Optimized joint noise suppression and dereverberation based on blind signal extraction for hands-free speech recognition system

In this paper, we build a method based on frequency-domain blind signal extraction (FD-BSE) to jointly suppress diffuse background noise and late reverberation for a hands-free automatic speech recognition (ASR) system. The proposed method utilizes FD-BSE to extract speech signal from the noisy mixture, and a multichannel postprocessing filter, such as a Wiener filter or generalized minimum mean-square error short-time spectral amplitude estimator, for dereverberation. We also apply parameter optimization that maximize the likelihood of the acoustic model of ASR. Experimental results confirm that the proposed method can improve the word recognition accuracy by up to approximately 17% compared with FD-BSE alone under a signal-to-noise ratio of 10 dB.