【作者】 王子腾； 纪璇； 王晓飞； 付强； 颜永红；

【Author】 Wang Ziteng;Ji Xuan;Wang Xiaofen;Fu Qiang;Yan Yonghong;Chinese Academy of Sciences;

【机构】 中国科学院声学研究所；

【摘要】 近年来,将深度神经网络(deep neural network,DNN)应用于语音增强取得了一定的效果。该文采用DNN,通过大量数据来学习带噪信号到干净信号之间的非线性映射关系,形成深度映射网络,以谱映射(spectral mapping)的方式实现了对目标信号的估计与重构,以特征映射(feature mapping)的方式实现了特征增强,并在实际噪声场景下的CHi ME-3数据集上获得了识别率的有效提升。同时,该文进一步对比谱映射与特征映射在自动语音识别应用中的性能表现,分析了将映射网络作为后处理方案和用于匹配训练时的性能差异。更多还原

【Abstract】 Recently, applying deep neural network(DNN) to solve the speech enhancement problem has made some progress. This article employs DNN to learn the nonlinear mapping function between noisy signals and clean signals, resulting in a deep mapping network, by which spectral mapping aims for estimation and reconstruction of target signal and feature mapping means for feature enhancement. The system achieves a higher recognition rate on the CHi ME-3 database, one of real environment noise. Also, experiments are designed to show the capability of spectral mapping and feature mapping, and performance difference of DNN as post-processing and match-training, in automatic speech recognition.更多还原