“2024-12-30”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（10位用户的15个中间修订版本未显示）
第6行： 第6行：
 
|Dong Wang
 
|Dong Wang
 
||
 
||
−
*
+
* AI handbook check & polish (primary version)
 +
 
 
||
 
||
 
*
 
*
第17行： 第18行：
 
|Lantian Li
 
|Lantian Li
 
||
 
||
−
*
+
* Complete 2025 AI calendar
 
||
 
||
 
*
 
*
第28行： 第29行：
 
|Ying Shi
 
|Ying Shi
 
||
 
||
−
*  
+
* Cosine-Guided Order VS Dominance Guided Order
 +
** Token error rate and  CTC-loss Confusion Matrix [https://z1et6d3xtb.feishu.cn/docx/HFJkdFp5XoM4wkxKb3Ucevhxnfb?from=from_copylink]
 
||
 
||
−
*
+
* Design Learnable order
 +
* Think about how to use Condition ...
 
||
 
||
 
*
 
*
第39行： 第42行：
 
|Zhenghai You
 
|Zhenghai You
 
||
 
||
−
*
+
* Paper reading for weekly report
 +
* Supplement the SPK-AUG experiment and plan to start writing the paper[https://z1et6d3xtb.feishu.cn/docx/K4SYdNM2QoBFqOxFp7Hc8Z3Mnpg]
 +
* Attempt to increase SSL loss on existing E2E-TSE{The current results are not good}
 
||
 
||
 
*
 
*
第74行： 第79行：
 
|Zehua Liu
 
|Zehua Liu
 
||
 
||
−
*
+
*Train LRS3 exp for AlignVSR
 +
*Iter Inference 7-times get result (43.88%) better than No Iter-Inference (45.74%)
 +
*Use Weaker Encoder to generate corrupted text seems slightly better than before(43.88% < 44.54%)[https://z1et6d3xtb.feishu.cn/docx/JBsidACDVojhCaxFQLbcCVbsnAc?from=from_copylink]
 
||
 
||
 
*
 
*
第97行： 第104行：
 
|Wan Lin
 
|Wan Lin
 
||
 
||
−
*
+
* NS: margin BCE loss & multi-enroll training
 
||
 
||
 
*
 
*
第107行： 第114行：
 
|-
 
|-
 
|Tianhao Wang
 
|Tianhao Wang
 +
||
 +
* Try using attention instead of FiLM:
 +
** Cross-Attention: validation loss: cross-attn: -8.065 vs. FiLM: -9.986
 +
** Self-Attention: loss doesn‘t decrease
 
||
 
||
 
*
 
*
 +
||
 +
*
 +
|-
 +
 +
 +
|-
 +
|Xiaoxue Luo
 +
||
 +
* prepare for the final exam
 
||
 
||
 
*
 
*
第152行： 第172行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*
+
* Multi policy pipeline building (Done Tech/Sentiment policy generation)
 +
* Some copyright stuff in Royal Flush
 
||
 
||
 
*
 
*
第174行： 第195行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*
+
* Prepare an ASR Java jar file and API doc for zhongchuan
 
||
 
||
 
*
 
*
第184行： 第205行：
 
|Turi
 
|Turi
 
||
 
||
−
*
+
* Prepared finetuning code for MMS
 +
** MMS from meta outperforms whisper and they provide adapters for every language supported which include Oromo language.
 +
** Server is busy now, plan to do experiment on MMS
 +
* Preparing ppt for midterm defense
 
||
 
||
−
*
+
* Midterm defense
 
||
 
||
−
*
+
*  
 
|-
 
|-
 
|Yue Gu
 
|Yue Gu
第201行： 第225行：
 
|Qi Qu
 
|Qi Qu
 
||
 
||
−
*  
+
* Quantization for NPU: metrics updated [https://b30lttjm7l.feishu.cn/docx/WORcdiE1io86Agxg9hOcKIO0nwe].
 +
* CED + classifier used as VAD: speech detection during inactive hours in dormitories.
 +
* QAT (Quantization-Aware Training) exploration.
 
||
 
||
 
*
 
*

2024年12月30日 (一) 10:55的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • AI handbook check & polish (primary version)
Lantian Li
  • Complete 2025 AI calendar
Ying Shi
  • Cosine-Guided Order VS Dominance Guided Order
    • Token error rate and CTC-loss Confusion Matrix [1]
  • Design Learnable order
  • Think about how to use Condition ...
Zhenghai You
  • Paper reading for weekly report
  • Supplement the SPK-AUG experiment and plan to start writing the paper[2]
  • Attempt to increase SSL loss on existing E2E-TSE{The current results are not good}
Junming Yuan
  • MT-Hubert paper writing(1/2)
  • Check the high school AI handbook(237/362)
Xiaolou Li
  • Data process server code upgrade, data transfer
  • Experiment on 1500h (in training...)
  • Odds and ends...Paper reading, Coursework, Calendar delivery...
  • Upgrade the AV-HuBERT training code and test it
Zehua Liu
  • Train LRS3 exp for AlignVSR
  • Iter Inference 7-times get result (43.88%) better than No Iter-Inference (45.74%)
  • Use Weaker Encoder to generate corrupted text seems slightly better than before(43.88% < 44.54%)[3]
Pengqi Li
  • Conduct a preliminary experiment for the proposal(IS25-XAI) (A clear doc must be produced by the end of this week.)
  • Check the high school AI handbook(1/3)
Wan Lin
  • NS: margin BCE loss & multi-enroll training
Tianhao Wang
  • Try using attention instead of FiLM:
    • Cross-Attention: validation loss: cross-attn: -8.065 vs. FiLM: -9.986
    • Self-Attention: loss doesn‘t decrease
Xiaoxue Luo
  • prepare for the final exam
Zhenyu Zhou
Junhui Chen
Jiaying Wang
Yu Zhang
  • Multi policy pipeline building (Done Tech/Sentiment policy generation)
  • Some copyright stuff in Royal Flush
Wenqiang Du
  • Nothing~
Yang Wei
  • Prepare an ASR Java jar file and API doc for zhongchuan
Turi
  • Prepared finetuning code for MMS
    • MMS from meta outperforms whisper and they provide adapters for every language supported which include Oromo language.
    • Server is busy now, plan to do experiment on MMS
  • Preparing ppt for midterm defense
  • Midterm defense
Yue Gu
  • read several papers about synthetic data for personalized ASR and do some exps. plan to report them on this friday or next monday.
Qi Qu
  • Quantization for NPU: metrics updated [4].
  • CED + classifier used as VAD: speech detection during inactive hours in dormitories.
  • QAT (Quantization-Aware Training) exploration.
取自“http://cslt.org/mediawiki/index.php?title=2024-12-30&oldid=42687”