抽象的な

Optimization of feature vectors for art classifier in language independent speaker recognition system for biometric security

A Jose Albin, NM Nandhitha

Automated speaker recognition from speech signals plays a major role in the field of biometrics to authenticate the speaker. Though considerable research is carried out in this area, sensitivity and specificity of the proposed technique are not satisfactory. In this paper, language independent speaker recognition system using spectral features is proposed to increase the sensitivity of speaker recognition. Flux, short time energy, centroid, pitch, period and number of peaks are extracted from time domain coefficients, autocorrelation coefficients, discrete wavelet coefficients and cepstrum coefficients. Adaptive Resonance Theory (ART) network is used for identifying the speaker from the above features. From the research, it is found that the combination of features provides the best sensitivity for speaker recognition. In addition to the spectral domain features, time domain features also add intelligence to the proposed technique by increasing the accuracy to above 98%.