“ASR Status Report 2016-12-26”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
第79行： 第79行：
 
* Deep speaker embedding.
 
* Deep speaker embedding.
 
* Write book.
 
* Write book.
 +
* Replay detection on INTERSPEECH chanllenge.
 
|-
 
|-
  

2016年12月26日 (一) 05:40的版本


Date People Last Week This Week
2016.12.26


Jingyi Lin
Yanqing Wang
  • implement the detection mechanism by socket
  • find best parameters to avoid over-fitting
  • add two-class-SVM to the program
  • make GUI more pretty and easy to use
  • improve the program's robustness
  • screenshot:
  • write a document on the program
Hang Luo
Ying Shi
  • crawl corpus from internet.(I don't know whether the corpus is right or not.......)
  • make new LM(complete)
  • train new AM(complete)
  • a part of TRP
  • finish the TRP
Yixiang Chen
  • Prepare the input of speech data (trick of block segmentation)
  • Complete the init version on max-margin SRE.
  • Write TRP-20160012 "基于Kaldi i-vector的说话人识别系统使用说明".
  • Prepare the thesis proposal.
  • Integrate CNN + max-margin.
Lantian Li
  • Deep speaker embedding
    • Prepare two datasets and make the i-vector baselines.
  • Write TRP-20160012 "基于Kaldi i-vector的说话人识别系统使用说明".
  • Write book of robustness SRE.
  • Wechat open account.
  • Deep speaker embedding.
  • Write book.
  • Replay detection on INTERSPEECH chanllenge.
Zhiyuan Tang
  • TRP of "How to Config Kaldi nnet3 (in Chinese)", not finished yet;
  • outline of TRP for "Multi-task Recurrent Model for True Multilingual Speech Recognition".
  • Finish the above 2 TRPs.





Date People Last Week This Week
2016.12.19


Jingyi Lin
  • --
  • Concentrate on checking the cslt.book.
  • Prepare for the annual convention.
Yanqing Wang
  • build a data sender ( read & generate txt files of distracted feature )
  • build a data analyzer ( detect the modification of files and make response ( show tokens ) )
  • screenshot:
  • (maybe) replace the detection mechanism by socket
  • find best parameters to avoid over-fitting
  • add two-class SVM to the program
  • make GUI more pretty and easy to use
Hang Luo
  • Compare decode result between mono and bi LM, and the decode result ues bi LM before and after joint
  • Choose wrong decode sentence and find its difference between baseline and shareGMM baseline
  • Finished ML book
  • Continue joint training analysis work, but I'm very confused about how to improve
Ying Shi
  • some work about kazak lm
  • crawl data from kazak internet
  • run new AM by current speech data
  • get more corpus from internet
  • use current corpus make LM and decode
Yixiang Chen
  • Leanring tensorflow
  • coding pair wise net use tensorflow
  • alter CNN
  • coding CNN connect pair wise
  • Dealing with the issue of different lengths of voice
Lantian Li
  • LRE challenge on AP16-OL7.
  • Jeju for APSIPA16.
  • LRE on AP16-OL7.
  • Deep speaker embedding.
Zhiyuan Tang
  • Jeju for APSIPA16.
  • A speech about recent ASR improvements.
  • A supplementary TRP for "Multi-task Recurrent Model for True Multilingual Speech Recognition".
取自“http://index.cslt.org/mediawiki/index.php?title=ASR_Status_Report_2016-12-26&oldid=25076”