|
|
| 第58行： |
第58行： |
| | * double check mixed Hubert code： | | * double check mixed Hubert code： |
| | ** fix some bugs (time-mask.etc) | | ** fix some bugs (time-mask.etc) |
| − | ** time-mask vs. feat mask: (27.98%, 23.17%) vs.(23.19%, 25.99%) | + | ** time-mask vs. feat mask: (Top-1 acc, EER): (27.98%, 23.17%) vs.(23.19%, 25.99%) |
| | ** softmax+CE --> sigmoid+BCE still have problem. | | ** softmax+CE --> sigmoid+BCE still have problem. |
| | || | | || |
| People |
This Week |
Next Week |
Task Tracking (DeadLine)
|
| Dong Wang
|
- AIGraph high education version
- Prepare AIgraph Large Model version
- NMI paper publication staff
|
|
|
| Lantian Li
|
- AI-Graph EN (1/4)
- Huawei Project Proposal v1.0
- First Lesson on 24-fall AI Undergraduates
|
|
|
| Ying Shi
|
- Huawei project proposal
- Optimize the Text-enroll KWS code
- improve readability.
- remove redundant code.
|
|
|
| Zhenghai You
|
- Exploring the generality of spk aug on different data and structures[1]
|
|
|
| Junming Yuan
|
- double check mixed Hubert code：
- fix some bugs (time-mask.etc)
- time-mask vs. feat mask: (Top-1 acc, EER): (27.98%, 23.17%) vs.(23.19%, 25.99%)
- softmax+CE --> sigmoid+BCE still have problem.
|
|
|
| Xiaolou Li
|
- Writing VTS documents
- Paper Reading & Preparing for Report
- Exp on LRS3
- LLM: LLaMA2 -> LLaMA3.1 (30h ↓0.4%)
- Grouping LLaMA2: (443h ↑0.5%, 30h ↓2.5%)
- Rethinking the method to inject information (ablation study first)
|
|
|
| Zehua Liu
|
|
|
|
| Pengqi Li
|
|
|
|
| Wan Lin
|
- VC2 pre-train; VB1+VC2 mix-turning
- Data filter in VB1: 1.25% EER in vox1-o
- VB1 pre-train; VC2 fine-tuning
- VB1 pre-train: 2.61% EER in vox1-o
- VC2 fine-tuning: maybe couldn't reach better performance
|
|
|
| Tianhao Wang
|
|
|
|
| Zhenyu Zhou
|
|
|
|
| Junhui Chen
|
|
|
|
| Jiaying Wang
|
|
|
|
| Yu Zhang
|
- Dataset collection from THS
- Retraining R^2 SAC paper, with same env still failed (TCN ACC: 0.708, RECALL: 0.183), will check with Han this week
- Paper reading and some plan (report this Fri)
|
|
|
| Wenqiang Du
|
|
|
|
| Yang Wei
|
- Train text enroll KWS model with aishell and kespeech data.
- Prepare live broadcast
|
|
|
| Lily
|
- Prepare for holiday course（October 2nd、3rd） and online-course
- AI radiance's daily work
|
|
|
| Turi
|
- Trained conformer on Sagalee data excluding utterances containing digits
- Achieved 21.28% WER, 2.65 WER reduction
- Preparing KWS data from Sagalee dataset using MFA
|
|
|
| Yue Gu
|
- paper writing
- open the code
- prepare for the presentation
|
|
|
| Qi Qu
|
- KWS:
- Finding ideal thresholds for b0-models in predefined scenes: Chinese Mandarin, Cantonese, Uyghur and Kazakh.
- Finding ideal thresholds for b6-models with fixed b0-model thresholds.
- AED:
- Fixing parameters of Fbank feature extraction for CED and retraining classifiers.
|
|
|
