“Zhiyong Zhang 2016-05-09”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
第3行: 第3行:
 
:# WER17.99 for 1e-7(size:13M) lm-threshold
 
:# WER17.99 for 1e-7(size:13M) lm-threshold
 
:# WER17.51 for 1e-8(size:54M) lm-threshold
 
:# WER17.51 for 1e-8(size:54M) lm-threshold
* All worse than character-lm without boundary which is WER13.24 (size:71M) lm-threshold:1e-7
+
:# All worse than character-lm without boundary which is WER13.24 (size:71M) lm-threshold:1e-7
  
 
# Investigate the Reranker Framework(ReFr) toolkit for discriminative language model training
 
# Investigate the Reranker Framework(ReFr) toolkit for discriminative language model training

2016年5月9日 (一) 00:50的版本

  • Weekly Report
  1. Test character-lm with word-boundary
  1. WER17.99 for 1e-7(size:13M) lm-threshold
  2. WER17.51 for 1e-8(size:54M) lm-threshold
  3. All worse than character-lm without boundary which is WER13.24 (size:71M) lm-threshold:1e-7
  1. Investigate the Reranker Framework(ReFr) toolkit for discriminative language model training
  • To do next week
  1. Try to use ReFr to train a discriminative language model