On the use of phase information-based joint factor analysis for speaker verification under channel mismatch condition

Recent studies have shown that phase information contains speaker characteristics. A new extraction method to extract pitch synchronous phase information has been proposed and shown that it was very effective under channel matched condition. However, phase changes between different channels. Therefore, the speaker recognition performance is drastically degraded under channel mismatch condition. On the other hand, joint factor analysis (JFA) is an approach that is robust for channel variability. In this paper, we propose phase information-based JFA for speaker verification under channel mismatch condition. Speaker verification experiments were performed using the NIST 2003 SRE database. Phase information-based JFA achieved a relative equal error rate reduction of 20.9% for male and 17.4% for female compared to the traditional system based on Gaussian mixture model and Universal background model (GMM-UBM) that influenced by channel variability. Furthermore, by combining phase information-based method with the MFCC-based method, we obtained the better result than that of the only MFCC-based method.