“2024-09-30”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
第49行： 第49行：
 
* Exploring the role of speaker encoder in TSE[https://z1et6d3xtb.feishu.cn/docx/GHF8doRjDo50ihxGUPpcsZgLncb]
 
* Exploring the role of speaker encoder in TSE[https://z1et6d3xtb.feishu.cn/docx/GHF8doRjDo50ihxGUPpcsZgLncb]
 
** Joint traing Spk Enc have better separation effect, but the EER is poor
 
** Joint traing Spk Enc have better separation effect, but the EER is poor
−
** Pretrain & Freeing Spk Enc EER well, but SI-SDR is poor
+
** Pretrain & Freezing Spk Enc EER well, but SI-SDR is poor
 
** Further explore the different impacts of using spk aug on different tasks
 
** Further explore the different impacts of using spk aug on different tasks
 
* The generality of SPK-AUG
 
* The generality of SPK-AUG

2024年9月30日 (一) 10:57的版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
Lantian Li
  • AI-Graph handbook v0.1
  • AI-Graph EN (12/50)
  • Huawei TiDing 3.0 - Model Quantization
  • BUPT/AI-Radiance trivial things
Ying Shi
  • Add 4 kinds of negative sampling strategies Optimized Text-enroll KWS code
    • (deletion, substitution, insertion, and shuffle) and verify them to ensure no bugs.
    • Find that new negative sampling will increase the difficulty of training which indicates that only depending on positional embedding is not enough.
  • Reproduce conditional chain overlap asr (Sequence to Multi-Sequence Learning via Conditional Chain Mapping for Mixture Signals)
    • According to Jiaying's work the code released by the published paper can not work
    • Write dominance-based conditional chain overlap asr by myself (in progress)
Zhenghai You
  • Exploring the role of speaker encoder in TSE[1]
    • Joint traing Spk Enc have better separation effect, but the EER is poor
    • Pretrain & Freezing Spk Enc EER well, but SI-SDR is poor
    • Further explore the different impacts of using spk aug on different tasks
  • The generality of SPK-AUG
    • Refactored DPRNN-TSE results are reliable and have been accelerated from 87 hours to 32 hours
Junming Yuan
Chen Chen
Xiaolou Li
  • Use MFA on LRS3 to cut it into small segments
  • Use discrete embedding of avhubert in vsp-llm training (Still training)
  • Some idea of align video feature and LLM (Dense Connector, CL methods)
  • Handover the data collection and get familiar with the process
  • Data Collection: 3138 h (need to re-check, DDL: 10.15)
Zehua Liu
Pengqi Li
Wan Lin
  • Voxblink1 model training and testing [2]
Tianhao Wang
  • AudioSep reproduction
    • problem: LAION CLAP needs 48kHz audio so the data needs to be up-resample
Xiaoxue Luo
  • AI-Graph High school handbook(v0.1)
Zhenyu Zhou
  • Model Quantization document submit
  • Review conditional chain code
Junhui Chen
Jiaying Wang
Yu Zhang
  • Fri Report
  • Change SocioDojo Agent from ChatGPT-3.5-Turbo to Llama-3.1-8B (still working)
Wenqiang Du
  • Check primary school handbook(43/45)
  • Release chinese and haining KWS model
Yang Wei
Lily
  • APSIPA workshop Tianjin and Prepare Friday's report
  • Prepare for online-course
  • AI radiance's daily work
Turi
  • Segmented audios in dataset into individual words.
  • Paper reading
Yue Gu
  • Almost complete the revisions of my journal paper
Qi Qu
  • KWS
    • Testing zh48 models on dataset of Mandarin Chinese w/ Guangdong accent: recall drops significantly.
  • AED
    • Evaluating third-party solution of baby crying detection.
  • Misc.
    • Preparing for live talk.
取自“http://index.cslt.org/mediawiki/index.php?title=2024-09-30&oldid=42286”