基于声门波码本受限的迭代维纳滤波语音增强  

Glottal codebook constrained iterative Wiener filtering speech enhancement

作　　者：戴明扬[1] 周毅[1] 徐柏龄[1] 

机构地区：[1]南京大学声学研究所近代声学国家重点实验室,南京210093

出　　处：《声学学报》2003年第1期21-27,共7页Acta Acustica

基　　金：国家自然科学基金资助项目(60272037)

摘　　要：对基于全极点模型的迭代维纳滤波语音增强方法进行深入研究,指出声门波波形失真和畸变是导致该种方法增强语音听感质量不好的主要原因。为进一步提高迭代维纳滤波增强语音的听感质量,本文从声门波的恢复和增强入手,提出了基于声门波码本受限的迭代维纳滤波语音增强方法:首先对干净训练语音逆滤波,获取其声门波信号;并基于参数模型分析该声门波,提取其特征参数;再根据声门波及其特征参数对声门波训练集进行κ-均值聚类,生成干净的声门波码本,该码本充分反映了干净语音的有效激励声源模式;用该码本对声门波加以规范和约束,使迭代维纳滤波过程中增强语音的激励声门波处干干净语音有效激励声源模式空间内。模拟实验表明,在同样输入条件下,采用声门波码本受限的迭代维纳滤波方法增强的语音比基于全极点模型的迭代维纳滤波方法增强的语音具有更小的失真,提高了增强语音的听感质量。After deeply researching into the All-Pole based iterative Wiener filtering speech enhancement algorithm, we obtained the result that low quality of the enhanced speech was mainly due to glottal wave distortion, so we probed into the crude way to improve enhanced speech quality by glottal wave restoration. We proposed the glottal codebook constrained iterative Wiener filtering speech enhancement algorithm. This method adapts the clean training speech to get it's glottal wave, and analyses the clean training glottal wave frame by frame pitch synchronously based on LF model and modified turbulent noise model, and then clusters training glottal wave frames according to extracted glottal features using k-means algorithm to produces the glottal codebook, which represents the effective exciting model of clean human speech. We imposes constraints on the glottal wave of enhanced speech during iterative Wiener filtering according as the glottal codebook to ensure glottal wave laid in the clean exciting glottal space. As shown in the simulation experiments, this method can reduce the glottal wave distortion and improve enhanced the speech quality effectively.

关 键 词：声门波码本受限 迭代维纳滤波 语音增强 全极点模型 语音信号处理 听感质量 

分 类 号：TN912.3[电子电信—通信与信息系统]