|
|
| 第153行： |
第153行： |
| | |Junhui Chen | | |Junhui Chen |
| | || | | || |
| − | * | + | * NS paper writing |
| | || | | || |
| | * | | * |
| People |
This Week |
Next Week |
Task Tracking (DeadLine)
|
| Dong Wang
|
- Paper review for ISCSLP
- AIGraph slides check (37/50)
|
|
|
- Finish AIGraph slides check
| Lantian Li
|
- GPU status [1]
- Review several chain-based separation papers
- AI graph
- Slides checking (23/50)
- High school handbook (11/40)
- Submit GTV
|
- High school handbook (20/40)
|
|
| Ying Shi
|
|
|
|
| Zhenghai You
|
- Reproduce DPRNN-IRA[2]
- TSE Project: Targetless model with tSDR and tMSE Loss
|
|
|
| Junming Yuan
|
|
|
|
| Chen Chen
|
|
|
|
| Xiaolou Li
|
- LRS-30h PALR2 (2 epoch result, still training)
- VSR: 30.56%
- Refinement: 30.40%
- Paper reading
|
|
|
| Zehua Liu
|
- LRS3-30h ICL (CER:30.3 > 29.1）
- LRS3-30h image adaptive mask code
- Papper reading[3] and Maybe Q-former can be useful
|
|
|
| Pengqi Li
|
|
|
|
| Wan Lin
|
|
|
|
| Tianhao Wang
|
- New task investigation: Sound Filter
- Neural Scoring: CN-Celeb exps refresh the results
- project: data washing and generator toolkit
|
|
|
| Zhenyu Zhou
|
- TSE stream test[5]
- Multi-speaker SS Paper reading
|
|
|
| Junhui Chen
|
|
|
|
| Jiaying Wang
|
- change to chain-based methods
- reproducing sequence to multi-sequence code
- read related papers
|
|
|
| Yu Zhang
|
- AED:
- retraining model without dilation convolution
- assist Huawei in model quantization and engineering
- Finance:
- retraining model with data totally align with JunWang
|
|
|
| Wenqiang Du
|
- Complete the company's model upgrade with Weiyang
- cn + minnanyu + Uyghur
- cn + Uyghur + Kazakh
- primary school handbook (12/46)
|
|
|
| Yang Wei
|
- AIBabel KWS
- Train a new version of Chinese, Uyghur, Kazakh model. [6]
|
|
|
| Lily
|
|
|
|
| Turi
|
- Completed data collection
- Trained conformer on full data (100h) for 200 epochs. WER 23.9% on 2hrs of test data.
|
|
|
| Yue Gu
|
- complete the introduction、related work and 1/3 results of the journal paper (8/9).
|
|
|
| Qi Qu
|
- KWS:
- Documents: data preparation, test method.
- Datasets prepared: 1. Mandarin Chinese mild-accent (20 keywords, 14 persons); 2. FAs from production (150k, from Jul. 1 to Jul. 15).
- AED:
- AudioSet as negative samples.
|
|
|
