结合卷积平滑耳蜗谱和深度网络的语音增强技术  

Speech Enhancement Based on Convolution Smoothing Cochleagram and Deep Networks

作　　者：柏梁泽 高勇 BAI Liangze;GAO Yong(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,China)

机构地区：[1]四川大学电子信息学院,四川成都610065 [2]四川大学电子信息学院通信教研室,四川成都610065

出　　处：《无线电工程》2020年第12期1055-1062,共8页Radio Engineering

基　　金：四川大学科研资助项目(0020505501744)。

摘　　要：为了进一步提高基于深度神经网络的语音增强方法的性能,针对语音声学特征和网络的设计两方面进行优化。引入了一种声学特征卷积平滑耳蜗谱,组合了3个不同平滑度的耳蜗谱,能够有效捕捉语音局部和整体信息。网络结构上,使用深度双向长短时记忆网络可以有效提高网络的学习能力。采用一种后处理差分平均滤波,提高了特征参数在语音增强中的效果。实验结果表明,相比于传统算法和深度学习的方法,增强后的语音不管是清晰度还是质量都有显著提升。In order to further improve the performance of speech enhancement method based on deep neural network,two aspects of speech acoustic features and network design are optimized.This paper introduces an acoustic feature convolution smoothing cochleagram.This feature combines three cochleagram spectra with different smoothness,which can effectively capture the local and global information of speech.In terms of network structure,deep bidirectional long short memory network can effectively improve the learning ability of the network.In addition,a post-processing differential average filter is used to improve the effect of feature parameters in speech enhancement.Experiment results show that compared with the traditional algorithm and deep learning method,the proposed method improves the speech clarity and quality significantly.

关 键 词：语音增强 卷积平滑耳蜗谱 深度双向长短时记忆网络 差分平均滤波 

分 类 号：TN912.3[电子电信—通信与信息系统]