Robust speaker verification using phase information of speech

Phase information provide useful acoustic cue for identifying individual speakers. Speaker verification employing the instantaneous phase-related features that perform well in clean or matched noise/channel conditions degrades dramatically when encounter unexpected communication environments. These adverse effects can distort the short-term distributions of the speaker parameters. It is observed that by mapping each feature stream to a target distribution over a specific time interval, their robustness to environmental or channel mismatch is enhanced. Through speaker verification experiments on microphone and telephone data, it is observed that the proposed robust feature extraction front-end consistently reduces the equal error rate.