基于信号子空间和DNN的语音增强方法  

Speech Enhancement Method Based on Signal Subspace and DNN

作　　者：董胡 陈伟 彭高丰 陈耀东 刘刚 DONG Hu;CHEN Wei;PENG Gaofeng;CHEN Yaodong;LIU Gang(School of Information Science and Engineering,Changsha Normal University,Changsha 410100,China;School of Electronic Information and Electrical Engineering,Changsha University,Changsha 410022,China;School of Physics,Central South University,Changsha 410083,China)

机构地区：[1]长沙师范学院,信息科学与工程学院,湖南长沙410100 [2]长沙学院,电子信息与电气工程学院,湖南长沙410022 [3]中南大学,物理学院,湖南长沙410083

出　　处：《微型电脑应用》2025年第1期32-34,38,共4页Microcomputer Applications

基　　金：教育部人文社会科学研究青年基金项目(22YJCZH025);湖南省教育厅重点研究项目(21A0618);长沙市自然科学基金项目(kq2202313)。

摘　　要：虽然传统的单通道语音增强技术使用简单,但其计算复杂度低且存在音乐噪声等问题。虽然基于深度神经网络(DNN)的语音增强方法有相对好的语音增强性能,但是它需要适当的训练数据集和较大的计算量。为了解决这些问题,提出一种结合子空间与DNN的语音增强方法。使用子空间去除部分噪声,再使用DNN进一步去除残留噪声,以增强语音的可理解性。实验结果表明,在非平稳低信噪比环境中,提出的语音增强方法在PESQ和STOI 2个方面都优于传统的子空间和DNN方法。Although the traditional single-channel speech enhancement technology is simple to use,it has low computational complexity and has problems such as music noise.Although the speech enhancement method based on deep neural network(DNN)has relatively good speech enhancement performance,it requires appropriate training data set and a large amount of computation.This paper proposes a speech enhancement method combining subspace and DNN.The subspace is used to remove part of the noise,and the DNN is used to further remove residual noise to enhance the intelligibility of speech.The experimental results show that the speech enhancement method proposed is superior to the traditional subspace and the DNN method in both PESQ and STOI aspects in the non-stationary low signal to noise ratio environment.

关 键 词：语音增强 子空间 深度神经网络 可懂度 

分 类 号：TN912.35[电子电信—通信与信息系统]