|
|
| 第251行： |
第251行： |
| | |Qi Qu | | |Qi Qu |
| | || | | || |
| − | * | + | * Knock detection: output every knock's offset so that a shorter audio can be built to speed up human verification. |
| | + | * Text-enroll KWS: model i/o optimized; 2.5x faster than the first version. |
| | + | * KWS: Chongqing dialect train dataset (15 keywords, ~24.5k utterances). |
| | + | * Exp. using new FunASR model (SeACoParaformer) for cloud verification, which handles hotwords better. |
| | + | * Exp. using B0-based KWS model for local verification after detection from Chipintelli's chip. |
| | || | | || |
| | * | | * |
| People |
This Week |
Next Week |
Task Tracking (DeadLine)
|
| Dong Wang
|
- 2nd round check for middle-school AI handbook
- AI training for teachres of Tsinghua Middle School
|
|
|
| Lantian Li
|
- AI-Graph EN Chapter 3 Done
- 2025 Daily Sign v1.0 Done
- CSTR Report
|
|
|
| Ying Shi
|
- Test Google's product about sound separation
- Correct the test results of the previous condition overlap asr model here
|
|
|
| Zhenghai You
|
|
|
|
| Junming Yuan
|
- reproduce cocktail-Hubert
- feat-mask MT-Hubert
- result in [1]
|
|
|
| Chen Chen
|
|
|
|
| Xiaolou Li
|
- Finally finish the VTS report
- Data preparation
- CVS3 process 1/4
- take over webVideo from SUN CHANG and preprocess it through auto-avsr pipline
- Code preparation
- Finish the Conformer/CTC pretraining code
- Still debuging AVHuBERT pretraining code
- Paper reading...
|
|
|
| Zehua Liu
|
- VTS Documents Revise with Xiaolou
- Iterative inference training
- LLM Different context length[2]
|
|
|
| Pengqi Li
|
- Summarize recently work and report.
|
- Mapping to IPA from diff language.
- Write Paper.
|
|
| Wan Lin
|
|
|
|
| Tianhao Wang
|
- organizing the exp plan and modify the code for In-context-Audio-Retrieval (in training)
|
|
|
| Xiaoxue Luo
|
- prepare the code for CED+AudioSep
- participate in an AI competition with Wenqiang and Zhangyu
|
|
|
| Zhenyu Zhou
|
- Huawei project
- read papers
- code review(Design new ordering method)
|
|
|
| Junhui Chen
|
|
|
|
| Jiaying Wang
|
|
|
|
| Yu Zhang
|
- ICCIP 2024
- Paper reading about LLM Market Simulation
|
|
|
| Wenqiang Du
|
|
|
|
| Yang Wei
|
|
|
|
| Lily
|
|
|
|
| Turi
|
|
|
|
| Yue Gu
|
- synthesis some audios for target speakers [3]
- paper writing
|
|
|
| Qi Qu
|
- Knock detection: output every knock's offset so that a shorter audio can be built to speed up human verification.
- Text-enroll KWS: model i/o optimized; 2.5x faster than the first version.
- KWS: Chongqing dialect train dataset (15 keywords, ~24.5k utterances).
- Exp. using new FunASR model (SeACoParaformer) for cloud verification, which handles hotwords better.
- Exp. using B0-based KWS model for local verification after detection from Chipintelli's chip.
|
|
|
