“2024-06-17”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（15位用户的21个中间修订版本未显示）
第11行： 第11行：
 
* Refine "the principle of AI education in primary schools"
 
* Refine "the principle of AI education in primary schools"
 
* A few public talks
 
* A few public talks
−
 
 
||
 
||
 
*
 
*
第22行： 第21行：
 
|Lantian Li
 
|Lantian Li
 
||
 
||
−
*
+
* GPU status [https://z1et6d3xtb.feishu.cn/wiki/XGcGwRK5viJmpRkjH9AczIhynCh]
 +
* Completed all teaching for this semester.
 +
* Projects
 +
** AED -> System integration with Huawei, Bullying patent for FYT
 +
** TSE -> Preparing for the first phase delivery.
 +
** VSR -> 1200h+
 +
** Finance -> Reproducing R^2 SAC, Overview time-series modelling
 +
* Papers
 +
** NeuralScoring -> In progress
 +
** IS24 Camera-ready paper
 +
** CNVSRC 2024 baseline paper
 
||
 
||
 
*
 
*
第33行： 第42行：
 
|Ying Shi
 
|Ying Shi
 
||
 
||
−
*
+
* Text enroll mix speech keyword spotting
 +
* Cohort ASR with conditional Chain [https://z1et6d3xtb.feishu.cn/docx/VXhcdlaLto5HT1xR8Fec5Dtsnxh?from=from_copylink here]
 
||
 
||
 
*
 
*
第44行： 第54行：
 
|Zhenghai You
 
|Zhenghai You
 
||
 
||
−
*
+
* Start training on complete data for Huawei TSE project
 +
* Change adaptlayer from cancat to FiLm [https://z1et6d3xtb.feishu.cn/docx/U8CmdZfKzowpgtxzKuvczUCKnnh]
 +
* Test inference time and SISDR in online model
 +
* Consider a TSE network that combines mixture and enrollment with attractor to extract speaker information
 
||
 
||
 
*
 
*
第65行： 第78行：
 
|Chen Chen
 
|Chen Chen
 
||
 
||
−
*
+
* Release UY/CH-CHILD dataset
 +
* help with NCMMSC paper about CNVSRC 2024
 +
* CN-CVS2
 +
** 1200+ hours data, phase 1 finished in June 26th
 +
** hand over the CN-CVS2 website things
 
||
 
||
−
*
+
* ISCSLP paper (7.7 20:00 ddl)
 
||
 
||
 
*
 
*
第76行： 第93行：
 
|Xiaolou Li
 
|Xiaolou Li
 
||
 
||
−
*
+
* LRS2 full test[https://z1et6d3xtb.feishu.cn/docx/MjMpdxyjAoK5I7xuwThcqdfkngd#LlBLdS9qXoCAGuxahHScL0BInPe]
 +
* Paper reading
 
||
 
||
 
*
 
*
第87行： 第105行：
 
|Zehua Liu
 
|Zehua Liu
 
||
 
||
−
*
+
* NCMMSC papper
 +
* change parameter result seems good ,(but still training)[https://z1et6d3xtb.feishu.cn/docx/ZaTFd3A5EoK982xWBVschloanee?from=from_copylink]
 
||
 
||
 
*
 
*
第98行： 第117行：
 
|Pengqi Li
 
|Pengqi Li
 
||
 
||
−
*
+
* PhD mid-term assessment
 +
* two NC-papers
 
||
 
||
 
*
 
*
第120行： 第140行：
 
|Tianhao Wang
 
|Tianhao Wang
 
||
 
||
−
*
+
* Neural Scoring exps[https://z1et6d3xtb.feishu.cn/docx/BywjdkGvNou12sxQ4dAcxYa9noh]
 +
** share encoder
 +
** channel attention (similar to EA-ASP, useless)
 +
** early frequency attention (fbank level, training)
 
||
 
||
 
*
 
*
第131行： 第154行：
 
|Zhenyu Zhou
 
|Zhenyu Zhou
 
||
 
||
−
*
+
*Huawei projetc
 +
**Summary of recent experimental results[https://z1et6d3xtb.feishu.cn/docx/U8CmdZfKzowpgtxzKuvczUCKnnh]
 
||
 
||
 
*
 
*
第142行： 第166行：
 
|Junhui Chen
 
|Junhui Chen
 
||
 
||
−
*
+
* Neural Scoring supplementary experiments
 +
** Share Encoder NS: NS > Share Encoder NS > EA-ASP (Importance of decoupling)
 +
** Ways of attention (F-bank Enroll-Aware, seems useful)
 
