“2025-02-17”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（17位用户的24个中间修订版本未显示）
第6行： 第6行：
 
|Dong Wang
 
|Dong Wang
 
||
 
||
−
*  
+
* Middle-School education PPT
  
 
||
 
||
第18行： 第18行：
 
|Lantian Li
 
|Lantian Li
 
||
 
||
−
*  
+
* Final version proofreading of the high-school book (17/40)
 +
* Polish IS2025 papers.
 
||
 
||
 
*
 
*
第29行： 第30行：
 
|Ying Shi
 
|Ying Shi
 
||
 
||
−
*  
+
* Conditional chain overlap asr
 +
** pooling  order
 +
** padding chain
 +
** augmentation padding
 +
** [https://z1et6d3xtb.feishu.cn/docx/Wg3ldbCKeoKhBzxZ2lccTO8ZnTf?from=from_copylink here]
 
||
 
||
 
*  
 
*  
第41行： 第46行：
 
|Zhenghai You
 
|Zhenghai You
 
||
 
||
−
*  
+
* Completed the revision of the second edition of the paper
 
||
 
||
 
*
 
*
第53行： 第58行：
 
* The results of MT-Hubert/Cocktail-Hubert/Hubert on LS960[https://z1et6d3xtb.feishu.cn/docx/DRICd10VCodDfOxXbDMcI97snod]
 
* The results of MT-Hubert/Cocktail-Hubert/Hubert on LS960[https://z1et6d3xtb.feishu.cn/docx/DRICd10VCodDfOxXbDMcI97snod]
 
**15-shot finetuning(clean/Mixup/MT):MT-HuBERT > Cocktail-Hubert > Hubert
 
**15-shot finetuning(clean/Mixup/MT):MT-HuBERT > Cocktail-Hubert > Hubert
−
** MPC-Hubert is in progress
+
** MPC-Hubert is still in the training queue
 
||
 
||
 
*
 
*
第64行： 第69行：
 
|Xiaolou Li
 
|Xiaolou Li
 
||
 
||
−
*  
+
* Paper modification
 +
* Data processing (2000h / 4000h)
 +
* Some new demand in data collection server
 +
* Paper reading
 
||
 
||
 
*  
 
*  
第75行： 第83行：
 
|Zehua Liu
 
|Zehua Liu
 
||
 
||
−
*
+
*CNVSRC 2024 and VSR-LLM paper writing and revise, and do some relevant Experiments.
 
||
 
||
 
*
 
*
第86行： 第94行：
 
|Pengqi Li
 
|Pengqi Li
 
||
 
||
−
*  
+
* XAI of Speaker Verification[https://z1et6d3xtb.feishu.cn/docx/QnZvdrNAboVCCVxgVRzcpsgRn6d]
 +
** some experiment looks like successful
 +
** Writing paper same time(30%)
 +
*Report recently work on Friday
 
||
 
||
 
*
 
*
第97行： 第108行：
 
|Wan Lin
 
|Wan Lin
 
||
 
||
−
*  
+
* alter training strategy to 2-positive multi-enroll for mix-training
 +
* find and fix some tiny bugs
 +
* results [https://z1et6d3xtb.feishu.cn/docx/MxBNdPbLao0tsoxkBVCcUgUoneh?from=from_copylink]
 
||
 
||
 
*
 
*
第108行： 第121行：
 
|Tianhao Wang
 
|Tianhao Wang
 
||
 
||
−
*  
+
* filter the label for AudioSet data (use fine-grained label and avoid duplicated separation)
 +
* import CED model to the training code
 
||
 
||
 
*
 
*
第119行： 第133行：
 
|Xiaoxue Luo
 
|Xiaoxue Luo
 
||
 
||
−
*  
+
* Check the pictures in AI high school handbook,will be completed in these days
 +
* The learning and production of AI daily signature
 
||
 
||
 
*
 
*
第130行： 第145行：
 
|Zhenyu Zhou
 
|Zhenyu Zhou
 
||
 
||
−
*
+
* Graduation article
 +
* interspeech paper with zhenghai
 +
* code double check with jiaying
 
||
 
||
 
*
 
*
第141行： 第158行：
 
|Junhui Chen
 
|Junhui Chen
 
||
 
||
−
*
+
* Read paper about e2e SV / loss (little useful content for NS).
 +
* Continue to think and try other tricks, but no meaningful results yet.
 
