Improving the performance of speaker verification systems under noisy conditions using low level features and score level fusion

This paper provides an overview of low-level features for speaker recognition, with an emphasis on the recently proposed MFCC variant based on asymmetric tapers (MFCC asymmetric from now on); which has proven high noise robustness in the context of speaker verification. Using the TIMIT corpus the performance of the MFCC-asymmetric is compared with: the standard Mel-Frequency Cepstral Coefficients (MFCC) and The Linear Frequency Cepstral Coefficients (LFCC) under clean and noisy environments. To simulate real world conditions, the verification phase was tested with two noises (babble and factory) at different Signal-to-Noise Ratios (SNR) issued from NOISEX-92 database. The experimental results showed that MFCCs-asymmetric tapers (k=4) outperform other features in noisy condition. Finally, we have investigated the impact of consolidating evidences from different features by score level fusion. Preliminary results show promising improvement on verification rate with score fusion.