Evaluation of Hands-Free Large Vocabulary Continuous Speech Recognition by Blind Dereverberation Based on Spectral Subtraction by Multi-channel LMS Algorithm

Abstract

Previously, Wang et al. [1] proposed a blind dereverberation method based on spectral subtraction using a multi-channel least mean squares (MCLMS) algorithm for distant-talking speech recognition. Preliminary experiments showed that this method is effective for isolated word recognition in a reverberant environment. However, robustness and effect factors of the dereverberation method based on spectral subtraction were not investigated. In this paper, we analyze the effect factors of compensation parameter estimation for the dereverberation method based on spectral subtraction, such as the number of channels (the number of microphones), the length of reverberation to be suppressed, and the length of the utterance used for parameter estimation, and evaluate these on large vocabulary continuous speech recognition (LVCSR). We conducted speech recognition experiments on a distorted speech signal simulated by convolving multi-channel impulse responses with clean speech. The proposed method with beamforming achieves a relative word error reduction rate of 19.2% relative to conventional cepstral mean normalization with beamforming for LVCSR. The experimental results also show that our proposed method is robust in a variety of reverberant environments for both isolated and continuous speech recognition and under various parameter estimation conditions.