基于稀疏表示的鲁棒性说话人识别系统  

Robust Speaker Recognition System Based on Sparse Representation

作　　者：于云[1] 周伟栋[1] 

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003

出　　处：《计算机技术与发展》2015年第12期41-46,共6页Computer Technology and Development

基　　金：国家自然科学基金资助项目(61271335);国家"973"重点基础研究发展计划项目(2011CB302303);江苏省自然科学基金项目(BK20140891)

摘　　要：基于稀疏表示的说话人识别方法在无噪的环境下已经达到了理想的效果,然而在背景噪声下,此方法的识别性能大幅度下降。为了提高系统的鲁棒性,提出了一种新型的基于稀疏表示的鲁棒性说话人识别系统模型。此系统结合多状态训练和语音增强谱减法,在训练阶段和测试阶段同时利用语音增强技术,然后对增强后的语音进行多状态训练,以便提高训练特征数据集和测试特征数据集之间的匹配度。实验分析和结果表明,所提出的新型模型在所研究的白噪声和有色噪声下达到了很好的抗噪性能,具有很强的鲁棒性。Robust speaker recognition method based on sparse representation in the absence of noise has reached ideal performance. However, speaker recognition based on sparse representation doesn＇ t perform well where background noise exists. To improve the robustness of this system,describe a new robust speaker recognition system based on sparse representation. The system combines multi-condition training and spectrum subtraction,which is thought to be a preprocessing block not only for the testing stage, but also for the training stage. Then propose to make multi-condition training where various sets of features are extracted, so as to improve the matched degree be- tween training data and testing data. Experimental analysis and results show that the proposed new model under white and colored noises can get the great anti -noise performance, and obviously improve the robustness of the speaker recognition under background noisy envi- ronments.

关 键 词：说话人识别 稀疏表示 多状态训练 谱减法 

分 类 号：TN912.3[电子电信—通信与信息系统]