Difference between revisions of "FreeNeb Status Report 2017-11-20"

From cslt Wiki
Jump to: navigation, search
(以“This Week: {| class="wikitable" !People !! Last Week !! This Week !! Next Week !! Task Tracing(<font color="red">DeadLine</font>) |- |Mengyuan Zhao || * Engineering...”为内容创建页面)
 
 
(3 intermediate revisions by 2 users not shown)
Line 17: Line 17:
 
|-
 
|-
 
|Zhiyong Zhang ||
 
|Zhiyong Zhang ||
* Train multi-speaker TTS based on Huilian and roobo data
+
* Fix code bug, and can reproduce the TTS experiment as Doc. Wang scripts based on huilian female1000
:* Base model done, but the synthesised wav is not good. It seems the acoustic model does not converge.
+
:* Large dataset test delayed due to the problems of grid
  
  
  
 
||
 
||
* Continue to find the problem of poor acoustic predicting of Multi-speaker TTS;
+
* Re-produce the Multi-speaker TTS;
 +
* To test two hybrid of the speakers.
 
* To train duration-model using 16k data.
 
* To train duration-model using 16k data.
  
Line 51: Line 52:
 
|-
 
|-
 
|Zhenlong Han||
 
|Zhenlong Han||
* Finish training Japanese acoustic model with transfer learning. Now the MPE is on training.
+
* Finish training Japanese acoustic model MPE, but It's get worse along epoch increase. Mengyuan told me that I should tuning all parameters in the Nnet. I will try again with present degs.
 
+
* Finish Uyghur embedded and cloud model xent training. But It is small worse than thuyg20 challenges, both Mengyuan and I dose not know more about the thuyg20 challenges model.
                                  ECO135 ECO_77 ETT SPC160 <br>
+
baseline_4X1200X9391_xent           21.06 13.23 26.73 16.6  <br>
+
embedded_6X512X800_xent                   26.85 18.38 33.22 24.02 <br>
+
embedded_6X400X800_transfer_learning_xent  23.8         15.71 29.66 19.64 <br>
+
 
+
* VAD is finished.
+
 
||
 
||
 
* English embedded model training.
 
* English embedded model training.
Line 69: Line 64:
 
|Shuai Zhang||
 
|Shuai Zhang||
 
* Add the intention of statistics into the graph and test it.
 
* Add the intention of statistics into the graph and test it.
 +
* Complete the first version VVParrot project
 
||
 
||
 
* According to the feedback, modify the project
 
* According to the feedback, modify the project

Latest revision as of 01:59, 20 November 2017

This Week:

People Last Week This Week Next Week Task Tracing(DeadLine)
Mengyuan Zhao
  • Engineering
  1. Draft API document of embedded-ASR engine.
  2. Draft API document of Deep Feature Extractor(deepfe).
  3. Draft a server version ASR DEMO, but still have bugs.
  • Engineering
  1. Try to finish server version of TTS DEMO.
Zhiyong Zhang
  • Fix code bug, and can reproduce the TTS experiment as Doc. Wang scripts based on huilian female1000
  • Large dataset test delayed due to the problems of grid


  • Re-produce the Multi-speaker TTS;
  • To test two hybrid of the speakers.
  • To train duration-model using 16k data.
Yang Wei
  • Write test specification for FreeNeb TTS engine.
  • Test FreeNeb TTS engine.
Dong Wang
  • ICASSP
  • OC2017


Zhenlong Han
  • Finish training Japanese acoustic model MPE, but It's get worse along epoch increase. Mengyuan told me that I should tuning all parameters in the Nnet. I will try again with present degs.
  • Finish Uyghur embedded and cloud model xent training. But It is small worse than thuyg20 challenges, both Mengyuan and I dose not know more about the thuyg20 challenges model.
  • English embedded model training.
  • Uyghur embedded model training.
Shuai Zhang
  • Add the intention of statistics into the graph and test it.
  • Complete the first version VVParrot project
  • According to the feedback, modify the project
  • Complete the VVParrot project
  • Add the documents about the two project


Yanchi Jin
  • Complete the megrez_tool (Fizzim) output format.
  • Start the compilation of the unit testing and overall test framework.
  • Complete the compilation of the framework.