“2024-04-08”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
 
(12位用户的22个中间修订版本未显示)
第34行: 第34行:
 
|Ying Shi
 
|Ying Shi
 
||  
 
||  
*
+
* Detect wake-up words from Continuous speech Down
 +
* SPL Paper, almost down
 +
* [https://z1et6d3xtb.feishu.cn/wiki/DYSjw7LfviU7u1kWbhEcFjE1nld?from=from_copylink Group Work]
 
||
 
||
 
*  
 
*  
第67行: 第69行:
 
|Chen Chen
 
|Chen Chen
 
||  
 
||  
*  
+
* Finished my thesis :)
 +
* Group work [https://z1et6d3xtb.feishu.cn/docx/FvXjdKWH1oejYgxnjFwcQKFan8g?from=from_copylink]
 +
** finish training for most of the structures, but all performs a little worse
 +
** no good news from KD
 +
** good primary experiment result from char as modeling unit, need to check after whole training
 +
** 58 hours for CN-CVS II (too slow)
 
||
 
||
*  
+
* Try some data aug methods for VSR (Crop size)
 +
* Help with entropy analyze of child data
 +
* START ICASSP2024 PAPER READING
 
||
 
||
 
*   
 
*   
第78行: 第87行:
 
|Xiaolou Li
 
|Xiaolou Li
 
||  
 
||  
*  
+
* Reproduce different VSR structure [https://z1et6d3xtb.feishu.cn/docx/BgD4djeTioCgd6xjFXNcBXXLnKg?from=from_copylink]
 +
** interCTC, Resnet3D frontend, Branchformer, E-Branchformer
 +
* Paper reading
 +
** Mamba, icassp2024
 
||
 
||
*  
+
* modify E-Branchformer
 +
* Resnet3D frontend + Branchformer test
 +
* s4 decoder (maybe), mamba paper learning
 
||
 
||
 
*   
 
*   
第89行: 第103行:
 
|Zehua Liu
 
|Zehua Liu
 
||  
 
||  
*
+
* read papper
 +
* auxiliary loss code[https://z1et6d3xtb.feishu.cn/docx/ZaTFd3A5EoK982xWBVschloanee?from=from_copylink]
 
||
 
||
 
*  
 
*  
第100行: 第115行:
 
|Pengqi Li
 
|Pengqi Li
 
||   
 
||   
*  
+
* Summary[https://z1et6d3xtb.feishu.cn/docx/Mt46dqE8UoBnuBxu7q8cPletnzb] of speech processing XAI for NSFC
 +
** polish, reference, v1(90%)
 +
* Workshop report(video, slide, poster)
 
||
 
||
 
*
 
*
第111行: 第128行:
 
|Wan Lin
 
|Wan Lin
 
||  
 
||  
*  
+
* Graduation paper
 
||
 
||
 
*
 
*
第122行: 第139行:
 
|Tianhao Wang
 
|Tianhao Wang
 
||  
 
||  
* EA-ASP (from SJTU) reproduced successfully
+
* EA-ASP (from SJTU) reproduced successfully [https://z1et6d3xtb.feishu.cn/docx/BywjdkGvNou12sxQ4dAcxYa9noh]
** EA-ASP module implement, wespeaker toolkit modification totally according to the paper
+
** EA-ASP implement, wespeaker toolkit modification, training pairs construction totally according to the paper
** vox2 training pairs construction following the paper
+
 
** get the better EER(4.021%) comparing to the paper (5.212%) on Vox1-O-Overlap
 
** get the better EER(4.021%) comparing to the paper (5.212%) on Vox1-O-Overlap
 
** evaluation on our trials (concat and weak_overlap are better, overlap and mix are worse)
 
** evaluation on our trials (concat and weak_overlap are better, overlap and mix are worse)
第148行: 第164行:
 
|-
 
|-
 
|Junhui Chen
 
|Junhui Chen
 +
||
 +
* Neural scoring with Frequency-channel attention
 +
** Our overlap test trial: EER 7.382% -> 7.132%
 +
* Graduation paper
 
||
 
||
 
*  
 
*  
||
 
*
 
 
||
 
||
 
*   
 
*   
第160行: 第178行:
 
|Jiaying Wang
 
|Jiaying Wang
 
||  
 
||  
* HHuawei project:train speakerbeam on -4,4,converge,some bug in test(dataloader)
+
* HHuawei project:train speakerbeam on -4,4,converge,some bug in test
 
