Optimizing Features Extraction Parameters for Speaker Verification

In this paper the role of the frame length on the computation of Mel Frequency Cepstral Coefficients (MFCCs) is investigated in a text-dependent speaker verification system. The variations of vocal characteristics of subjects along the time and the related information conveyed in the MFCCs cause a significant degradation on verification performance. In our experiments we tested the use of different frame lengths for feature extraction in the training and in the verification phases, for a set of speakers whose speech productions were spanned over approximately 3 months. Results show that a suitable choice of the frame lengths combination for training and testing phases can improve performance. The approach shows its potentialities up to 40% in ER reduction for female speakers and up to 58% for the male subset. Key-Words: Speaker Verification, Text Dependent, Mismatch, Frame Length, CD-HMM