Speaker Verification using Weighted Local MFCC Features Extracted by Minimum Verification Error Learning

摘要

Text-independent speaker verification using adaptively weighted Mel Frequency Cepstrum Coefficients (MFCC) over multiple neighboring frames, and Gaussian Mixture for likelihood estimation is introduced. For each registrant, optimal linear weightings of multiple speech frames are searched based on Minimum Verification Error (MVE) learning, generalizing the scheme of the use of ΔMFCC feature which attempts to capture inter-frame characteristics. In the verification experiments, the proposed method was found to improve the verification performance under noisy environments and use via phone line, when compared with the conventional methods.