Text-Independent Speaker Identification Using Residual Feature Extraction Technique

The Mel Frequency Cepstral Coefficients (MFCC) parameters are derived mainly to represent the spectral envelope or formant structure of the vocal tract system. In this paper, a new feature extraction technique WOCOR is proposed to capture the spectro temporal source excitation characteristics embedded in the linear predictive (LP) residual signal. The vocal Source Wavelet Octave Coefficients Of Residues (WOCOR) information contains pitch frequency and phase in the residual signal. WOCOR features are called vocal source feature because they are dependent on the source of the speech namely the pitch being generated by the vocal folds. WOCOR is generated by applying pitch synchronous wavelet transform to the residual signal. Pitch Synchronous wavelet transform is used to capture the spectro temporal characteristics of the excitation signal. Experimental evaluation is carried out on TIMIT database with 630 speakers using Gaussian Mixture Model (GMM) and Naive Bayesian Classifier. Experimental results show that, speaker identification based on GMM modeling out performs Naive Bayesian classifier based speaker identification. Comparatively an increased in speaker identification efficiency of 6.69% is achieved with GMM modeling for WOCOR feature extraction.