Multichannel speech enhancement algorithm based on hybrid reverberation model

Abstract

To solve the speech enhancement problem in reverberation and noise scenarios, a new speech enhancement model was constructed integrating multichannel linear prediction model and spatial coherence model, and then a multichannel speech enhancement algorithm based on a hybrid reverberation model was designed. The post-reverberation was divided into two components, which were modeled using a multichannel linear prediction model and a spatial coherence model, respectively. To optimize the model parameters, a Kalman filter was used to update the model parameters and polynomial matrix eigenvalue decomposition was used for spatial, temporal, and frequency decorrelation to achieve reverberation and noise reduction. Experimental results show that the proposed algorithm can enhance speech in high and low-reverberation noise environments, and its enhancement effect is superior to popular speech enhancement algorithms, the performance indicators of speech enhancement, perceptual evaluation of speech quality score (PESQ) value and short-time objective intelligibility (STOI) value, have increased by 30% and 20%, respectively.

Keywords

• multichannel speech enhancement
• Kalman filter
• polynomial matrix eigenvalue decomposition