基于Beta混合模型的声学回声抑制  

Acoustic echo suppression based on Beta mixture model

作　　者：童颖[1,2] 顾亚平[3] 屈科[4] 

机构地区：[1]中国科学院声学研究所,北京100190 [2]中国科学院大学,北京100049 [3]中国科学院声学研究所东海研究站,上海201815 [4]广东海洋大学广东省近海海洋变化与灾害预警重点实验室,广东湛江524088

出　　处：《华中科技大学学报（自然科学版）》2017年第6期10-14,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：广东省自然科学基金资助项目(2014A030310256)

摘　　要：为解决全双工模式下声学回声抑制问题,采用Beta混合模型对两个互相关函数特征量建模.通过估计近端目标语音存在概率,控制双端讲话状态下回声抑制算法的房间冲激响应函数的更新.同时引入一种改进的软判决方法,将语音存在概率与增益函数结合,提高回声估计的准确性,抑制更多的残留回声.客观评价准则评估表明:采用基于Beta混合模型的双端讲话检测算法具有比较低的错误概率,且对房间冲激响应的变化具有较好的鲁棒性,在提高近端目标语音质量的同时达到了更好的回声抑制效果.To solve the problem of acoustic echo suppression in full duplex mode,Beta mixture model was adopted to model the two cross-correlation functions.The near-end target speech presence probability was estimated to decide the update of room impulse response function of the echo suppression algorithm in double talk state.A modified soft decision method combining the speech presence probability with the gain function was introduced to improve the accuracy of echo estimation and apply greater attenuation of residual echo.According to the objective evaluation criteria,lower error probability is obtained using double talk detection algorithm based on Beta mixture model,which is robust for the change of room impulse response,and the quality of the near-end target speech signal is improved while achieving a better suppression performance.

关 键 词：声学信号处理 回声抑制 Beta混合模型 双端讲话检测 频谱修正方法 

分 类 号：TN912.3[电子电信—通信与信息系统]