“2026-01-05”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（11位用户的16个中间修订版本未显示）
第6行： 第6行：
 
|Dong Wang
 
|Dong Wang
 
||
 
||
−
*
+
* English version of the AI handbook  - middle/high school almost done
 +
 
 +
 
 
||
 
||
 
*
 
*
第41行： 第43行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*
+
* Some additional result analysis and some revisions on manuscript according to the reviewer' comments. [https://z1et6d3xtb.feishu.cn/docx/Woz0dFwsaoWZuOxZ2KGcm9MxnKe?from=from_copylink][https://z1et6d3xtb.feishu.cn/docx/VIZadovoZohLscxQUYNcu59HnGd?from=from_copylink]
 
||
 
||
 
*
 
*
第63行： 第65行：
 
|Pengqi Li
 
|Pengqi Li
 
||
 
||
−
*
+
* Draft Paper: The framework and core content are complete.
 +
* AI Handbook: Checking and verifying references.
 +
* Winter Camp Preparation: Preparing materials for the Tsinghua University Middle School AI Winter Camp.
 
||
 
||
 
*
 
*
第74行： 第78行：
 
|Junming Yuan
 
|Junming Yuan
 
||
 
||
−
*
+
* Aug-MT-HuBERT:
 +
** Fixed the pre-training to 100K steps, compared 3 noise ratios（50%, 80%, and 100%）and 2 different SNR sampling strategies([-3, 3] uniform / [0, 10] Gauss).
 +
*** Best setting, PR performance(PER): 8.12 -> 8.08, ASR performance(WER): 9.22 -> 9.07
 +
*** There is still around 3% gap compared to HuBERT.
 +
* SS adaptation(SI-SDRi):
 +
** Cocktail-HuBERT(14.76) > MT-HuBERT(13.93) > WavLM(13.83) > ConvTasNet(10.39)
 +
* Draft Paper( CN version almost done)
 
||
 
||
 
*
 
*
第85行： 第95行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*
+
* GPU Util [https://z1et6d3xtb.feishu.cn/wiki/XX4NwX3tJiBDcgkMi0hcFUtInHh?from=from_copylink]
 +
* working on the big assignment and preparing for the final exam.
 
||
 
||
 
*
 
*
第96行： 第107行：
 
|Junhui Chen
 
|Junhui Chen
 
||
 
||
−
*
+
* prepare for final exam
 
||
 
||
 
*
 
*
第107行： 第118行：
 
|Xiaolou Li
 
|Xiaolou Li
 
||
 
||
−
*
+
* Prepared applications for 4 CDT programmes
 
||
 
||
 
*  
 
*  
第129行： 第140行：
 
|Tianhao Wang
 
|Tianhao Wang
 
||
 
||
−
*
+
* additional experiments for ChainSep [https://z1et6d3xtb.feishu.cn/docx/RG5Qd9ashoGhw5xc9M2cAVcHnLc]
 
||
 
||
 
*
 
*
第140行： 第151行：
 
|Xiaoxue Luo
 
|Xiaoxue Luo
 
||
 
||
−
*
+
* 2-5mix attractor-based USS model for Huawei project
 +
** SI-SDR : 2mix: 4.725，  3mix: 0.866，  4mix: -1.155，  5mix: -2.709
 +
* design a better solution for Huawei project
 +
** multi-head separation (two groups: speakers and events)
 
||
 
||
 
*
 
*
第173行： 第187行：
 
|Bochao Hu
 
|Bochao Hu
 
||
 
||
−
*
+
* P2S wer: 45%.data:20% GT 20% real seq 60% synthetic seq. llm can't correct with high per(> 30%)
 
||
 
||
 
*
 
*
第184行： 第198行：
 
|Hongcheng Zhang
 
|Hongcheng Zhang
 
||
 
||
−
* finish code of ASU-LLM (CLAP + HuBERT)
+
* finish the main code of ASU-LLM (CLAP + HuBERT). The core architecture consists of a dual-stream encoder combining CLAP and Audio Hubert, integrated with LLaMA
 
||
 
||
 
*
 
*

2026年1月5日 (一) 11:36的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • English version of the AI handbook - middle/high school almost done


Lantian Li
  • Final review of my MLA book (3/10)
  • MoE daily work
Wenqiang Du
  • Check AI course PPT(4/14)
  • Completed AI100 question task and will be uploaded to the platform this month
Yang Wei
  • Some additional result analysis and some revisions on manuscript according to the reviewer' comments. [1][2]
Ying Shi
Pengqi Li
  • Draft Paper: The framework and core content are complete.
  • AI Handbook: Checking and verifying references.
  • Winter Camp Preparation: Preparing materials for the Tsinghua University Middle School AI Winter Camp.
Junming Yuan
  • Aug-MT-HuBERT:
    • Fixed the pre-training to 100K steps, compared 3 noise ratios（50%, 80%, and 100%）and 2 different SNR sampling strategies([-3, 3] uniform / [0, 10] Gauss).
      • Best setting, PR performance(PER): 8.12 -> 8.08, ASR performance(WER): 9.22 -> 9.07
      • There is still around 3% gap compared to HuBERT.
  • SS adaptation(SI-SDRi):
    • Cocktail-HuBERT(14.76) > MT-HuBERT(13.93) > WavLM(13.83) > ConvTasNet(10.39)
  • Draft Paper( CN version almost done)
Yu Zhang
  • GPU Util [3]
  • working on the big assignment and preparing for the final exam.
Junhui Chen
  • prepare for final exam
Xiaolou Li
  • Prepared applications for 4 CDT programmes
Jiaying Wang
  • spk order separation training
Tianhao Wang
  • additional experiments for ChainSep [4]
Xiaoxue Luo
  • 2-5mix attractor-based USS model for Huawei project
    • SI-SDR : 2mix: 4.725， 3mix: 0.866， 4mix: -1.155， 5mix: -2.709
  • design a better solution for Huawei project
    • multi-head separation (two groups: speakers and events)
Yue Gu
  • write Phd defense slides
Lily
Bochao Hu
  • P2S wer: 45%.data:20% GT 20% real seq 60% synthetic seq. llm can't correct with high per(> 30%)
Hongcheng Zhang
  • finish the main code of ASU-LLM (CLAP + HuBERT). The core architecture consists of a dual-stream encoder combining CLAP and Audio Hubert, integrated with LLaMA
取自“http://cslt.org/mediawiki/index.php?title=2026-01-05&oldid=43799”