Usefulness of Residual-based Features in Speaker Verification and Their Combination Way With Linear Prediction Coefficients

摘要

This thesis focuses on usefulness of the LPC-residue and LPC coefficients in the speaker verification system. First step, in the front-end, feature extraction get the magnitude spectrum of the speech signal from a 32ms short-time segment of speech that is pre-emphasized and processed by a mel-scale filterbank. And the output of the filterbank are then cosine-transformed to produce the cepstral coefficients. After the coefficients have gotten, they are passed to a Gaussian mixture model (GMM). The GMM is used to represent the claimed speaker's acoustic classes. GMM will produce a maximum-likelihood value. If the value is greater than a predefined threshold, the claimed speaker is accepted. The input of our proposed system have two elements; one is the original speech, and the other is the residual signal. In our study, we create a new feature vector. It is composed of the cepstral coefficients (denoted as LPCC), derived from the LPC, and the MFCC of the residual signal. We find that this new feature vector perform the best comparing to the LPCC and residual-MFCC.