Pitch Synchronous Based Feature Extraction for Noise-Robust Speaker Verification

This paper proposes a pitch synchronous based feature extraction method for speaker verification under noisy conditions. Before feature extraction, a pitch synchronous analysis procedure is employed. Each speech frame of the input utterance is divided into a periodic part and a non-periodic part by pitch synchronous analysis. Because that the nonperiodic part causes harmonic leakage of the speech frame, only the periodic part is preserved for concatenating feature extraction. Two dominant spectral and cepstral analysis methods, MFCC and LPCC, are employed to represent the features of each frame. Speaker verification experiments under noisy conditions are conducted on NTT speech database. The performances of pitch synchronous based MFCC (p-syn MFCC), pitch synchronous based LPCC (psyn LPCC), MFCC and LPCC are compared with each other. Experiments results indicate that the proposed pitch synchronous based feature extraction method is efficient for speaker verification under noisy conditions.