Speech Recognition by Dereverberation Method Based on Multi-channel LMS Algorithm in Noisy Reverberant Environment

1 Introduction In a distant-talking environment, channel distortion drastically degrades speech recognition performance because of mismatches between the training and test environments. The current approaches focusing on robustness issues for automatic speech recognition (ASR) in noisy reverberant environments can be classified as speech enhancement, robust feature extraction, or model adaptation methods. In this paper, we focus on speech enhancement in a distant-talking environment. Previously, Wang et al. [1] proposed a robust distant-talking speech recognition method based on power spectral subtraction (SS) employing the adaptive multi-channel least mean squares (MCLMS) algorithm. In their study, late reverberation was treated as additive noise, and a noise reduction technique based on power SS was proposed to estimate the power spectrum of clean speech using an estimated power spectrum of the impulse response. To estimate the power spectra of the impulse responses, they extended the MCLMS algorithm for identifying impulse responses in a time domain [2] to a frequency domain. We proposed a blind dereverberation method based on generalized SS (GSS), which has been shown to be effective for noise reduction, instead of power SS [3]. The dereverberation method based on GSS with beamforming achieved a relative word error reduction rate of 9.8% and 31.4% compared to the dereverberation method based on power SS with beamforming and the conventional cepstral mean normalization (CMN) with beamforming, respectively. However, both the power SS-based method [1] and GSS-based method [3] were evaluated in a simulated reverberant environment without additive noise. In this paper, we evaluate the blind dereverberation methods in a real noisy reverberant environment with stationary noise or non-stationary noise. To suppress the stationary noise and non-stationary noise, we use a noise reduction technique based on GSS and a blind source separation based on independent component analysis (ICA), respectively. FastICA is one of the most popular algorithm for ICA. Our experimental result shows that combining the Efficient FastICA (EFICA) [4], which is an improved version of FastICA, is effective for our dereverberation method [6]. The schematic diagram of our dereverberation method is shown in Fig. 1. The late reverberation are reduced from the spectrum of multi-channel distorted speech by our dereverberation method using the estimated spectrum of impulse response. Thereafter, the early reverberation is normalized by CMN at the feature extraction stage. 2.1 Dereverberation based on GSS If speech s[t] is corrupted by convolutional noise h[t], the observed speech x[t] becomes x[t] = h[t] * s[t], where …