Compensation Approaches for Distant Speaker Identification under Reverberant Environments

Speech recorded in real environments by distant microphones is degraded by factor like reverberation. This degradation strongly affects the performance of the speaker identification system. Three compensation approaches are investigated to improve the robustness of speaker identification in such scenarios. The first approach applies spectral subtraction before feature extraction in order to reduce the late reverberation effect. The second approach makes use of feature warping as robust features of distant speaker identification under mismatched training-testing conditions. The third approach presents a novel GMM parameters initialization method: combination division and k-means clustering. The experiment results show that the compensated system as compared with baseline system, the channel average identification rate has an increase of 11.4%, 15.4%, 17%, 17.8% on TIMIT database and 6.82%, 6.36%, 9.34%, 14% on JNAS database.