摘要
为提高监督性语音分离在多种训练目标下的分离性能,提出一种基于特征组合的多目标监督性语音分离方法.针对声学特征之间的不同特性,采用group lasso方法对时频域特征进行选择及组合;利用深度神经网络(Deep Neural Network,DNN)模型实现组合特征在多目标下的分离实验.实验结果表明,使用特征组合得到的分离性能有显著的提高,且在多种训练目标下都显示出了可行性与优越性.
In order to improve the separation performance of supervised speech separation under vs?nous training targets,a multitargets supervised speech separation method based on feature combination is proposed.According to the different characteristics of acoustic features,the group lasso method is used to select and combine the timefrequency domain features.The deep neural network(DNN)model is used to realize the separation experiments of combined features under multiple targets.The experimental results show that the separation periormance index obtained by the combination of features is significantly improved,and it also have feasibility and superiority under various training objectives.
作者
兰琼琼
陆志华
叶庆卫
周宇
LAN Qiong-qiong;LU Zhi-hua;YE Qing-wei;ZHOU Yu(Faculty of Information Science and Engineering,Ningbo University,Ningbo 31521 I,China)
出处
《无线通信技术》
2019年第3期17-22,共6页
Wireless Communication Technology
基金
国家自然科学基金资助项目(No.61801255)
关键词
语音分离
监督性学习
特征组合
DNN
speech separation
supervised learning
feature combination
DNN