基于聚类分析与说话人识别的语音跟踪  

Speech Tracking Based on Cluster Analysis and Speaker Recognition

作　　者：郝敏 刘航 李扬 简单 王俊影 HAO Min;LIU Hang;LI Yang;JIAN Dan;WANG Jun-ying(School of Electro-Mechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学机电工程学院,广东广州510006

出　　处：《计算机与现代化》2020年第4期7-13,18,共8页Computer and Modernization

基　　金：广东省省级科技计划项目(2013B011304008,2013B090600031);佛山市产学研专项资金项目(2012HC100195)。

摘　　要：目前语音跟踪在说话人干扰的条件下,即一段语音中存在多个说话人的混合语音信号时,语音跟踪质量会严重下降。针对这种情况,提出一种基于聚类分析与说话人识别的语音跟踪算法。算法首先使用改进的聚类分析方法进行语音分离,具体包括在K-means聚类中对质心进行缓存并降低采样率,以及在embedding特征空间引入正则项。其次,算法采用GMM-UBM说话人模型进行语音跟踪。实验结果表明改进的聚类分析方法可以有效提高算法的实时性及其语音分离质量,GMM-UBM模型在3 s语音的测试中具有84%的识别率。At present,the speech tracking quality will be seriously reduced under the condition of speaker interference,that is,mixed speech signals of multiple speakers in a speech segment.Aiming at this situation,a speech tracking algorithm based on cluster analysis and speaker recognition is proposed.Firstly,the improved clustering analysis method is used for speech separation.Specifically,it includes caching the center of mass and lowering the sampling rate in K-means clustering,and introducing regular terms into embedding feature space.Secondly,the GMM-UBM speaker model is used for speech tracking.The experimental results show that the improved cluster analysis method can effectively improve the real-time performance of the algorithm and the quality of speech separation,the GMM-UBM model has an 84%recognition rate in 3 s speech test.

关 键 词：单信道语音跟踪 智能语音 聚类分析 高斯混合模型 长短期记忆网络 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]