语音增强：使用burg谱先验信噪比估计消除“音乐噪声”

Speech Enhancement:Using Burg-Based A Priori SNR Estimator to Eliminate Musical Noise Phenomenon

针对谱衰减增强算法中存在的"音乐噪声"问题,提出一种新的估计先验信噪比的方法。文章深入分析了谱衰减算法中"音乐噪声"的产生原因和抑制机理,并基于对先验信噪比的平滑,准确性和分辨率三者之间的关系以及它们对增强算法的影响等方面的综合考虑,选择burg方法结合直接判决准则进行估计,从而得到更加接近于真实情况下的先验信噪比。实验证明,在最小均方误差(MMSE)和对数谱最小均方误差(LSA-MMSE)增强算法中使用文章提出的先验信噪比估计,可以有效抑制"音乐噪声",同时,语音失真也进一步降低。

This paper presents a novel a priori SNR estimator for speech enhancement,which is applied to eliminate the musical noise phenomenon remaining in the results enhanced by spectral attenuation noise suppressor. For the purpose to find a proper estimator closest to the ideal one ,we have studied the cause of musical noise and its suppression mechanism. With emphasis on estimator properties of smoothness, accuracy and resolution, the burg-based method is built on the research of their relationships and influence on de-noising. The experimental demonstrated that： if the proposed estimator is incorporated into the noise suppressor using minimum meansquare error （MMSE） method or log-spectral minimum mean-square error （LSA-MMSE） method, better results will be achieved with the musical noise greatly eliminated and distortion decreased further.