Maximum Likelihood Discriminant Feature for Text-Independent Speaker Verification

Feature extraction is an essential first step in speaker verification applications. In addition to static features extracted from each frame of speech data, it is beneficial to use dynamic features that use information from neighboring frames. In this paper a new feature estimation method based on maximum likelihood discriminant analysis is presented. We compare it to traditional MFCC features in a NIST 2006 SRE core task. Experiments show that the proposed scheme provides more discriminative feature vectors. The features obtained with the new estimation method show a 10% - 15% relative improvement in EER and MinDCF over traditional MFCC features.