Parameter Mask Speech Enhancement for Robust Automatic Speech Recognition

来源 :2014年国际计算机科学与软件工程学术会议 | 被引量 : 0次 | 上传用户:ikkonen
下载到本地 , 更方便阅读
声明 : 本文档内容版权归属内容提供方 , 如果您对本文有版权争议 , 可与客服联系进行内容授权或下架
论文部分内容阅读
  A parameter mask is proposed and analyzed in this paper to speech enhancement for robust automatic speech recognition (ASR).With the frame work of computational auditory scene analysis (CASA),ideal binary mask (IBM) is used to get the signal to noise ratio (SNR) improvement,but not the ASR performance improvement.The gap between the SNR and ASR improvement is great.To conventional ASR system,the main goal is providing the similar energy distribution to the clean target speech and no matter the energy comes from the speech or noise.We use the SNR in time frequency (T-F) unit to generate the parameter mask (PM) which is used to estimate the clean speech energy from the mixture signals.Experiment results show the higher ASR performance of the proposed method than IBM with very small SNR performance decrease.
其他文献
There are various of materials science data resources in steel domain,and most of these open data resources are available.However,these open data resources may reside in different web sites,literature
The learning of Credit Scoring has recently gained much attention,and many methods based on machine learning approaches have been proposed.Based on the above research,most of existing Credit Scoring m
For the discrete resource allocation problem with alternative plans and the benefits increasing with increasing resource usage,the optimization model for integer variable is created.Due to the variabl
With the development of Chinas economy,credit scoring has become important.The general credit scoring model is to solve the two classification problems,but in real life we often encounter multiple cla
This paper addresses the issues of programming a multi-level parallel computer.This computer has an architecture that combines multi-level parallelism for efficient implementation.To exploit the full
Environmental audio classification has been the focus in the field of speech recognition.Random forest is a powerful machine learning classifier compared to other conventional pattern recognition tech
Mining potential users in micro-blogging network,establishing the appropriate follow predictors and models for active micro-bloggers are very important for increasing the number of fans,and enhancing
With the rapid development and update in electronic products nowadays,it has become a general trend to make the circuit simulation in design of electronic product using EDA software in order to optimi
In order to make the printing enterprise production management more convenient and normalized,the enterprise takes its production order management as the key of enterprise production management and ev
In order to solve problems exist in spatial data watermarking technique with rapid expansion,this study aims to develop a watermarking system of vector maps that is based on spatial relationship.It hi