Abstract: Aiming at the residual musical noise problem which is caused by the amplitude spectrum LOG_MMSE estimation algorithm at low SNR, this paper proposes a new speech enhancement algorithm based on energy entropy statistical model. The approximate solution is derivated. Weighting factors are dynamically adjusted according to the difference between entropy, to further improve the weak speech segment denoising effect. Experimental results show that the algorithm in the SegSNR improvement and the residual musical noise are better than the traditional amplitude spectrum LOG_MMSE algorithm.

Key words: Voice Activity Detection(VAD), energy entropy, MMSE, noise estimation

摘要： 针对幅度谱LOG_MMSE估计算法在低信噪比时的残留音乐噪声问题，提出基于能量熵统计模型的LOG_MMSE估计，推导其近似解。根据熵的差值动态调整加权因子，进一步改善弱语音段的去噪效果。实验结果表明，该算法在改善信噪比和残留音乐噪声方面均优于传统的幅度谱LOG_MMSE算法。

关键词: 语音活动检测, 能量熵, 最小均方误差, 噪声估计

CLC Number: 

• TN912.35