||
 
||
 
*
 
*
第152行： 第170行：
 
|Jiaying Wang
 
|Jiaying Wang
 
||
 
||
−
*
+
* conditional chain with ctc experiment
 +
** loss decreased slowly in the previous few epochs, almost stop at around -3
 +
** checked the model code: no bug
 +
** test results of these epoch to determine the problem
 
||
 
||
 
*
 
*
第163行： 第184行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*  
+
* Add Backtest report to debate (no improvement)
 +
* Change LLM to DeepSeek R1 70B
 +
** Sharpe Ratio 0.401 (from 0.256), S&P 500 is 0.576.
 +
* Add more quantile factor and information source (still working on this)
 
||
 
||
 
*
 
*
第174行： 第198行：
 
|Wenqiang Du
 
|Wenqiang Du
 
||
 
||
−
*  
+
*Continue to check AI primary handbook（done）
 +
*Continue to check AI middle handbook（done）
 
||
 
||
 
*
 
*
第185行： 第210行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*  
+
* Finetune text enroll kws model with different accent keyword data.
 
||
 
||
 
*
 
*
第195行： 第220行：
 
|Turi
 
|Turi
 
||
 
||
−
*  
+
* Thesis writing
 +
* Trained the language model, some bugs when decoding with LM
 
||
 
||
 
*  
 
*  
第203行： 第229行：
 
|Yue Gu
 
|Yue Gu
 
||
 
||
−
*  
+
* fine-grained personality-gating method：finish the code and model training，ready to test
 
||
 
||
 
*
 
*
第212行： 第238行：
 
|Qi Qu
 
|Qi Qu
 
||
 
||
−
*  
+
* Successful porting of text-enroll KWS models (hybrid quantization) to mr536 w/ low precision loss, same RT as previous version (which shows drastic precision loss). [https://b30lttjm7l.feishu.cn/docx/WORcdiE1io86Agxg9hOcKIO0nwe?from=from_copylink]
 
||
 
||
 
*
 
*

2025年2月17日 (一) 11:00的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • Middle-School education PPT
Lantian Li
  • Final version proofreading of the high-school book (17/40)
  • Polish IS2025 papers.
Ying Shi
  • Conditional chain overlap asr
    • pooling order
    • padding chain
    • augmentation padding
    • here
Zhenghai You
  • Completed the revision of the second edition of the paper
Junming Yuan
  • The results of MT-Hubert/Cocktail-Hubert/Hubert on LS960[1]
    • 15-shot finetuning(clean/Mixup/MT):MT-HuBERT > Cocktail-Hubert > Hubert
    • MPC-Hubert is still in the training queue
Xiaolou Li
  • Paper modification
  • Data processing (2000h / 4000h)
  • Some new demand in data collection server
  • Paper reading
Zehua Liu
  • CNVSRC 2024 and VSR-LLM paper writing and revise, and do some relevant Experiments.
Pengqi Li
  • XAI of Speaker Verification[2]
    • some experiment looks like successful
    • Writing paper same time(30%)
  • Report recently work on Friday
Wan Lin
  • alter training strategy to 2-positive multi-enroll for mix-training
  • find and fix some tiny bugs
  • results [3]
Tianhao Wang
  • filter the label for AudioSet data (use fine-grained label and avoid duplicated separation)
  • import CED model to the training code
Xiaoxue Luo
  • Check the pictures in AI high school handbook,will be completed in these days
  • The learning and production of AI daily signature
Zhenyu Zhou
  • Graduation article
  • interspeech paper with zhenghai
  • code double check with jiaying
Junhui Chen
  • Read paper about e2e SV / loss (little useful content for NS).
  • Continue to think and try other tricks, but no meaningful results yet.
Jiaying Wang
  • conditional chain with ctc experiment
    • loss decreased slowly in the previous few epochs, almost stop at around -3
    • checked the model code: no bug
    • test results of these epoch to determine the problem
Yu Zhang
  • Add Backtest report to debate (no improvement)
  • Change LLM to DeepSeek R1 70B
    • Sharpe Ratio 0.401 (from 0.256), S&P 500 is 0.576.
  • Add more quantile factor and information source (still working on this)
Wenqiang Du
  • Continue to check AI primary handbook（done）
  • Continue to check AI middle handbook（done）
Yang Wei
  • Finetune text enroll kws model with different accent keyword data.
Turi
  • Thesis writing
  • Trained the language model, some bugs when decoding with LM
Yue Gu
  • fine-grained personality-gating method：finish the code and model training，ready to test
Qi Qu
  • Successful porting of text-enroll KWS models (hybrid quantization) to mr536 w/ low precision loss, same RT as previous version (which shows drastic precision loss). [4]
取自“http://cslt.org/mediawiki/index.php?title=2025-02-17&oldid=42843”