李涛1,
孙晓月1,
杨登才2,
王琪1
1.北京工业大学信息学部人工智能研究院和信息与通信工程学院 ??北京 ??100124
2.北京工业大学科技发展研究院 北京 100124
基金项目:国家自然科学基金(51477028)，北京市教委科技计划面上项目(KM201510005007)
详细信息
作者简介:李如玮：女，1972年生，博士，副教授，硕士生导师，研究方向为语音信号处理
李涛：男，1994年生，硕士生，研究方向为语音信号处理
孙晓月：女，1995年生，硕士生，研究方向为语音信号处理
杨登才：男，1978年生，博士，副研究员，信号与光信号处理
王琪：女，1991年生，在站博士后，研究方向为语音信号处理
通讯作者:李如玮 liruwei@bjut.edu.cn
中图分类号:TN912.3计量
文章访问数:1329
HTML全文浏览量:984
PDF下载量:37
被引次数:0
出版历程
收稿日期:2018-12-06
修回日期:2019-05-21
网络出版日期:2019-06-04
刊出日期:2019-12-01
Binaural Target Sound Source Localization Based on Time-frequency Units Selection
Ruwei LI1,,,Tao LI1,
Xiaoyue SUN1,
Dengcai YANG2,
Qi WANG1
1. Laboratory of Speech and Audio Signal Processing and Institute of Artificial Intelligence, Faculty of Information Technology, Beijing University of Technology, Beijing 100124, China
2. Institute of Science and Technology Development, Beijing University of Technology, Beijing 100124, China
Funds:The National Natural Science Foundation of China(51477028), The Scientific Research Program of Beijing Municipal Commission of Education (KM201510005007)
摘要
摘要:针对复杂声学环境下，现有目标声源定位算法精度低的问题，该文提出了一种基于时频单元选择的双耳目标声源定位算法。该算法首先利用双耳目标声源的频谱特征训练1个基于深度学习的时频单元选择模型，然后使用时频单元选择器从双耳输入信号中提取可靠的时频单元，减少非目标时频单元对定位精度的负面影响。同时，基于深度神经网络的定位系统将双耳空间线索映射到方位角的后验概率。最后，依据与可靠时频单元相对应的后验概率完成目标语音的声源定位。实验结果表明，该算法在低信噪比和各种混响环境，特别是存在与目标声源类似的噪声环境下目标声源的定位精度得到明显改善，性能优于对比算法。
关键词:目标声源定位/
深度学习/
时频单元选择
Abstract:The performance of the existing target localization algorithms is not ideal in complex acoustic environment. In order to improve this problem, a novel target binaural sound localization algorithm is presented. First, the algorithm uses binaural spectral features as input of a time-frequency units selector based on deep learning. Then, to reduce the negative impact of the time-frequency unit belonging to noise on the localization accuracy, the selector is emploied to select the reliable time-frequency units from binaural input sound signal. At the same time, a Deep Neural Network (DNN)-based localization system maps the binaural cues of each time-frequency unit to the azimuth posterior probability. Finally, the target localization is completed according to the azimuth posterior probability belonging to the reliable time-frequency units. Experimental results show that the performance of the proposed algorithm is better than comparison algorithms and achieves a significant improvement in target localization accuracy in low Signal-to-Noise Ratio(SNR) and various reverberation environments, especially when there is noise similar to the target sound source.
Key words:Target sound localization/
Deep learning/
Time-frequency units selection
PDF全文下载地址:
https://jeit.ac.cn/article/exportPdf?id=835ee074-ec93-4f9f-a2f0-3a466d341a36
