“Sinovoice-2014-01-06”版本间的差异
来自cslt Wiki
第31行: | 第31行: | ||
* Start the training in 2 days | * Start the training in 2 days | ||
− | =Decoder= | + | =DNN Decoder= |
* Chao need to investigate the code change with Dr. Chen. | * Chao need to investigate the code change with Dr. Chen. | ||
* The work items involve (1) Kaldi tree loading (2) bigLM composition (3) DNN feature computing | * The work items involve (1) Kaldi tree loading (2) bigLM composition (3) DNN feature computing |
2014年1月6日 (一) 10:51的版本
目录
Project management
- Working items negotiation done
- 2014 contract setup
- The first amount (150k) delivered.
- Project team setup
DNN training
Environment setting
- Wiki setup
- Weekly meeting setup
- SGE environment settled in Sinovoice
Corpora
- New standard for data labeling is set
- The current standard involves regular sentences and noise, and the former may involve noise words
470 hour training
- 470h training started in Sinovoice server. Reached the 11th iteration of DNN. Training acc 48 and cv acc 47.15.
- 470h training with 8400 states also runs in the Sinovoice cluster.
- Parallel 470h training just started in CSLT cluster.
- Xiaoming will prepare the test set.
- More configurations on schedule.
6000 hour trainin
- Data preparation should be done in 1 day
- Start the training in 2 days
DNN Decoder
- Chao need to investigate the code change with Dr. Chen.
- The work items involve (1) Kaldi tree loading (2) bigLM composition (3) DNN feature computing