“2024-08-26”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
 
(9位用户的17个中间修订版本未显示)
第32行: 第32行:
 
|Ying Shi
 
|Ying Shi
 
||
 
||
*  
+
* Fenyinta stuff
 +
* reproduce cohort-SOT overlap ASR and some analysis
 +
* Text enroll keywords spotting intermediate PIT-SOT CTC + high layer cross-attention is in progress  [https://z1et6d3xtb.feishu.cn/docx/DI3UdF496ojxCQxTqUycsjDQnxf?from=from_copylink here]
 
||
 
||
 
*
 
*
第43行: 第45行:
 
|Zhenghai You
 
|Zhenghai You
 
||
 
||
*
+
* Speaker Augument: Completed experiments in Libri2mix, Low SISDR testset lower speaker confusion rate
 +
* ExFormer: Always inferior to the SOTA[https://z1et6d3xtb.feishu.cn/docx/ZbtsdTGuQo4IXnxuxHXcpvBynoe]
 
||
 
||
 
*
 
*
第56行: 第59行:
 
** To fully reproduce the official model, it would take approximately 32 days.
 
** To fully reproduce the official model, it would take approximately 32 days.
 
* Investigate how to train Hubert with Mix-speech (in progress)
 
* Investigate how to train Hubert with Mix-speech (in progress)
||
 
*
 
||
 
*
 
|-
 
 
 
|-
 
|Chen Chen
 
||
 
*
 
 
||
 
||
 
*
 
*
第77行: 第69行:
 
|Xiaolou Li
 
|Xiaolou Li
 
||
 
||
*
+
* LLM long context test
 +
* Poster for IS24
 +
* Paper reading
 
||
 
||
 
*
 
*
第108行: 第102行:
 
*
 
*
 
||
 
||
*
+
* 9.20(one month)
 
|-
 
|-
  
第115行: 第109行:
 
|Wan Lin
 
|Wan Lin
 
||
 
||
*
+
* Neural Scoring: vox2+voxblink1 [https://z1et6d3xtb.feishu.cn/docx/BywjdkGvNou12sxQ4dAcxYa9noh?from=from_copylink]
 
||
 
||
 
*
 
*
第131行: 第125行:
 
*
 
*
 
||
 
||
*
 
 
|-
 
|-
  
第138行: 第131行:
 
|Zhenyu Zhou
 
|Zhenyu Zhou
 
||
 
||
*
+
*Some thinking about onnx quantization[https://z1et6d3xtb.feishu.cn/docx/S9ChdyH7go490txZ2ZHcNjXTn2b]
 
||
 
||
 
*
 
*
第161行: 第154行:
 
|Jiaying Wang
 
|Jiaying Wang
 
||
 
||
*
+
* re-write conditional chain code(can be finished this week)
 +
* check wsj data
 +
 
 
||
 
||
 
*
 
*
第196行: 第191行:
 
|Yang Wei
 
|Yang Wei
 
||
 
||
*
+
* Check the badcase of KWS model test.
 
||
 
||
 
*
 
*
第226行: 第221行:
 
||
 
||
 
* write the introduction
 
* write the introduction
* test the adaptation model on the same accent data:
+
* test the adaptation model on the same accent data:[https://www.yuque.com/g/shibeiing/angax2/efpdhqvsxdi4phua/collaborator/join?token=ysvAeigXC9KFi4CF&source=doc_collaborator]
 
+
(got sick today)
 
||
 
||
 
*
 
*
第236行: 第231行:
 
|Qi Qu
 
|Qi Qu
 
||
 
||
*  
+
* KWS:
 +
** zh48 test dataset updated: 29 speakers in 3 locations, ~600 utterances per keyword.
 +
** Recall ~ FA relations plotted.
 
||
 
||
 
*
 
*

2024年8月26日 (一) 11:03的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • Primary school book (17)
  • College AI education
Lantian Li
  • GPU status [1]
  • AI primary
    • High school handbook (40/40)
Ying Shi
  • Fenyinta stuff
  • reproduce cohort-SOT overlap ASR and some analysis
  • Text enroll keywords spotting intermediate PIT-SOT CTC + high layer cross-attention is in progress here
Zhenghai You
  • Speaker Augument: Completed experiments in Libri2mix, Low SISDR testset lower speaker confusion rate
  • ExFormer: Always inferior to the SOTA[2]
Junming Yuan
  • Confirmed that the performance gap of the 10% is determined by the impact of GPUs[3].
    • To fully reproduce the official model, it would take approximately 32 days.
  • Investigate how to train Hubert with Mix-speech (in progress)
Xiaolou Li
  • LLM long context test
  • Poster for IS24
  • Paper reading
Zehua Liu
  • CNVSRC 2024 Website
  • Data transfer to HUAWEI
  • LLM in Chinese VSR(In-context-learning)
Pengqi Li
  • Extend Proposal for 'HOW PHONEMES CONTRIBUTE TO DEEP SPEAKER MODELS?'[4]
    • Reviewing code, paper.
    • Analyzing di-phones in Audio-Mnist.
    • Start exp with TIMIT dataset.
  • 9.20(one month)
Wan Lin
  • Neural Scoring: vox2+voxblink1 [5]
Tianhao Wang
  • AudioSep reproducing
  • IS24 poster
Zhenyu Zhou
  • Some thinking about onnx quantization[6]
Junhui Chen
  • Neural Scoring:
    • Vox2+Voxblink-clean test[7]
Jiaying Wang
  • re-write conditional chain code(can be finished this week)
  • check wsj data
Yu Zhang
  • AED engineering problem assist
  • Prepare for report
Wenqiang Du
  • Complete the unified format and recheck of Primary school handbook
  • Write middle school handbook(29-41)
  • Training Chinese and Cantonese KWS model
Yang Wei
  • Check the badcase of KWS model test.
Lily
Turi
  • Added more sections to the draft paper
    • Need to refine and do more experiments
Yue Gu
  • write the introduction
  • test the adaptation model on the same accent data:[8]

(got sick today)

Qi Qu
  • KWS:
    • zh48 test dataset updated: 29 speakers in 3 locations, ~600 utterances per keyword.
    • Recall ~ FA relations plotted.