基于MMSE-MLSA与感知滤波的语音增强算法

Speech Enhancement Algorithm Based on MMSE-MLSA and Perceptual Filtering

在语音通信过程中,纯净的语音信号可能受到各种不同类型的干扰噪声信号的影响,例如白噪声、色噪声等。针对常见语音增强算法在低信噪比的复杂噪声环境下语音增强后存在语音失真及残余噪声的问题,提出了一种结合改进对数谱幅度的最小均方误差(MMSE-MLSA)谱估计与感知滤波的语音增强算法。该算法采用MMSE-MLSA对含噪语音作初级谱估计增强处理,使用次级感知滤波器进一步掩蔽初级增强信号中的残余音乐噪声。仿真实验结果表明,在低信噪比的复杂噪声环境下,该算法能有效降低语音失真及去除残余音乐噪声,与另外两种语音增强算法比较,增强效果更加突出。

In the process of speech communication,pure speech signal may be affected by various types of interference noise signals,such as white noise,color noise,etc.In order to solve the problem of speech distortion and residual noise in speech enhancement algorithms under complex noise environment with low SNR,a speech enhancement algorithm based on minimum mean square error-modified log-spectral amplitude(MMSE-MLSA)spectrum estimation and perceptual filtering is proposed.The noisy speech is enhanced by MMSE-MLSA as primary spectrum estimation,then a secondary perceptual filter is used to mask the residual music noise after primary enhancement.The simulation shows that the proposed algorithm can effectively reduce speech distortion and remove residual music noise in complex noise environment with low SNR.Compared with the other two algorithms of speech enhancement,the enhancement effect of the method proposed is more prominent.