Feature Extraction from Temporal Phase for Speaker Recognition

Feature extraction is important for pattern recognition problems in speech research. Most of the methods of feature extraction primarily exploit spectral information than the phase information. Even though phase is an important characteristics of the speech signal, its use is not much exploited. In this work, in addition to state-of-the-art Mel Frequency Cepstral Coefficients (MFCC), we use features derived from the temporal phase (i.e., T-Phase) of the speech signal for speaker recognition application. The proposed method extracts Linear Prediction Coefficients (LPC) from T-Phase of the speech signal at the frame-level. Experiments are carried on standard NIST 2002 Speaker Recognition Evaluation (SRE) using standard Gaussian Mixture Model - Universal Background Model (GMM-UBM) system. It is observed that the score-level fusion of MFCC and T-Phase feature sets gives 76.18 % identification rate which is a 4% and 8% improvement than MFCC and LPC alone, respectively. In addition, experiments show that score-level fusion reduces the % Equal Error Rate (EER) by 2% and 4% than MFCC and LPC alone, respectively.