Feature extraction using wavelet packet decomposition based on MPEG-I

This paper presents a method of features extraction in speaker recognition. This method divides voice signal into two parts based on the MPEG Psychological Model I, and processes them respectively. And then we analyze the extracted MFCC parameters, and compare the partition of frequency-band between Mel-spaced filter group and wavelet packet decomposition. We extract the coefficients by wavelet package transform (called ldquoWPTCrdquo) as diagnostic parameters used in the speaker recognition. The results of experiments indicated that it perform well than MFCC while using the WPTC based on MPEG-I.