提高耳语音可懂度的非对称压缩语音增强方法  

An asymmetric attenuated speech enhancement approach for improving intelligibility of noisy whisper

作　　者：周健[1,2] 郑文明[3] 王青云[2] 赵力[2] 

机构地区：[1]安徽大学计算智能与信号处理教育部重点实验室,合肥230031 [2]东南大学水声信号处理教育部重点实验室,南京210096 [3]东南大学儿童发展与学习科学教育部重点实验室,南京210096

出　　处：《声学学报》2014年第4期501-508,共8页Acta Acustica

基　　金：国家自然科学基金(61301295;61231002;61273266;61003131);安徽省自然科学基金(1308085QF100;1408085MF113);安徽大学博士科研启动经费资助

摘　　要：提出两种基于非对称代价函数的耳语音增强算法,将语音增强过程中的放大失真和压缩失真区分对待。Modified ItakuraSaito(MIS)算法对放大失真给予更多的惩罚,而Kullback-Leibler(KL)算法则对压缩失真给予更多的惩罚。实验结果表明,在低于—6 dB的低信噪比情况中,经MIS算法增强后的耳语音的可懂度相比传统算法有显著提高;而KL算法则获得了同最小均方误差语音增强算法近似的可懂度提高效果,证实了耳语音中的放大失真和压缩失真对于耳语音可懂度的影响并不相同,低信噪比时较大的压缩失真有助于提高耳语音可懂度,而高信噪比时的压缩失真对耳语音可懂度影响较小。Two asymmetric cost function for whispered speech enhancement methods are proposed. The cost of the amplification distortion and the attenuation distortion are different in both methods. The Modified Itakura-Saito （MIS） distance function gives more penalties to speech amplification distortion while the Kullback-Leibler （KL） divergence function gives more penalties to speech attenuation distortion. The experimental results show that the MIS method gains larger intelligibility improvement of the whispered speech than the conventional speech enhancement algorithms in much lower Signal to Noise Ratio （SNR） less than -6 dB, and the KL method has similar intelligibility improvement performance to the Minimum Mean Square Error （MMSE） speech enhancement method. The results confirm that the amplification distortion and the attenuation distortion have different effects on the intelligibility of the enhanced whisper. Specifically, larger attenuation distortion can improve speech intelligibility in lower SNR condition and it has a little influence on speech intelligibility in high SNR condition.

关 键 词：耳语音 可懂度 语音增强 非对称 放大失真 代价函数 噪声谱 信噪比 高斯噪声 最小均方误差 

分 类 号：TN912.35[电子电信—通信与信息系统]