Dereverberantion based on generalized spectral subtraction for distant-talking speaker recognition

A dereverberation method based on generalized spectral subtraction (GSS) using multi-channel least mean square (MCLMS) was proposed previously. The results on speech recognition experiments showed that this method achieved a significant improvement compare to the conventional methods. In this paper, we employ this method to distant-talking speaker recognition. However, the GSS-based dereverberation method using clean speech models degrades the speaker recognition performance while it is very effective for speech recognition. One of the reason may be that the GSS-based dereverberation method causes some distortions such as speaker characteristics distortion between clean speech and dereverberant speech. In this paper, we address this problem by training speaker models using dereverberant speech which is obtained by suppressing reverberation from arbitrary artificial reverberant speech. The speaker recognition experiment was performed on a large scale far-field speech with different reverberant environments to the training environments. The proposed method achieved a relative error reduction rate of 88.2% compared to conventional CMN with beamforming using clean speech models and 32.8% compared to reverberant speech models, respectively.