结合人耳听觉感知的两级语音增强算法  

A Two-Stage Speech Enhancement Algorithm Combined with Human Auditory Perception

作　　者：张勇[1,2] 刘轶 刘宏[1] 

机构地区：[1]北京大学信息科学技术学院,北京100871 [2]深港产学研基地深圳市智能媒体和语音重点实验室,深圳518057

出　　处：《信号处理》2014年第4期363-373,共11页Journal of Signal Processing

基　　金：深圳市技术研究开发计划基础研究(JCYJ20120831171439442)

摘　　要：针对现有语音增强算法面临残留噪声这一问题,结合人耳听觉系统的掩蔽特性,本文提出了一种优化的语音增强算法。算法分为两级,第一级利用MMSE-LSA谱估计法对带噪语音进行降噪处理,经过处理后,带噪语音信号的信噪比得到了提高。然后,针对第一级增强语音信号中的残余噪声利用人耳听觉掩蔽特性掩蔽掉。为此,算法结合人耳听觉掩蔽特性设计了感知增强滤波器,该滤波器能够有效去除第一级增强语音信号中的残留噪声。仿真实验表明,在各种复杂背景噪声以及信噪比环境下,经过本文算法处理后的增强语音信号残留噪声明显减少,算法提升了增强语音的主观感知质量。For the residual noise problem faced by existing speech enhancement algorithms,an optimal speech enhancement algorithm was presented,which was combined with the masking properties of human auditory system.The Algorithm consisted of two stages,and the MMSE-LSA spectral estimation method was applied to the noisy speech for noise reduction in the first stage.After the first stage processing,the signal-to-noise ratio of the noisy speech was improved.Then,for the residual noise of the enhancement speech signal processed by the first stage,the masking properties of human auditory system was used in the second stage.A perceptual enhancement filter was designed based on the human auditory perception,which could effectively remove the residual noise.Experimental results under different kinds of background noise and different signal-to-noise ratios demonstrate that the residual noise is significantly reduced and the subjective perception quality of the enhanced speech signal is increased.

关 键 词：语音增强 对数谱幅度最小均方误差 感知增强滤波 

分 类 号：TN912.3[电子电信—通信与信息系统]