Speech Dereverbration Using Multi-Channel CDR Linear Prediction

In order to avoid the impact of reverberation in enclosed environments on the intelligent speech processing system , an adaptive speech dereverbration method was proposed in this thesis. Given that some values of the covariance matrix constructed by the speech signal on microphone array might be zero, which may cause distortion of the output speech signal. The speech dereverbration using multi-channel linear prediction (MCLP) were adopted in the thesis, based on double-channel coherent-to-diffuse power ratio (CDR) covariance initialization and employing estimated value of the covariance matrix of the array received signals obtained by recursive averaging. The experimental results show that the improved MCLP algorithm is more effective by comparing different methods in different reverberation environments.