||
 
||
 
*
 
*
第153行： 第179行：
 
|Jiaying Wang
 
|Jiaying Wang
 
||
 
||
−
*
+
* debug cohort transformer structure (confused why transformer does not work)
 +
** deeper network: 2 attention head, 8 layer/block, 4 blocks in total(failed)
 +
** use only MF training set (failed)
 +
** use position encoding and transformer block in speechbrain(failed both pit and sisdr loss)
 +
 
 
||
 
||
 
*
 
*
第164行： 第194行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*
+
* Implement R2SAC
 +
* Retrain Huawei Quantization Model
 +
* Paper reading
 
||
 
||
 
*
 
*
第187行： 第219行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*
+
* Huilan TTS
 +
** Export ONNX model from original format. Still deal with inferring error.
 
||
 
||
 
*
 
*
第216行： 第249行：
 
|Yue Gu
 
|Yue Gu
 
||
 
||
−
*
+
* try to fill the gap between CEM recall and utterance recall, then I want achieve more better performance
 
||
 
||
 
*
 
*
第225行： 第258行：
 
|Qi Qu
 
|Qi Qu
 
||
 
||
−
*  
+
* AED
 +
** Model tested on different data.
 +
** Tried some other models, i.e. Zipformer from sherpa-onnx.
 +
* KWS
 +
** Data collected and processed to account for poor performance.
 
||
 
||
 
*
 
*

2024年6月17日 (一) 12:59的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • Review of a few papers from NCMMSC, MDPI etc.
  • Review papers regarding AI for Medicine
  • Refine "the principle of AI education in primary schools"
  • A few public talks
Lantian Li
  • GPU status [1]
  • Completed all teaching for this semester.
  • Projects
    • AED -> System integration with Huawei, Bullying patent for FYT
    • TSE -> Preparing for the first phase delivery.
    • VSR -> 1200h+
    • Finance -> Reproducing R^2 SAC, Overview time-series modelling
  • Papers
    • NeuralScoring -> In progress
    • IS24 Camera-ready paper
    • CNVSRC 2024 baseline paper
Ying Shi
  • Text enroll mix speech keyword spotting
  • Cohort ASR with conditional Chain here
Zhenghai You
  • Start training on complete data for Huawei TSE project
  • Change adaptlayer from cancat to FiLm [2]
  • Test inference time and SISDR in online model
  • Consider a TSE network that combines mixture and enrollment with attractor to extract speaker information
Junming Yuan
  • SSL model finetuning analysis v1[3].Need to check.
Chen Chen
  • Release UY/CH-CHILD dataset
  • help with NCMMSC paper about CNVSRC 2024
  • CN-CVS2
    • 1200+ hours data, phase 1 finished in June 26th
    • hand over the CN-CVS2 website things
  • ISCSLP paper (7.7 20:00 ddl)
Xiaolou Li
  • LRS2 full test[4]
  • Paper reading
Zehua Liu
  • NCMMSC papper
  • change parameter result seems good ,(but still training)[5]
Pengqi Li
  • PhD mid-term assessment
  • two NC-papers
Wan Lin
Tianhao Wang
  • Neural Scoring exps[6]
    • share encoder
    • channel attention (similar to EA-ASP, useless)
    • early frequency attention (fbank level, training)
Zhenyu Zhou
  • Huawei projetc
    • Summary of recent experimental results[7]
Junhui Chen
  • Neural Scoring supplementary experiments
    • Share Encoder NS: NS > Share Encoder NS > EA-ASP (Importance of decoupling)
    • Ways of attention (F-bank Enroll-Aware, seems useful)
Jiaying Wang
  • debug cohort transformer structure (confused why transformer does not work)
    • deeper network: 2 attention head, 8 layer/block, 4 blocks in total(failed)
    • use only MF training set (failed)
    • use position encoding and transformer block in speechbrain(failed both pit and sisdr loss)
Yu Zhang
  • Implement R2SAC
  • Retrain Huawei Quantization Model
  • Paper reading
Wenqiang Du
  • Preparing for the final exam
Yang Wei
  • Huilan TTS
    • Export ONNX model from original format. Still deal with inferring error.
Lily
  • Thesis
  • Prepare slides for Xinjiang teacher's course
Turi
  • End of semester course project presentations
Yue Gu
  • try to fill the gap between CEM recall and utterance recall, then I want achieve more better performance
Qi Qu
  • AED
    • Model tested on different data.
    • Tried some other models, i.e. Zipformer from sherpa-onnx.
  • KWS
    • Data collected and processed to account for poor performance.
取自“http://index.cslt.org/mediawiki/index.php?title=2024-06-17&oldid=41873”