结合听觉掩蔽效应的时频自适应小波阈值增强  

Speech enhancement using time-frequency wavelet threshold with auditory masking

作　　者：丁卫[1] 王忠[1] 

机构地区：[1]四川大学电气信息学院,四川成都610064

出　　处：《计算机工程与设计》2011年第11期3768-3771,3856,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(49901013);中国博士后基金项目(2005038615);四川科技厅应用基础基金项目(2008JY0043-1)

摘　　要：针对浊音、清音和噪声的不同特性,结合听觉掩蔽并使用随尺度变化的多阈值对语音信号进行处理。提出了多小波门限估计法,该方法针对不同声音成分,使用不同的与尺度有关的缩小因子调节门限值;通过估计各频带内的信噪比,实现了阈值的时频自适应变化;用巴克小波包分解法模拟人耳临界带特性,用小波谱减法对带噪语音进行预增强,采用Johnston模型推导遮蔽阈值,并推导小波系数增益因子,用其调节时频自变化门限,即用听觉掩蔽效应进一步调节阈值。实验结果表明,提出的方法降低了噪声干扰,并减少了语音失真。For the different characteristics of voiced speech, unvoiced speech and noise, a multi-threshold with auditory masking dependent on the wavelet packet scale is used. Firstly, the multi-threshold estimation based on wavelet is proposed, combined with a variety of methods on threshold estimation changing with the scale. This threshold is adjusted by different reduction factor related to the scale, according to the different components of voice. Secondly, by estimating the signal to noise ratio in different frequency band, the time-frequency adaptive threshold is achieved. Finally, critical band characteristics of the human ear are simulated using bark wavelet packet decomposition. At first, the noisy speech is enhanced using spectral subtraction based on wavelet coefficients. The masking threshold is derived using Johnston model, and the wavelet coefficients gain factor is derived using this masking threshold. Finally, the time-frequency threshold is adjusted by the gain factor. Experimental results show that this method reduces the noise interference, voice distortion is also reduced.

关 键 词：语音增强 Bark小波包变换 噪声估计 小波阈值处理 掩蔽效应 

分 类 号：TN912.35[电子电信—通信与信息系统]