“2025-03-03”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
第55行: 第55行:
 
* Double check the related experimental code.
 
* Double check the related experimental code.
 
** MT-HuBERT(in progress) & Cocktail-HuBERT need re-pretrain.
 
** MT-HuBERT(in progress) & Cocktail-HuBERT need re-pretrain.
** The results of other baseline on [https://z1et6d3xtb.feishu.cn/docx/VThUd30RPoTBR4xOiKYc4gQTnsb here]
+
** The results of other baseline in [https://z1et6d3xtb.feishu.cn/docx/VThUd30RPoTBR4xOiKYc4gQTnsb here]
 
||
 
||
 
*
 
*

2025年3月3日 (一) 10:50的版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
Lantian Li
Ying Shi
  • Prepare Ascend Sever environment
  • training Conditional Chain overlap ASR model with Hierachical-Transformer here
Zhenghai You
Junming Yuan
  • Finish MPC-HuBERT pretrain.
  • Double check the related experimental code.
    • MT-HuBERT(in progress) & Cocktail-HuBERT need re-pretrain.
    • The results of other baseline in here
Xiaolou Li
Zehua Liu
  • Paper Reading and Sharing in last Friday
  • Writing Vision Language Model code
  • Writing NSFC document
Pengqi Li
  • Prepare the AI course for Tsinghua University Junior High School.
  • Using t-SNE to visualize the factorized content vector.
    • Next step is to color(speaker information importance or not) each point.
Wan Lin
  • try some methods for clean performance(no improvement)
  • supply experiments for other tests
Tianhao Wang
  • sound separation: 2-mix and 3-mix model training
  • weekly report
  • subset data training
Xiaoxue Luo
  • generation of multi-mix audio data and did some test experiments.
  • read papers
Zhenyu Zhou
Junhui Chen
  • Reproducing speaker diarization method for NS (debugging...)
  • read paper
Jiaying Wang
Yu Zhang
Wenqiang Du
  • Primary handbook's PPT (24/44)
  • Continue to check Primary and middle handbook(Completed this week)
  • Speech cloning sample for the company
Yang Wei
Turi
Yue Gu
  • finish some exps, but nothing is improved.
  • finish a proposal,I will present it recently
Qi Qu
  • Applying pre-prod eval routine on text-enroll KWS models: the ideal thresholds for each keyword vary significantly.