“FreeNeb Status Report 2017-12-25”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（3位用户的4个中间修订版本未显示）
第22行： 第22行：
 
|-
 
|-
 
|Zhiyong Zhang ||
 
|Zhiyong Zhang ||
−
* Finished 100h 16k TTS training
+
* Retrain 100h 16k TTS training
−
:* Poor synthesised wavs, the reason may be
+
:* The synthesised waves still not seems well, hold
−
::* Noise corrupted wav may not fit TTS task, the corrupted spectrum would be learned too;
+
* Contact with THS about exist technical problems, still not receive feedback
−
::* I assumed one utterence for one specified person, not so accurate, may another reason?
+
:* I have finished part of the technical solution
 +
 
 +
* Contact with Aiweier, and end the cooperation.
  
 
||
 
||
−
* To train duration-model using 16k data.
+
 
  
 
||
 
||
 
* Aiweier
 
* Aiweier
−
# Contact with Aiweier.
+
# Contact with Aiweier. --Done
 
* Myhexin
 
* Myhexin
 
# Draft a document and contact with Myhexin.
 
# Draft a document and contact with Myhexin.
第39行： 第41行：
 
|-
 
|-
 
|Yang Wei ||
 
|Yang Wei ||
−
* Write test specification for FreeNeb TTS pyserver.
+
* Locate a failed case
−
* Test pyserver.
+
* Test the cases that client close incidentally.
 
||
 
||
−
* Add new cases to test FreeNeb TTS pyserver.
+
* Test the cases that network is not stable.
 
||
 
||
 
* Continue test TTS pyserver
 
* Continue test TTS pyserver
第61行： 第63行：
 
|-
 
|-
 
|Zhenlong Han||
 
|Zhenlong Han||
−
* write document.
+
* English embedded model finish, and cloud model have training 700 model
−
* train some model with 8k and 16k, and add ivector to it.
+
                              libri_clean wsj_eval92 wsj_eval93 voxforge libri_other
 +
recipe_result                       6.95           6.4                         16.29
 +
embedded_6X400X800_tl_xent       13.49           8.05         10.9         12.83         32.82
 +
embedded_6X400X800_tl_xent_kt       13.71           8.12         11.44         12.92         31.86
 +
cloud_6X1200X8000_tl_xent_kt_700mdl    10.78              6.42         8.63         11.59         27.29
 +
 
 
||
 
||
 
* English embedded model training.
 
* English embedded model training.
第71行： 第78行：
 
|-
 
|-
 
|Shuai Zhang||
 
|Shuai Zhang||
−
* Add the function of asking.
+
* Test the Megrez project overall
−
* Add the edit-distance algorithm.
+
 
* fixed the bugs in answering the question.
 
* fixed the bugs in answering the question.
 
||
 
||
−
* Test the Megrez project overall.
 
 
* Complete the VVParrot project.  
 
* Complete the VVParrot project.  
−
* Add the documents about the two project.  
+
* Modify the edit-distance algorithm.
 
||
 
||
−
 
+
# graph to json toolkit v0.1 release. Dec.27.
 
||
 
||
 
* the test by ourselves about the Megrez will be done 20th Dec.  
 
* the test by ourselves about the Megrez will be done 20th Dec.  

2017年12月25日 (一) 03:09的最后版本

This Week:

People Last Week This Week Meet Minutes Task Tracing(DeadLine)
Mengyuan Zhao
  • Engineering
  1. Optimize TTS-pyserver to speed up.
  2. Working on online-ASR DEMO.
  • Engineering
  1. Try to finish first version of online-ASR APP.
  2. Clean up code and draft documents.
  • Roobo
  1. Prepare a workshop(half a day) in Jan 2018 after Jan.20,2018.
  • Engineering
  1. Improve speed of TTS engine
  2. Draft documents for many engines: fnasr, fntts, deepfe....
  1. Online-TTS DEMO (Jan.20,2018)
  2. Online-ASR DEMO (Jan.20,2018)
Zhiyong Zhang
  • Retrain 100h 16k TTS training
  • The synthesised waves still not seems well, hold
  • Contact with THS about exist technical problems, still not receive feedback
  • I have finished part of the technical solution
  • Contact with Aiweier, and end the cooperation.


  • Aiweier
  1. Contact with Aiweier. --Done
  • Myhexin
  1. Draft a document and contact with Myhexin.
Yang Wei
  • Locate a failed case
  • Test the cases that client close incidentally.
  • Test the cases that network is not stable.
  • Continue test TTS pyserver
Dong Wang
  • Some discussion with Fenyinta
  • Some discussion with Japan Tongfang
  • More people will join!
  • APSIPA 2017


Zhenlong Han
  • English embedded model finish, and cloud model have training 700 model

libri_clean wsj_eval92 wsj_eval93 voxforge libri_other recipe_result 6.95 6.4 16.29 embedded_6X400X800_tl_xent 13.49 8.05 10.9 12.83 32.82 embedded_6X400X800_tl_xent_kt 13.71 8.12 11.44 12.92 31.86 cloud_6X1200X8000_tl_xent_kt_700mdl 10.78 6.42 8.63 11.59 27.29

  • English embedded model training.
  • Prepare all the ASR models:
Shuai Zhang
  • Test the Megrez project overall
  • fixed the bugs in answering the question.
  • Complete the VVParrot project.
  • Modify the edit-distance algorithm.
  1. graph to json toolkit v0.1 release. Dec.27.
  • the test by ourselves about the Megrez will be done 20th Dec.
  • the VVParrot will be finished the standard function 28th Dec which will include login, register, record the score and get the sentence randomly.
Yanchi Jin
  • Already update the database.
  • Complete the Megrez_tool and the chart of structure.
  • Continue update the database.
取自“http://cslt.org/mediawiki/index.php?title=FreeNeb_Status_Report_2017-12-25&oldid=29589”