“2024-03-18”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
 
(10位用户的13个中间修订版本未显示)
第9行: 第9行:
 
* Design/Discussion AI popular science
 
* Design/Discussion AI popular science
 
* Conjecture for minmum loss training
 
* Conjecture for minmum loss training
 
 
||
 
||
 
*
 
*
第20行: 第19行:
 
|Lantian Li
 
|Lantian Li
 
||
 
||
*  
+
* GPU status [https://z1et6d3xtb.feishu.cn/wiki/XGcGwRK5viJmpRkjH9AczIhynCh]
 +
* INTERSPEECH 2024
 +
* ASIP-BUPT (CohortTSE, SE-Adapter, SpeakerAug, NeuralScoring)
 
||
 
||
 
*  
 
*  
第51行: 第52行:
 
* Some evaluations about TSE speaker encoder
 
* Some evaluations about TSE speaker encoder
 
* Huawei project (Phase 1st)
 
* Huawei project (Phase 1st)
* Some doubts about the paper due to the latest testing
+
* Some doubts about the paper due to the latest testing in minimum loss
 
||
 
||
 
* Change the speakerbeam speaker encoder to frequency domain
 
* Change the speakerbeam speaker encoder to frequency domain
第79行: 第80行:
 
|Chen Chen
 
|Chen Chen
 
||  
 
||  
*  
+
* Finish IS24 paper
 +
* Some documents for VTS X project
 +
* Proposal for next stage work on VSR/VTS
 +
** Focus on two task: 1) CNCVS2 dataset 2) Mandarin VSR Benchmark [https://z1et6d3xtb.feishu.cn/docx/PUNcdn0mZoYciuxhNogcu1TMnQd?from=from_copylink] on CNCVS1&2&CNVSRC
 +
** Aim at a solid benchmark with data/code/model
 +
** Perhaps a long journal paper
 +
||
 +
* Conditional entropy analysis of VTS task
 +
** MFA is done
 +
** TODOs: feature/embedding extracting, clustering, discrete conditional entropy calculating
 
||
 
||
 
*  
 
*  
||
 
 
 
|-
 
|-
  
第90行: 第98行:
 
|Xiaolou Li
 
|Xiaolou Li
 
||  
 
||  
*  
+
* Finish INTERSPEECH2024 paper
 +
* review code of cnvsrc
 +
* Next step:
 +
** Focus on model structure of VSR Benchmark
 
||
 
||
 
*  
 
*  
第101行: 第112行:
 
|Zehua Liu
 
|Zehua Liu
 
||  
 
||  
*
+
*Finish IS24
 +
*VSR work continues
 
||
 
||
 
*  
 
*  
第112行: 第124行:
 
|Pengqi Li
 
|Pengqi Li
 
||   
 
||   
*  
+
* Extending workshop paper
 +
** Finish slide for workshop paper.
 +
** make plan, investigate, prepare dataset for extending paper.
 +
** Rethink how to design a method that can globally PID
 +
* Team Working[https://z1et6d3xtb.feishu.cn/docx/T3U2dTs5poiIgtxtM2Sc0QennWe]
 
||
 
||
 
*
 
*
第123行: 第139行:
 
|Wan Lin
 
|Wan Lin
 
||  
 
||  
*  
+
* Neural scoring [https://z1et6d3xtb.feishu.cn/docx/TQvWdk8LVo9ONaxQ5Qac9A2Dn3d?from=from_copylink]
 
||
 
||
 
*
 
*
第159行: 第175行:
 
|Junhui Chen
 
|Junhui Chen
 
||
 
||
*  
+
* Neural scoring
 +
* Interim report
 
||
 
||
 
*
 
*
第170行: 第187行:
 
|Jiaying Wang
 
|Jiaying Wang
 
||  
 
||  
*  
+
* weekly report
 +
* PIT baseline: ConTasNet (finish tonight)
 +
* test whether the separation target is the closer one to the cohort embedding: the rate is around 0.5
 +
** confused about the efficiency of cohort
 +
** Further experiment:TasNet with minimal loss
 
