非因果先验信噪比估计的LSA算法改进

摘要

For most speech enhancement process, the accuracy of the a priori SNR and the background noise estimation has great influence. This paper derives traditional MMSE-LSA algorithm based on noncausal a priori SNR estimation. In contrast to the LSA algorithm, it introduces noncausal estimation for the a priori SNR, and utilizes smoothing parameters to change the value of noise spectral which can adaptively adjust to the noise environment. Experimental results show the new algorithm yields a higher improvement in the segmental SNR, lower residual noise and betterenhancement of speech quality.%对于大多数的语音增强算法,先验信噪比及背景噪音频谱估计的准确与否,对语音增强的效果影响至关重要.为此,在传统MMSE-LSA算法的基础上,提出一种基于非因果先验信噪比估计的LSA 改进算法,较好地弥补了传统 LSA 算法在先验信噪比上估计的不足,同时采用平滑系数动态更新噪音频谱值,使估计值能更好地跟踪噪音的变化.实验结果表明,改进算法能有效减少残余噪音量,提高语音分段信噪比,改善语音质量.