“2026-01-26”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
第100行： 第100行：
 
** PR(PER%), LNTL-MT-HuBERT: 7.24%, Clean-HuBERT: 5.41%, WavLM: 4.84%
 
** PR(PER%), LNTL-MT-HuBERT: 7.24%, Clean-HuBERT: 5.41%, WavLM: 4.84%
 
** ASR(WER%), LNTL-MT-HuBERT: 8.43%, Clean-HuBERT: 6.42%, WavLM: 6.21%
 
** ASR(WER%), LNTL-MT-HuBERT: 8.43%, Clean-HuBERT: 6.42%, WavLM: 6.21%
−
** SD(DER%), LNTL-MT-HuBERT: 4.31%, Cocktail-HuBERT: 4.07%
+
** SD(DER%), LNTL-MT-HuBERT: 4.31%, Cocktail-HuBERT: 4.07%, MT-HuBERT: 3.93%
 
* revise ICASSP final paper
 
* revise ICASSP final paper
 
||
 
||

2026年1月26日 (一) 10:41的版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • AI middle school hand book / English version almost completed.
Lantian Li
  • Recheck AI textbooks for senior middle school.
  • MoE daily work (busy week)
  • Go on the final review of my MLA book (6/10)
Wenqiang Du
  • Project work about AIbabel
    • Upgraded online model
    • New demand technology survey
Yang Wei
Ying Shi
Yue Gu
  • write my thesis
Lily
Pengqi Li
Junming Yuan
  • learn-not-to-listen MT-HuBERT
    • PR(PER%), LNTL-MT-HuBERT: 7.24%, Clean-HuBERT: 5.41%, WavLM: 4.84%
    • ASR(WER%), LNTL-MT-HuBERT: 8.43%, Clean-HuBERT: 6.42%, WavLM: 6.21%
    • SD(DER%), LNTL-MT-HuBERT: 4.31%, Cocktail-HuBERT: 4.07%, MT-HuBERT: 3.93%
  • revise ICASSP final paper
Yu Zhang
  • GPU Util: [1]
  • LLM
    • Completed the code for Swarm MMLU metric-based optimization together with @chenjunhui
    • Still working on modifying the Swarm topology code (for more complex swarm topology)
Junhui Chen
  • (caught a cold)
  • MMLU
    • difference in the ECS distribution between good nodes and bad nodes.
    • metrics loss helps select high-quality nodes and eliminate low-quality ones. [2]
Jiaying Wang
Xiaoxue Luo
  • prepare 2mix speech separation results for ASR test
  • 2-5mix multi_head separation model for Huawei project
    • data: remove data that the target audio is mixed audio(input equals output, model does not learn)
    • loss: add MSE_loss to the original SISDR_loss to control the energy amplitude of output audio
Bochao Hu
Hongcheng Zhang
Weiman Sun
  • Ran a completely opposite ASR task on the same data
    • Consistent with last week's result
  • Analyze different models
取自“http://index.cslt.org/mediawiki/index.php?title=2026-01-26&oldid=43902”