Gammatone域特征在IRM-DBN语音增强中的有效性研究  

Research on the Effectiveness of Speech Features in Gammatone Domain for IRM-DBN Speech Enhancement

作　　者：王卫梅 王雁 贾海蓉 WANG Wei-mei;WANG Yan;JIA Hai-rong(College of Information and Computer,Taiyuan University of Technology,Taiyuan 030024,China)

机构地区：[1]太原理工大学信息与计算机学院

出　　处：《内蒙古大学学报（自然科学版）》2019年第6期666-673,共8页Journal of Inner Mongolia University：Natural Science Edition

基　　金：国家自然科学基金项目(61371193);山西省自然科学基金项目(201701D121058)

摘　　要：对于在噪声背景下的深度学习来说,好的特征提取能极大地提高语言增强的性能.研究在深度信念网络中,以目前性能最好的理想浮值掩蔽为学习目标,验证Gammatone域特征的语音增强效果优于其他域特征.首先,分别提取在不同噪声不同信噪比下的基于Gammatone域的语音特征,根据纯净语音和噪声计算得到初始理想浮值掩蔽;其次,采用深度信念网络作为学习模型,从训练带噪语音特征中学习得到估计的理想浮值掩蔽;最后,利用测试语音特征和估计的理想浮值掩蔽合成增强语音,分析所用特征的有效性.实验结果表明:基于Gammatone域的语音特征比其他域特征的各种性能评价指标值更高,明显提高了语音质量,增强效果更佳.For deep learning in noisy backgrounds,a good feature extraction can greatly improve the performance of speech enhancement.The ideal ratio mask with the best performance as the learning target in deep belief networks,it is proved that the speech enhancement effect of Gammatone domain features is better than that of other domain features.Firstly,the speech features in Gammatone domain under different noises and different signal-to-noise ratios are extracted,and the initial ideal ratio mask is calculated according to clean speech and noise.Secondly,using the deep belief network as a learning model,an estimated ideal ratio mask is obtained from the training noisy speech features.Finally,an enhanced speech is synthesized by using the test speech features and the estimated ideal ratio mask,and the validity of the features used is analyzed.The experimental results show that the speech features based on Gammatone domain have higher performance evaluation values than other domain features.The speech quality is obviously improved and the enhancement effect is better.

关 键 词：特征提取 深度信念网络 理想浮值掩蔽 Gammatone域 语音增强 

分 类 号：TN912.35[电子电信—通信与信息系统]