改进相关统计模型的语音增强算法

Improved LSA-MMSE algorithm based on relative statistic model

语音和噪声的时频相关特性研究表明,"音乐噪声"区别于语音的一个重要特征是"音乐噪声"谱时频不相关。根据这一特点,在传统先验信噪比估计相关统计模型基础上给出了两点相关性补充假设。在此基础上,通过改进对数谱最小均方误差语音增强(LSA-MMSE)算法中的D-D先验信噪比估计,提出了改进对数谱最小均方误差语音增强算法。仿真实验采用了主观综合评分测度(MOS)和MBSD两种评价机制,实验结果表明,新模型和算法可以有效地抑制"音乐噪声"现象。

The researches on the relativity character of speech and noise in time-frequency domain show that one of the differences between music noise and speech is the irrelevance of music noise in time-frequency domain.From the viewpoint,two additional assumptions are given for the statistic model of priori SNR estimator.Via revising the decision-directed（D-D） estimator in conventional Log-spectral amplitude（LSA） MMSE speech enhance algorithm,a novel LSA-MMSE speech enhance algorithm is put forward.In the experimental simulation,two estimation mechanisms as MOS and MBSD are applied,and the experimental results show that the proposed algorithm offers more pleasant enhanced speech with residual musical noise explicitly suppressed.