)2 and novel fuzzy vector quantization (NFVQ). On simulation, the NFVQ shows significant improvement in performance over fuzzy c-means and FVQ2. The experimental evidence demonstrates that for two seconds of training and one second of testing data, the efficiency of the NFVQ, with a minimum objective function of Jm = 0.073 and distortion D = 4.334, for a set of 100 speakers chosen from the Texas Instruments and Massachusetts Institute of Technology (TIMIT) database and self-collected database is 98.8% and 98.1%, respectively