基于深度神经网络的单通道语音分离算法  

A monaural speech separation algorithm based on deep neural networks

作　　者：梁尧 朱杰[1] 马志贤[1] LIANG Yao;ZHU J ie;MA Zhi-xian(Department of Electronic Engineering,Shanghai Jiaotong University,Shanghai 200240,Chin)

机构地区：[1]上海交通大学电子工程系,上海200240

出　　处：《信息技术》2018年第7期24-27,共4页Information Technology

基　　金：国家自然科学基金项目(61371147;11433002);国家重点研发计划资助(2017YFF0210903)

摘　　要：单通道包含的信息非常有限使得单通道语音分离是一个非常具有挑战性的研究领域。文中提出一种基于深度神经网络(DNN)的监督学习算法,将混合语音信号从单通道中分离出来。首先,将原始信号混合并提取特征,输入到分离网络中输出分离后的信号。通过设计实验,对网络的最佳参数进行选择,并确定特征向量和上下文窗口大小。最后采用TIMIT语料库来评估提出的算法,与基于非负矩阵分解(NMF)的模型相比,基于DNN的模型在SDR指标上提高了1.67d B。Monaural speech separation has become a very challenging research area,since there is very limited information from only one channel. In this paper,a supervised learning algorithm,typically a deep neural network（DNN） based model is proposed to separate mixed speech signals from one channel.Firstly,the original signals are mixed and features are extracted,which are then input into the separation network to output the divided signals. Moreover,the experiments are conducted to select the best parameters of the network,as well as the feature vectors and context window sizes. At last,evaluations are conducted using the TIMIT speech corpus. The DNN based model achieves 1. 67 dB SDR gain compared to nonnegative matrix factorization（NMF） based models.

关 键 词：单通道语音分离 监督学习 深度神经网络 

分 类 号：TN912.3[电子电信—通信与信息系统]