基于噪声谱约束的二值掩码估计语音增强算法  

Speech Enhancement Algorithm of Binary Mask Estimation Based on Noise Spectrum Constraints

作　　者：叶琪[1,2] 陶亮[1] 周健[1,2] 王华彬[1] 

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,安徽合肥230031 [2]安徽大学媒体计算研究所,安徽合肥230601

出　　处：《信号处理》2016年第1期70-76,共7页Journal of Signal Processing

基　　金：国家自然科学基金(61301295;61372137);安徽省自然科学基金(1308085QF100);安徽大学博士启动资金资助项目

摘　　要：提出一种基于噪声谱约束的二值掩码估计语音增强算法,用以提高低信噪比情况下的语音可懂度。首先分析了低信噪比时,先验信噪比过估对噪声谱估计函数的影响;再分别对先验信噪比和噪声谱估计函数进行修正;最后,根据修正后的噪声谱估计函数和先验信噪比判断出噪声谱被欠估的时频单元,估计出二值掩码值,并对相应的增强后语音时频单元进行幅度谱约束。仿真结果表明,在几种常见背景噪声的低信噪比情况下,相比于传统维纳滤波法,本文算法效果更好,能有效的提高语音可懂度。In order to improve the speech intelligibility with low signal-to-noise ratio, a speech enhancement algorithm of the binary mask estimation based on noise spectrum constraints is proposed. We first analyzed the impact of over-estimation of the a priori signal-to-noise ratio on the noise spectrum estimation function with low signal-to-noise ratio. And then we modified the a priori signal-to-noise ratio and the noise spectrum estimation function respectively. Finally, we used the modified noise spectnnn estimation function and priori signal-to-noise ratio to extract the time-frequency units where the noise spectrums were under-estimated, and then to estimate the binary mask values according to the different time-frequency units. We used the bi- nary mask values to the estimated speech time-frequency units to make a constraint. Simulation results show that under several common background noises with low signal-to-noise ratio, the performance of proposed approach is more excellent and can im- prove the speech intelligibility effectively compared with traditional Wiener filtering method.

关 键 词：二值掩码 语音可懂度 先验信噪比 噪声谱估计 

分 类 号：TN912.35[电子电信—通信与信息系统]