一种深度降噪自编码器的语音增强算法  

A Speech Enhancement Algorithm for Deep Noise Reduction Autoencoder

作　　者：刘鹏 LIU Peng(Department of Information Engineering and Big Data Science,Shanxi Institute of Technology,Yangquan 045000)

机构地区：[1]山西工程技术学院信息工程与大数据科学系,阳泉045000

出　　处：《计算机与数字工程》2021年第5期875-879,共5页Computer & Digital Engineering

基　　金：山西工程技术学院科研课题(编号:2020004)资助。

摘　　要：依据带噪语音中不同类型语音分段(segment)对语音整体的可懂度影响不同,提出了一种基于语音分段来分类训练深度降噪自编码器(DDAE)的语音增强算法。该算法使得DDAE模型在尽可能减小Dropout所引入的扰动对带噪语音噪声特性破坏的同时,提高了对带噪语音可懂度关键分段(中均方根分段)语音特性学习的鲁棒性,提高了增强语音的可懂度。实验结果表明,该算法较现有方法提高了增强语音可懂度的NCM值。According to the principle that different types of speech segments in noisy speech have different influence on the in⁃telligibility of the speech,a speech enhancement algorithm based on the Deep Denoising AutoEncoder(DDAE)trained by the clas⁃sified speech segments is proposed.This algorithm makes the DDAE model minimize the disruption of the noise characteristics intro⁃duced by the Dropout.Meanwhile it makes the DDAE model improve the robustness of the speech feature learning of the key seg⁃ments(the middle-level RMS segments)of noisy speech intelligibility,the intelligibility of enhanced speech is improved.The ex⁃perimental results show that the proposed algorithm improves the NCM value of enhanced speech intelligibility compared with the ex⁃isting methods.

关 键 词：语音分段 深度降噪自编码器 DROPOUT 语音可懂度 

分 类 号：TN912.35[电子电信—通信与信息系统]