A blind dereverberation method based on power spectral subtraction (SS) using a multi-channel least mean\r\nsquares algorithm was previously proposed to suppress the reverberant speech without additive noise. The results\r\nof isolated word speech recognition experiments showed that this method achieved significant improvements\r\nover conventional cepstral mean normalization (CMN) in a reverberant environment. In this paper, we propose a\r\nblind dereverberation method based on generalized spectral subtraction (GSS), which has been shown to be\r\neffective for noise reduction, instead of power SS. Furthermore, we extend the missing feature theory (MFT), which\r\nwas initially proposed to enhance the robustness of additive noise, to dereverberation. A one-stage\r\ndereverberation and denoising method based on GSS is presented to simultaneously suppress both the additive\r\nnoise and nonstationary multiplicative noise (reverberation). The proposed dereverberation method based on GSS\r\nwith MFT is evaluated on a large vocabulary continuous speech recognition task. When the additive noise was\r\nabsent, the dereverberation method based on GSS with MFT using only 2 microphones achieves a relative word\r\nerror reduction rate of 11.4 and 32.6% compared to the dereverberation method based on power SS and the\r\nconventional CMN, respectively. For the reverberant and noisy speech, the dereverberation and denoising method\r\nbased on GSS achieves a relative word error reduction rate of 12.8% compared to the conventional CMN with\r\nGSS-based additive noise reduction method. We also analyze the effective factors of the compensation parameter\r\nestimation for the dereverberation method based on SS, such as the number of channels (the number of\r\nmicrophones), the length of reverberation to be suppressed, and the length of the utterance used for parameter\r\nestimation. The experimental results showed that the SS-based method is robust in a variety of reverberant\r\nenvironments for both isolated and continuous speech recognition and under various parameter estimation\r\nconditions.
Loading....