||
 
||
 
*  
 
*  
第206行: 第227行:
 
|Yang Wei
 
|Yang Wei
 
||  
 
||  
*  
+
* Read training code of Paraformer model, in order to get intermediate data
 +
* Prepare Huilan product training, and deal with problems of ASR and TTS service
 
||
 
||
 
*
 
*
第216行: 第238行:
 
|Lily
 
|Lily
 
||
 
||
*  
+
* Paper reading
 +
* Prepare for overview paper
 
||
 
||
 
*
 
*

2024年3月18日 (一) 11:39的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • Interspeech 2024 paper refinement
  • Design/Discussion AI popular science
  • Conjecture for minmum loss training
Lantian Li
  • GPU status [1]
  • INTERSPEECH 2024
  • ASIP-BUPT (CohortTSE, SE-Adapter, SpeakerAug, NeuralScoring)
Ying Shi
  • Finish INTERSPEECH paper
  • Investigate random order SOT for multi-talker ASR task
  • 3-mix 0s offset test condition
    • DOM-SOT 20.51
    • PIT-SOT 23.26
    • random-order SOT 26.20
  • group work
Zhenghai You
  • Weekly report
  • Some evaluations about TSE speaker encoder
  • Huawei project (Phase 1st)
  • Some doubts about the paper due to the latest testing in minimum loss
  • Change the speakerbeam speaker encoder to frequency domain
  • Train a SID with a speakerbeam structure
Junming Yuan
  • Finish INTERSPEECH paper
  • Make the plan for the large vocabulary pretraining task.
    • Focus on the experimental details of the few-shot paper from Google.
    • Try to address the 3 questions:
      • How to change MT pretraining model structure?
      • How to train three strictly comparable pretraining models based on MT, Hubert, and wav2vec?
      • Why does Hubert+MT perform significantly better?
Chen Chen
  • Finish IS24 paper
  • Some documents for VTS X project
  • Proposal for next stage work on VSR/VTS
    • Focus on two task: 1) CNCVS2 dataset 2) Mandarin VSR Benchmark [2] on CNCVS1&2&CNVSRC
    • Aim at a solid benchmark with data/code/model
    • Perhaps a long journal paper
  • Conditional entropy analysis of VTS task
    • MFA is done
    • TODOs: feature/embedding extracting, clustering, discrete conditional entropy calculating
Xiaolou Li
  • Finish INTERSPEECH2024 paper
  • review code of cnvsrc
  • Next step:
    • Focus on model structure of VSR Benchmark
Zehua Liu
  • Finish IS24
  • VSR work continues
Pengqi Li
  • Extending workshop paper
    • Finish slide for workshop paper.
    • make plan, investigate, prepare dataset for extending paper.
    • Rethink how to design a method that can globally PID
  • Team Working[3]
Wan Lin
  • Neural scoring [4]
Tianhao Wang
  • Finish INTERSPEECH paper
  • Code reorganization
Zhenyu Zhou
  • InterSpeech2024 submission
  • Code reorganization
  • Neuro scoring reviewing
Junhui Chen
  • Neural scoring
  • Interim report
Jiaying Wang
  • weekly report
  • PIT baseline: ConTasNet (finish tonight)
  • test whether the separation target is the closer one to the cohort embedding: the rate is around 0.5
    • confused about the efficiency of cohort
    • Further experiment:TasNet with minimal loss
Yu Zhang
  • Portfolio backtesting report
  • stock trade API
Wenqiang Du
  • Aibabel
    • Control Uyghur KWS model FA,but not get a good performance yet.
    • Continue test and update CN KWS model
Yang Wei
  • Read training code of Paraformer model, in order to get intermediate data
  • Prepare Huilan product training, and deal with problems of ASR and TTS service
Lily
  • Paper reading
  • Prepare for overview paper
Turi
  • Data collection app[5]
  • Course works