TEXT TEXT-DEPENDENT SPEAKER RECOGNITION BY COMPRESSED FEATURE FEATURE-DYNAMIC DYNAMICS DER S DERIV IVED FROM ED SINUSOIDAL REPRESENT REPRESENTATION OF SPEECH

摘要

Prevalent speaker recognition me methods use only spectral thods spectral-envelope based features such as MFCC, ignoring the richspeaker identity information contained in the temporal temporal-spectral dynamics of the entire speech signal. We propose anew feature for speaker recognition based on sinusoidalrepre representation of spe sentation speech ech called compressed spectral d dy- ynamics( namics Sino Sinogram ram-CSD) CSD), which effectively captures suchspectral dynamics and the inherent speaker identity. Thediscriminative pow power of CSD allows er classifica classification tion to r re- emamain simple. The proposed CSD in CSD-MSRI method uses a si sim- mplenearest neigh ple neighbor classifier to deliver bor performance co com- mpetitivepetitive to conventional MFCC+DTW based text text-dependentspeaker recogni recognition methods tion at signif significantly lower icantly co com- mplexity.Plexity.