Pitch prosodic feature for speaker verification

High-level information, such as speaking rate, idiolect, artificial feature, though proved useful for speaker verification and complementary to short-term spectral features, is difficult to extract from speech. Pitch is the frequency of vibration of vocal cord and pitch contour is its variation as a function of time, conveying much speaker specific information. We propose a method to exact pitch prosodic feature (PPF) from pitch contour by means of wavelet analysis. Approximation coefficients and detail coefficients from multiscale analysis with Harr wavelet compose prosodic feature, which describes both general shape and dynamics of pitch contour. Experimental results on NIST06 SRE 8side-1side task show that we can achieve an equal error rate (EER) of 18.42% with PPF, a 23.66% relative reduction comparing with pitch baseline system. Simple fusion of the prosodic system with the MFCC system can further reduce the EER by 5% and minimum DCF by 7% comparing with single MFCC system.