中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
首页 机构 成果 学者
  
  1. CAS IR Grid
  2. 自动化研究所
  3. 中国科学院自动化研究所
  4. 多模态人工智能系统全国重点实验室
  5. 智能交互
Improving speech enhancement by focusing on smaller values using relative loss

文献类型：期刊论文

作者Li, Hongfeng1,2; Xu, Yanyan1,2; Ke, Dengfeng3; Su, Kaile4
刊名IET SIGNAL PROCESSING
出版日期2020-08-01
卷号14期号:6页码:374-384
关键词speech enhancement speech intelligibility performance evaluation learning (artificial intelligence) neural nets absolute differences speech quality relative loss single-channel speech enhancement noisy speech ideal ratio mask phase-sensitive mask mean square error loss function absolute error values magnitude spectra deep learning clean speech recovery short-time objective intelligibility signal-to-distortion ratio segmental signal-to-noise ratio performance evaluation
ISSN号1751-9675
DOI10.1049/iet-spr.2019.0290
通讯作者Xu, Yanyan(xuyanyan@bjfu.edu.cn)
英文摘要The task of single-channel speech enhancement is to restore clean speech from noisy speech. Recently, speech enhancement has been greatly improved with the introduction of deep learning. Previous work proved that using ideal ratio mask or phase-sensitive mask as intermediation to recover clean speech can yield better performance. In this case, the mean square error is usually selected as the loss function. However, after conducting experiments, the authors find that the mean square error has a problem. It considers absolute error values, meaning that the gradients of the network depend on absolute differences between estimated values and true values, so the points in magnitude spectra with smaller values contribute little to the gradients. To solve this problem, they propose relative loss, which pays more attention to relative differences between magnitude spectra, rather than the absolute differences, and is more in accordance with human sensory characteristics. The perceptual evaluation of speech quality, the short-time objective intelligibility, the signal-to-distortion ratio, and the segmental signal-to-noise ratio are used to evaluate the performance of the relative loss. Experimental results show that it can greatly improve speech enhancement by focusing on smaller values.
WOS关键词DEEP NEURAL-NETWORK ; SEPARATION
资助项目World-Class Discipline Construction and Characteristic Development Guidance Funds for Beijing Forestry University[2019XKJS0310]
WOS研究方向Engineering
语种英语
WOS记录号WOS:000555924300006
出版者INST ENGINEERING TECHNOLOGY-IET
资助机构World-Class Discipline Construction and Characteristic Development Guidance Funds for Beijing Forestry University
源URL[http://ir.ia.ac.cn/handle/173211/40366]  
专题模式识别国家重点实验室_智能交互
通讯作者Xu, Yanyan
作者单位1.Natl Forestry & Grassland Adm, Engn Res Ctr Forestry Oriented Intelligent Inform, 35 Qing Hua East Rd, Beijing 100083, Peoples R China
2.Beijing Forestry Univ, Sch Informat Sci & Technol, 35 Qing Hua East Rd, Beijing 100083, Peoples R China
3.Chinese Acad Sci, Inst Automat, Natl Lab Pattern Recognit, 95 Zhong Guan Cun East Rd, Beijing 100190, Peoples R China
4.Griffith Univ, Inst Integrated & Intelligent Syst, 170 Kessels Rd, Nathan, Qld 4111, Australia
推荐引用方式
GB/T 7714		 Li, Hongfeng,Xu, Yanyan,Ke, Dengfeng,et al. Improving speech enhancement by focusing on smaller values using relative loss[J]. IET SIGNAL PROCESSING,2020,14(6):374-384.
APA Li, Hongfeng,Xu, Yanyan,Ke, Dengfeng,&Su, Kaile.(2020).Improving speech enhancement by focusing on smaller values using relative loss.IET SIGNAL PROCESSING,14(6),374-384.
MLA Li, Hongfeng,et al."Improving speech enhancement by focusing on smaller values using relative loss".IET SIGNAL PROCESSING 14.6(2020):374-384.

入库方式： OAI收割

来源：自动化研究所

浏览0
下载0
收藏0
其他版本

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。