基于并行卷积循环网络的单通道语音增强系统  

Parallel convolutional recurrent network based monaural speech enhancement system

作　　者：李鑫元 黄鹤鸣[1,2] LI Xin-yuan;HUANG He-ming(School of Computer Science,Qinghai Normal University,Xining 810008,China;The State Key Laboratory of Tibetan Intelligent Information Processing and Application,Qinghai Normal University,Xining 810008,China)

机构地区：[1]青海师范大学计算机学院,青海西宁810008 [2]青海师范大学藏语智能信息处理及应用国家重点实验室,青海西宁810008

出　　处：《计算机工程与设计》2023年第4期1181-1188,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(62066039)。

摘　　要：为提升语音增强系统的收敛速度和泛化性,降低对训练数据的要求,提出一种基于并行卷积循环网络的语音增强系统。在卷积循环网络的基础上,使用归一化门控线性单元提升性能和收敛速度;使用并行循环层结构同时处理原始语音特征和经过编码器处理后的语音特征,通过后处理模块处理并行结构的输出。在THCHS30和LibriSpeech语音库及NOISEX92和PNL100 NS噪声库上的实验结果表明,与多个目前最先进的语音增强系统相比,该方法获得了最高36.92%的性能提升和62.36%的收敛速度提升。To improve the convergence speed and generalization of the speech enhancement system while reducing the requirement for training data,a speech enhancement system based on parallel convolutional recurrent network(PCRN)was proposed.Based on the CRN,normalized gated linear units(NGLU)was used to enhance performance and convergence speed,a parallel recurrent layer structure was used to process both the original and the encoder-processed speech features,and the output of the parallel structure was processed through a post-processing module.Experimental results on such speech datasets as THCHS30 and LibriSpeech and such noise datasets as NOISEX92 and PNL 100 Nonspeech Sounds verify that,compared with several state-of-the-art speech enhancement systems,the proposed method achieves up to 36.92%performance improvement and it also improves the convergence speed by 62.36%.

关 键 词：语音增强 单通道语音增强 深度学习 卷积循环网络 并行网络 门控线性单元 低资源训练 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]