“Hulan-2013-10-18”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
ASR Kernel development
TTS
第7行: 第7行:
 
==TTS==
 
==TTS==
  
* CD lab files done. Refining the script.  
+
* full-lab training is ready. Trained the first full-lab system with 16k/pseduo 48k data.
* Training toolkit is cleaned up. Now no alignment is required. Parallel training is done.
+
* re-recording 48k data using F00 (500 sentences) and retrain the model. The quality of the signal sounds better, while the quality of pitch is a bit strange. Need more investigation on parameter settings.
* Tried syllable based system instead of phones.
+
 
* Collected an online-novel reading.
+
  
 
Next week:
 
Next week:
  
* Refine the script
+
* Check the signal parameters and solve the problem of pitch.
* Clean up the online reading.
+
* Prepare the large data training with both all-F 863 data.
 +
* Prepare the large data training with online novel.
  
 
=Dialog system=
 
=Dialog system=

2013年10月18日 (五) 01:58的版本

ASR

ASR Kernel development

[ASR group weekly report]

TTS

  • full-lab training is ready. Trained the first full-lab system with 16k/pseduo 48k data.
  • re-recording 48k data using F00 (500 sentences) and retrain the model. The quality of the signal sounds better, while the quality of pitch is a bit strange. Need more investigation on parameter settings.


Next week:

  • Check the signal parameters and solve the problem of pitch.
  • Prepare the large data training with both all-F 863 data.
  • Prepare the large data training with online novel.

Dialog system

  • The search system migrated to the custom domain, with significant performance reduction
  Customs:
n	TF	TFIDF	
1	0.496	0.485
2	0.619	0.615
3	0.676	0.673
4	0.713	0.715
5	0.740	0.738

Agriculture:
n	TF	TFIDF
1	0.75	0.8
2	0.85	0.883
3	0.867	0.917
4	0.867	0.95
5	0.95	0.967
  • Two problems:
  1. short of semantic cluster.
  2. limited training data for idf.
  • Next week
  1. Analyse the QA database, to extract useful domain dependent data
  2. Analyse the data to expand the key words & phrases
  3. Analyse the data to attain better IDF.