关键词: 语音增强, 多带谱熵估计, 多窗谱估计, 相关加权, 分帧相减

Abstract: Multi-spectral entropy not only has the same frequency characteristics with the spectral entropy, but also can show the distribution of energy. Therefore, while conducting the voice detection, it tends to use the multi-spectral entropy to estimate. This paper proves that using multi-spectral entropy to estimate has the advantage over the spectral entropy in the non-stationary signal detection, and properly selects the multi-spectral entropy noise estimation algorithm which is suitable for tank environment. Combining multi-spectral entropy estimation, related weighted, sub-frame subtraction and other theories, it puts forward an algorithm of speech enhancement, based on multiple-window spectrum estimation. Simulation results show that the algorithm can not only suppress the background and musical noises better, but also keep a good intelligibility and naturalness of the voice.

Key words: speech enhancement, multi-spectral entropy estimation, multi-window spectral estimation, correlative weighting, sub-frame subtraction