“ASR Status Report 2017-7-31”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
 
(2位用户的4个中间修订版本未显示)
第6行: 第6行:
 
|Xiaofei Kang
 
|Xiaofei Kang
 
||  
 
||  
*  
+
* Finish the Speaker Recognition experiment:mouth with candy, normal chat
 
||  
 
||  
*  
+
* Understand all the scripts of the Speaker Recognition experiment, and then learn to modify it.
 
|-
 
|-
  
第48行: 第48行:
 
|Yixiang Chen   
 
|Yixiang Chen   
 
||  
 
||  
* check paper
 
 
* plot tsne picture for 863 & fisher-5000 data set
 
* plot tsne picture for 863 & fisher-5000 data set
 
* find why performance of wisper better than performance of chat
 
* find why performance of wisper better than performance of chat
 
||  
 
||  
*  
+
* check data and paper
 
|-
 
|-
  
第58行: 第57行:
 
|Lantian Li   
 
|Lantian Li   
 
||  
 
||  
*  
+
* T-sne plot for speaker segmentation preparation [http://cslt.riit.tsinghua.edu.cn/mediawiki/images/e/e2/Spk_seg.pdf].
 +
* check TASLP and NIPS paper.
 
||  
 
||  
*  
+
* deep spk recipe.
 
|-
 
|-
  

2017年7月31日 (一) 04:40的最后版本

Date People Last Week This Week
2017.7.31 Xiaofei Kang
  • Finish the Speaker Recognition experiment:mouth with candy, normal chat
  • Understand all the scripts of the Speaker Recognition experiment, and then learn to modify it.
Miao Zhang
  • finish the experiments on five kinds of speech
  • optimize the vad parameter to improve the performance
  • finish the new human test website
Yanqing Wang
  • retraining task: experiments are in progress, some time needed.
  • all experiments should be done.
  • TRP of retraining task.
Ying Shi
  • apply mongodb and ajax on the data checking website
    • with mongodb we are not depend on file lock anymore
    • there is no need to save web state(except some cookie) after employ ajax
  • continue to learn crawler
  • setup server for m2asr (use sheep02)
  • design crawler program
Yixiang Chen
  • plot tsne picture for 863 & fisher-5000 data set
  • find why performance of wisper better than performance of chat
  • check data and paper
Lantian Li
  • T-sne plot for speaker segmentation preparation [1].
  • check TASLP and NIPS paper.
  • deep spk recipe.
Zhiyuan Tang
  • Updated the auto-scoring system with the newest version of Kaldi. Several patches need to be repaired.
  • Kaldi book writing.
  • Initial version of auto-scoring system.
  • Kaldi book writing.




Date People Last Week This Week
2017.7.24 Xiaofei Kang
  • Prepare the data set of Speaker Recognition : pick out whisper
  • Learn the the nnet3 model, run the nnet3 experiment
  • Learn the Speaker Recognition model, run the Speaker Recognition experiment
Miao Zhang
  • joined a meeting in Chinese Academy of Social Sciences
  • worked out a recording plan
  • learnt kaldi and did experiments
  • test performances on 12 kinds of voices we have
Hui Tang
  • help jiayin to configure dnn and lstm in kaldi
  • left for postgraduate life
Yanqing Wang
  • change the source code of Kaldi to implement retraining ( with zero value fixed )
  • start to write a technical report of pruning the neural network ( not finished )
  • finish the retraining task
  • finish the technical report
Ying Shi
  • data checking website
  • learn how to write a crawler program
  • write a more general crawler
  • realign kazak train and test data with transfer learning model
Yixiang Chen
  • use wisper audio for speaker recognition
  • joined a meeting in Chinese Academy of Social Sciences
  • test performances on 12 kinds of voices
Lantian Li
  • deepspk on TASLP.
  • speaker segmentation.
  • recipe of deepspk.
Zhiyuan Tang
  • Replaced ATLAS lib with MKL lib for compiling auto-scoring system.
  • Kaldi book writing.
  • A basic demo for auto-scoring system.
  • Kaldi book writing.