“2024-08-19”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（15位用户的25个中间修订版本未显示）
第6行： 第6行：
 
|Dong Wang
 
|Dong Wang
 
||
 
||
−
*
+
* AI primary (middle-school) 1-6
 
||
 
||
 
*
 
*
第17行： 第17行：
 
|Lantian Li
 
|Lantian Li
 
||
 
||
−
*
+
* GPU status [https://z1et6d3xtb.feishu.cn/wiki/XGcGwRK5viJmpRkjH9AczIhynCh]
 +
* AI primary
 +
** High school handbook (30/40)
 
||
 
||
−
*
+
* High school handbook (40/40)
 
||
 
||
−
*
+
*  
 
|-
 
|-
  
第39行： 第41行：
 
|Zhenghai You
 
|Zhenghai You
 
||
 
||
−
*
+
* Continue the work of speaker augument and Ex-Former in TSE[https://z1et6d3xtb.feishu.cn/docx/ZbtsdTGuQo4IXnxuxHXcpvBynoe]
 
||
 
||
 
*
 
*
第45行： 第47行：
 
*
 
*
 
|-
 
|-
 +
  
 
|-
 
|-
−
|Junming Yuan
+
|Jiaying Wang
 
||
 
||
−
* Verified two parameters in Hubert pre-training config file that were confused with the original paper.[https://z1et6d3xtb.feishu.cn/docx/PaATdHi26oEc0Pxovd4cSyp0nQ2]
+
* reproduce conditional chain code
−
** Confirmed that in the second iteration of pretraining, features should be extracted from the 6-th layer of the transformer, not the 9-th layer.
+
** on both libri and wsj: training loss hard to reduce (around -3) and the corresponding test sisdr is positive
−
*** in 175k step, result of 6-th layer: 71.55/9.39, result of 9-th layer: 37.31/16.72
+
* rewriting the code (preserving the original model)
−
** Basically confirmed the setting of the parameter 'untie_final_proj' for the two iterations of pre-training.
+
 
||
 
||
 
*
 
*
第61行： 第63行：
  
 
|-
 
|-
−
|Chen Chen
+
|Junming Yuan
 
||
 
||
−
*
+
* Verified two parameters in Hubert pretraining config file that were confused with the original paper.[https://z1et6d3xtb.feishu.cn/docx/PaATdHi26oEc0Pxovd4cSyp0nQ2]
 +
** Confirmed that in the second iteration of pretraining, features should be extracted from the 6-th layer of the transformer, not the 9-th layer.
 +
*** in 175k step, result of 6-th layer: 71.55/9.39, result of 9-th layer: 37.31/16.72
 +
** Basically confirmed the setting of the parameter 'untie_final_proj' for the two iterations of pretraining.
 
||
 
||
 
*
 
*
第96行： 第101行：
 
|Pengqi Li
 
|Pengqi Li
 
||
 
||
−
*
+
* Investigating Extremely Short-Utterance in speaker recognition[https://z1et6d3xtb.feishu.cn/docx/JBGhdZDJUoSDwGx3wbFcDk9NnLb]
 
||
 
||
 
*
 
*
第105行： 第110行：
  
 
|-
 
|-
−
|Wan Lin
+
|Tianhao Wang
 
||
 
||
−
*
+
* reproducing CLIPSep on two datasets: MUSIC and VGGSound [https://z1et6d3xtb.feishu.cn/docx/HyLGduUu2oixxJxaFdwcDaIsnhd]
 +
** MUSIC: Text query: 10.06 SDR, Image query: 12.13 SDR
 +
** VGGSound: Text query: 2.78 SDR, Image query: 5.01 SDR
 
||
 
||
 
*
 
*
第116行： 第123行：
  
 
|-
 
|-
−
|Tianhao Wang
+
|Zhenyu Zhou
 
||
 
||
−
*
+
*conditional chain reproduce
 +
*model quantify[https://z1et6d3xtb.feishu.cn/docx/S9ChdyH7go490txZ2ZHcNjXTn2b]
 
||
 
||
 
*
 
*
第127行： 第135行：
  
 
|-
 
|-
−
|Zhenyu Zhou
+
|Wan Lin
 
||
 
||
−
*
+
* VoxBlink1
 +
** Data processing
 +
** Baseline(ResNet34) training and NS training [https://z1et6d3xtb.feishu.cn/docx/BywjdkGvNou12sxQ4dAcxYa9noh?from=from_copylink]
 
||
 
||
 
*
 
*
第140行： 第150行：
 
|Junhui Chen
 
|Junhui Chen
 
||
 
||
−
*
+
* VoxBlink1
 +
** Data processing
 +
** Baseline(ResNet34 ASP) training and NS training [https://z1et6d3xtb.feishu.cn/docx/BywjdkGvNou12sxQ4dAcxYa9noh#Ro69dyERUoN0HvxOGMWcOJjrnuf]
 
||
 
||
 
*
 
*
第147行： 第159行：
 
|-
 
|-
  
−
 
−
|-
 
−
|Jiaying Wang
 
−
||
 
−
*
 
−
||
 
−
*
 
−
||
 
−
*
 
−
|-
 
  
  
