基于K-SVD算法和组合字典的语音信号清浊音判决研究  

Judgement of Voiced and Unvoiced Sounds Based on K-SVD and Combined-Dictionary

作　　者：王莲子 李钟晓 陈倩倩 庄晓东[1] WANG Lianzi;LI Zhongxiao;CHEN Qianqian;ZHUANG Xiaodong(College of Electronic Information,Qingdao University,Qingdao 266071,China)

机构地区：[1]青岛大学电子信息学院,山东青岛266071

出　　处：《青岛大学学报（工程技术版）》2020年第2期17-23,共7页Journal of Qingdao University(Engineering & Technology Edition)

基　　金：国家自然科学基金资助项目(41804110)。

摘　　要：针对语音中清音和浊音特性的不同,本文提出了一种新的清浊音判别方法,利用K奇异值分解(K singular value decomposition,KSVD),分别对数据样本中的清音和浊音进行字典学习,训练出符合样本信号特性的浊音字典和清音字典,将多个单清音字典组合成组合清音字典,多个单浊音字典组合成组合浊音字典,并将待测信号在组合浊音字典和组合清音字典上进行稀疏表示,通过对比其系数的稀疏性来判别清浊音。研究结果表明,在相同条件下,与传统的清浊音判别方法相比,基于组合字典的判别方法对于多音素的清浊音判决具有更加准确的判决结果。该研究对语音识别和语音编码具有重要作用。According to the difference between voiced and unvoiced sounds,a new method has been proposed to judge them in this paper.The study selects enough voiced and unvoiced sounds as the object of dictionary learning and employs K-SVD algorithm to construct voiced dictionary and unvoiced dictionary respectively.And multiple single-voiced dictionaries are combined into a combined-voiced dictionary,while multiple single-unvoiced dictionaries are combined into a combined-unvoiced dictionary.Then the signals to be judged are sparse represented in combined-voiced dictionary and combined-unvoiced dictionary.The way to distinguish the voiced and unvoiced sounds is comparing the sparsity of coefficients in two dictionaries.The results have showed that comparing with the traditional methods by a large number of experimental results,the method based on combinatorial-dictionary is more accurate in the judgement of multiple phonemes under the same condition.This research is important for speech recognition and speech coding.

关 键 词：语音判别 字典学习 稀疏表示 组合字典 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术] TN912.34[自动化与计算机技术—计算机科学与技术]