Dereverberation for Speaker Identification in Meeting

Abstract

Current state-of-the-art speaker identification is a well-established research problem but reverberation is still a major issue used in real meeting scenarios. Dereverberation is essential for many applications such as speaker identification and speech recognition to improve the quality and intelligibility of speech signal interrupted by real reverberation environments. The classical approaches were focused on estimating desired speech signal with dereverberation by beamforming which is crucial for hands-free distant-speech interaction. Its performance degradation is caused when beamforming equipment is unable to comply with the restriction of being symmetric in time or synchronous in structure under real condition. In this paper, a new de-reverberated merging feature is presented for text-independent speaker identification issue applied as an important component of Multiple Distance Microphones (MDM) system used in real meeting scenario. This scenario poses new challenges: farfield, limited and short training and test data, and almost severe reverberation. To tackle this, we introduce a dimensionality reduction approach to extract informative low-dimension features from four kinds of MDM-based features. Experimental results on the MDM system processed reverberated signal show the effectiveness of the new approach and the presented performance evaluation demonstrates the robustness and effectiveness of the proposed approach with short test utterances.