基于稀疏编码的鲁棒说话人识别  

Robust Speaker Recognition Based on Sparse Coding

作　　者：何勇军[1] 孙广路[1] 付茂国 韩纪庆[2] 

机构地区：[1]哈尔滨理工大学计算机科学与技术学院,哈尔滨150080 [2]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001

出　　处：《数据采集与处理》2014年第2期198-203,共6页Journal of Data Acquisition and Processing

基　　金：国家自然科学基金(61305001;91120303;91120301)资助项目;高等学校博士学科点专项科研基金(20132303120003)资助项目;中国博士后基金(2013M531042)资助项目;黑龙江省博士后基金(LBH-Z13099)资助项目;黑龙江教育厅基金(12511096)资助项目;黑龙江省自然科学基金(F200936)资助项目

摘　　要：目前的说话人识别系统在噪声环境下性能将急剧下降,为了解决这一问题,提出了一种新的基于稀疏编码的说话人识别方法。该方法用一个通用背景字典(Universal background dictionary,UBD)刻画说话人语音的共性,并为每个说话人和环境噪声训练相应的字典来刻画说话人和环境的特殊变化。这些字典被拼接成一个大字典,然后将待测试语音稀疏分解在这个大字典上以实现说话人识别。为了提高说话人字典的区分能力,通过从说话人字典中移除与通用背景字典原子相似的原子来优化说话人字典。为了跟踪变化的噪声,采用混噪语音在线更新噪声字典。在各种噪声条件下的实验表明,所提出的方法在噪声环境下具有较强的鲁棒性。Speaker recognition suffers severe performance degradation under noisy environ- ments. To solve this problem, a novel method is proposed based on sparse coding of signals. This method employs a universal background dictionary （UBD） to model common variability of all speakers, a speaker dictionary to model special variability of each speaker and a noise dic- tionary to model variability of environmental noise. These three dictionaries are concatenated to be a big dictionary, over which test speech is sparsely represented and classified. To im- prove the discriminability of speaker dictionaries, the speaker dictionaries are optimized by re- moving speaker atoms which are close to the UBD atoms. To ensure the varied noises can be tracked, an algorithm is designed to update the noise dictionary with the noisy speech. Experi- mental results under various noise conditions show that the proposed method can obviously im- prove the robustness of speaker recognition under noisy environments.

关 键 词：说话人识别 稀疏表示 判别字典 形态成分分析 

分 类 号：R318.04[医药卫生—生物医学工程]