“2020-03-02”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
 
(8位用户的13个中间修订版本未显示)
第6行: 第6行:
 
|Dong Wang
 
|Dong Wang
 
||  
 
||  
*  
+
* Mostly completed the investigation on normalized scoring and complted the draft article
 +
* Investigated the DNF with simulation data
 
||
 
||
*  
+
* Keep on investigating the DNF model with simulation data
 +
* Coin some new approach for SID scoring, motivated by the NL interpertation of PLDA.
 
||
 
||
 
*   
 
*   
第17行: 第19行:
 
|Yunqi Cai
 
|Yunqi Cai
 
||  
 
||  
*  
+
* Finished patent
 +
* Investigate the total covariance constrain of DNF and showed some results
 +
* Preparing the weekly paper report
 
||
 
||
*  
+
* More investigation on DNF
 
||
 
||
 
*   
 
*   
第28行: 第32行:
 
|Zhiyuan Tang
 
|Zhiyuan Tang
 
||  
 
||  
*  
+
* Repeat glow and asr on WSJ dataset.
 
||
 
||
*  
+
* Different noise/domain on test.
 
||
 
||
 
*   
 
*   
第52行: 第56行:
 
|Ying Shi
 
|Ying Shi
 
||  
 
||  
*
+
* verify mel spectrum on DAE and double flow
 +
* train DAE and Double Flow(different layer and different flow type ) with spectrum feats(6 kinds of noise with random snr).
 +
* compute SDR and PESQ of DAE  and Double Flow
 
||
 
||
*
+
* verify the result
 +
* compute fwSNR
 +
* build more powerfull DAE baseline and flow model
 
||
 
||
 
*   
 
*   
第63行: 第71行:
 
|Wenqiang Du
 
|Wenqiang Du
 
||  
 
||  
*
+
*Using  the add noise、 spec-argument 、attention to improve model performance in different device channel
||
+
*
+
 
||
 
||
 
*   
 
*   
第85行: 第91行:
 
|Yue Fan
 
|Yue Fan
 
||  
 
||  
*
+
*Build the baseline models
 +
*Derive model formula
 
||
 
||
*
+
*Optimize model structure
 
||
 
||
 
*   
 
*   
第96行: 第103行:
 
|Jiawen Kang
 
|Jiawen Kang
 
||  
 
||  
*  
+
* Collect cnceleb speaker-birth map.
 +
* Finish length, age and 11*11 geners experiments.
 
||
 
||
*
+
* Make some distribution figures for different speakers and geners.
 +
* Design Cross-channel experiments.
 
||
 
||
 
*   
 
*   
第107行: 第116行:
 
|Ruiqi Liu
 
|Ruiqi Liu
 
||  
 
||  
*  
+
* Experiments on CN-Celeb different scenes and different length .
 +
* Continue training the model.
 
||
 
||
*  
+
* More experiments for different speakers and scenes.
 
||
 
||
 
*   
 
*   
第136行: 第146行:
 
|-
 
|-
  
 +
|-
 +
|Haolin Chen
 +
||
 +
* Double WaveGlow on noisy TIMIT
 +
* Test performance of Double Glow
 +
* Optimize model structure
 +
||
 +
* Comprehensive evaluation on two models
 +
||
 +
 +
|-
  
 
|}
 
|}

2020年3月2日 (一) 00:06的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • Mostly completed the investigation on normalized scoring and complted the draft article
  • Investigated the DNF with simulation data
  • Keep on investigating the DNF model with simulation data
  • Coin some new approach for SID scoring, motivated by the NL interpertation of PLDA.
Yunqi Cai
  • Finished patent
  • Investigate the total covariance constrain of DNF and showed some results
  • Preparing the weekly paper report
  • More investigation on DNF
Zhiyuan Tang
  • Repeat glow and asr on WSJ dataset.
  • Different noise/domain on test.
Lantian Li
  • Submit DNF patent.
  • Complete DNF repository.
  • Start frame-level DNF-SRE training.
  • Go on DNF-SRE.
Ying Shi
  • verify mel spectrum on DAE and double flow
  • train DAE and Double Flow(different layer and different flow type ) with spectrum feats(6 kinds of noise with random snr).
  • compute SDR and PESQ of DAE and Double Flow
  • verify the result
  • compute fwSNR
  • build more powerfull DAE baseline and flow model
Wenqiang Du
  • Using the add noise、 spec-argument 、attention to improve model performance in different device channel
Haoran Sun
Yue Fan
  • Build the baseline models
  • Derive model formula
  • Optimize model structure
Jiawen Kang
  • Collect cnceleb speaker-birth map.
  • Finish length, age and 11*11 geners experiments.
  • Make some distribution figures for different speakers and geners.
  • Design Cross-channel experiments.
Ruiqi Liu
  • Experiments on CN-Celeb different scenes and different length .
  • Continue training the model.
  • More experiments for different speakers and scenes.
Sitong Cheng
Zhixin Liu
Haolin Chen
  • Double WaveGlow on noisy TIMIT
  • Test performance of Double Glow
  • Optimize model structure
  • Comprehensive evaluation on two models