* cohort mini loss: check code & test
 
* cohort mini loss: check code & test
  
第173行: 第191行:
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
*  
+
* Financial Backtesting pipline
 +
** remake Stock and Industry return logic
 +
** debug Brinson Analysis result
 
||
 
||
*
+
* Use SAC as a baseline to run through the entire process from training to policy generation to backtesting
 
||
 
||
 
*   
 
*   
第196行: 第216行:
 
|Yang Wei
 
|Yang Wei
 
||  
 
||  
*  
+
* Fix problems about ASR model for mispronunciation detection task
 +
* Prepare the baseline system
 
||
 
||
 
*
 
*
第206行: 第227行:
 
|Lily
 
|Lily
 
||
 
||
 +
* Data Analysis
 
* Paper Reading[https://z1et6d3xtb.feishu.cn/docx/L0jGdCqEXouL8hx8kelcrJzjn8d?from=from_copylink]
 
* Paper Reading[https://z1et6d3xtb.feishu.cn/docx/L0jGdCqEXouL8hx8kelcrJzjn8d?from=from_copylink]
 
||
 
||

2024年4月8日 (一) 10:58的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • Primary AI design
  • NMI neuralmag paper submitted
Lantian Li
  • GPU status [1]
  • Projects (POC of Cough/Humming detection, TSE proposal)
  • ASIP-BUPT (NeuralScoring, CohortTSE)
  • AI Course Polish
  • Machine Moving
  • New machine (rabbit04)
Ying Shi
  • Detect wake-up words from Continuous speech Down
  • SPL Paper, almost down
  • Group Work
Zhenghai You
  • Checked the code of mini loss[2]
  • Test onnx of Huawei project[3]
Junming Yuan
  • Experimental report on "learn not to listen" v3 extend test[4]
Chen Chen
  • Finished my thesis :)
  • Group work [5]
    • finish training for most of the structures, but all performs a little worse
    • no good news from KD
    • good primary experiment result from char as modeling unit, need to check after whole training
    • 58 hours for CN-CVS II (too slow)
  • Try some data aug methods for VSR (Crop size)
  • Help with entropy analyze of child data
  • START ICASSP2024 PAPER READING
Xiaolou Li
  • Reproduce different VSR structure [6]
    • interCTC, Resnet3D frontend, Branchformer, E-Branchformer
  • Paper reading
    • Mamba, icassp2024
  • modify E-Branchformer
  • Resnet3D frontend + Branchformer test
  • s4 decoder (maybe), mamba paper learning
Zehua Liu
  • read papper
  • auxiliary loss code[7]
Pengqi Li
  • Summary[8] of speech processing XAI for NSFC
    • polish, reference, v1(90%)
  • Workshop report(video, slide, poster)
Wan Lin
  • Graduation paper
Tianhao Wang
  • EA-ASP (from SJTU) reproduced successfully [9]
    • EA-ASP implement, wespeaker toolkit modification, training pairs construction totally according to the paper
    • get the better EER(4.021%) comparing to the paper (5.212%) on Vox1-O-Overlap
    • evaluation on our trials (concat and weak_overlap are better, overlap and mix are worse)
Zhenyu Zhou
  • Finish NeuralScoring baseline[10]
  • ICASSP2024 report
Junhui Chen
  • Neural scoring with Frequency-channel attention
    • Our overlap test trial: EER 7.382% -> 7.132%
  • Graduation paper
Jiaying Wang
  • HHuawei project:train speakerbeam on -4,4,converge,some bug in test
  • cohort mini loss: check code & test
Yu Zhang
  • Financial Backtesting pipline
    • remake Stock and Industry return logic
    • debug Brinson Analysis result
  • Use SAC as a baseline to run through the entire process from training to policy generation to backtesting
Wenqiang Du
  • Some model update task
    • Chinese and Uyghur
Yang Wei
  • Fix problems about ASR model for mispronunciation detection task
  • Prepare the baseline system
Lily
  • Data Analysis
  • Paper Reading[11]
Qi Qu
  • First day
  • Dev environment setup
Turi
  • Prepared sentences
  • Prepared data collection app for release