第162行： 第164行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*
+
* AED engineering problem assist
 +
* paper reading (will report this FRI)
 
||
 
||
 
*
 
*
第173行： 第176行：
 
|Wenqiang Du
 
|Wenqiang Du
 
||
 
||
−
*
+
*Complete Primary school handbook draft (45 + 8)
 +
* Modify the format, expression, and distribution of knowledge points in the draft（40%）
 
||
 
||
 
*
 
*
第184行： 第188行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*
+
* Test KWS model on data v1.2.(https://z1et6d3xtb.feishu.cn/docx/Jkv2dnZVRo12eKx65tLc1vctnfb)
 
||
 
||
 
*
 
*
第204行： 第208行：
 
|Turi
 
|Turi
 
||
 
||
−
*
+
* Traveled back home and took rest
 +
* Now writing dataset paper.
 +
**Done with Intro, Literature review, Data collection sections. Experiment, Result and Conclusion sections remaining.
 +
* Wasn't able to do more experiment on dataset from Ethiopia due to poor network.
 +
 
 
||
 
||
 
*
 
*
第212行： 第220行：
 
|Yue Gu
 
|Yue Gu
 
||
 
||
−
*
+
* modify the introduction
 +
* complete the interspeech poster, and open source the paper code
 +
* rest for two days, next I will focus on my new work
 
||
 
||
 
*
 
*
第221行： 第231行：
 
|Qi Qu
 
|Qi Qu
 
||
 
||
−
*  
+
* Inactive due to absence.
 
||
 
||
−
*
+
* KWS:
 +
** zh48 test dataset to be updated: ~30 speakers in 3 locations.
 +
** yue10 (Cantonese 10 keywords) train dataset to be updated: ~120 speakers verified, more to come.
 +
** Try to find suitable keyword-wise thresholds based on Recall ~ FA relation.
 
||
 
||
 
*   
 
*   
 
|-
 
|-

2024年8月19日 (一) 10:59的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • AI primary (middle-school) 1-6
Lantian Li
  • GPU status [1]
  • AI primary
    • High school handbook (30/40)
  • High school handbook (40/40)
Ying Shi
Zhenghai You
  • Continue the work of speaker augument and Ex-Former in TSE[2]
Jiaying Wang
  • reproduce conditional chain code
    • on both libri and wsj: training loss hard to reduce (around -3) and the corresponding test sisdr is positive
  • rewriting the code (preserving the original model)
Junming Yuan
  • Verified two parameters in Hubert pretraining config file that were confused with the original paper.[3]
    • Confirmed that in the second iteration of pretraining, features should be extracted from the 6-th layer of the transformer, not the 9-th layer.
      • in 175k step, result of 6-th layer: 71.55/9.39, result of 9-th layer: 37.31/16.72
    • Basically confirmed the setting of the parameter 'untie_final_proj' for the two iterations of pretraining.
Xiaolou Li
Zehua Liu
Pengqi Li
  • Investigating Extremely Short-Utterance in speaker recognition[4]
Tianhao Wang
  • reproducing CLIPSep on two datasets: MUSIC and VGGSound [5]
    • MUSIC: Text query: 10.06 SDR, Image query: 12.13 SDR
    • VGGSound: Text query: 2.78 SDR, Image query: 5.01 SDR
Zhenyu Zhou
  • conditional chain reproduce
  • model quantify[6]
Wan Lin
  • VoxBlink1
    • Data processing
    • Baseline(ResNet34) training and NS training [7]
Junhui Chen
  • VoxBlink1
    • Data processing
    • Baseline(ResNet34 ASP) training and NS training [8]
Yu Zhang
  • AED engineering problem assist
  • paper reading (will report this FRI)
Wenqiang Du
  • Complete Primary school handbook draft (45 + 8)
  • Modify the format, expression, and distribution of knowledge points in the draft（40%）
Yang Wei
Lily
Turi
  • Traveled back home and took rest
  • Now writing dataset paper.
    • Done with Intro, Literature review, Data collection sections. Experiment, Result and Conclusion sections remaining.
  • Wasn't able to do more experiment on dataset from Ethiopia due to poor network.
Yue Gu
  • modify the introduction
  • complete the interspeech poster, and open source the paper code
  • rest for two days, next I will focus on my new work
Qi Qu
  • Inactive due to absence.
  • KWS:
    • zh48 test dataset to be updated: ~30 speakers in 3 locations.
    • yue10 (Cantonese 10 keywords) train dataset to be updated: ~120 speakers verified, more to come.
    • Try to find suitable keyword-wise thresholds based on Recall ~ FA relation.
取自“http://index.cslt.org/mediawiki/index.php?title=2024-08-19&oldid=42147”