A third-order tensor decomposition approach for adaptive speech dereverberation

Reverberation caused by sound reflections in enclosed spaces typically degrades the quality and intelligibility of the speech signal of interest, even leading to remarkable performance deterioration of voice communication and speech recognition systems. The adaptive weighted prediction-error (AWPE) algorithm can effectively suppress the late reverberation component of microphone speech, however, the prediction-error filter has to be long enough to accurately estimate the late reverberation component, which results in a high computational complexity, making it challenging to apply in the real-time systems. To address this issue, this paper proposes an improved method based on the third-order tensor decomposition. The long linear prediction-error filter in the AWPE algorithm is decomposed into three groups of short sub-filters through multiple Kronecker product. The high-dimensional correlation matrix is then converted into a set of low-dimensional correlation matrices, thus significantly reducing the computational complexity. Simulation results validate that the proposed approach not only achieves effective dereverberation performance, but greatly improves the computational efficiency.