|
|
| 第152行： |
第152行： |
| | || | | || |
| | * construct loudness training data: librimix with different loudness source | | * construct loudness training data: librimix with different loudness source |
| − | * loudness order exp: | + | * loudness order exp[https://z1et6d3xtb.feishu.cn/docx/TDQ9dMYiHoyZnLx6xJDcLg1ZnF3]: |
| | ** chain based structure: 2mix 11.92(150 epoch) | | ** chain based structure: 2mix 11.92(150 epoch) |
| | ** convtasnet structure: still training, 60epoch achieve 12.29 | | ** convtasnet structure: still training, 60epoch achieve 12.29 |
| People |
This Week |
Next Week |
Task Tracking (DeadLine)
|
| Dong Wang
|
|
|
|
| Lantian Li
|
- Review AI Book — College Edition (12/53).
- Project matters: HUAWEI SS/AutoBGM; FYT A/V GenreDetect.
- Prepare materials for my professional title evaluation.
|
|
|
| Ying Shi
|
|
|
|
| Zhenghai You
|
|
|
|
| Junming Yuan
|
- Grade 8 AI practice book(done!)
- SUPERB benchmark(Speech processing Universal Performance Benchmark)
- Source Separation downstream task(SI-SDRi at 50K training steps):
- MT-HuBERT: 10.77, HuBERT-BASE: 9.84
|
|
|
| Xiaolou Li
|
|
|
|
| Zehua Liu
|
|
|
|
| Pengqi Li
|
|
|
|
| Wan Lin
|
|
|
|
| Tianhao Wang
|
|
|
|
| Xiaoxue Luo
|
- read some papers on speech separation of unknown number of speakers
- Apply EDA(encoder-decoder based attractor calculation) method to speech separation
- Environment Configuration(done)
- Familiar with the code and make some adjustments to it(in progress)
|
|
|
| Junhui Chen
|
- LLM:
- 2 different reflection paper reading
- Baseline self-reflection metric collection — code completed, data collection in progress.
- Integrating new reflection methods (e.g. beamsearch-based reflection) into the current pipeline in collaboration with @Zhang Yu — work in progress.
|
|
|
| Jiaying Wang
|
- construct loudness training data: librimix with different loudness source
- loudness order exp[1]:
- chain based structure: 2mix 11.92(150 epoch)
- convtasnet structure: still training, 60epoch achieve 12.29
- ctc order exp:
- only use ctc for order, chain based structure: 10.65
- a speculation: Compared with semantic information (provided by CTC), acoustic-oriented information may be more suitable as a basis for separation.
|
- try to use speaker info as order
|
|
| Yu Zhang
|
- LLM:
- Baseline self-reflection metric collection — code completed, data collection in progress.
- Integrating new reflection methods (e.g. beamsearch-based reflection) into the current pipeline in collaboration with @chenjunhui — work in progress.
- AED:
- Added humming test and analysis for Huawei.
|
|
|
| Wenqiang Du
|
|
|
|
| Yang Wei
|
|
|
|
| Yue Gu
|
- polish the structure of my thesis and continue find jobs
|
|
|
| Qi Qu
|
|
|
|
