Mengyuan Zhao 15-08-03

来自cslt Wiki
2015年8月3日 (一) 14:18Zhaomy讨论 | 贡献的版本

(差异) ←上一版本 | 最后版本 (差异) | 下一版本→ (差异)
跳转至: 导航搜索
  • Xiaomi project
  1. test pair method for new words, didn't acquire good result.
  2. tune weight of some words in G.fst. result looks good.
  • LSTM training
  1. Track 8k AM (1400h+300h) xEnt training, and 16k AM (1700h+776h) xEnt training.
  2. Start MPE training on PingAnBaoXian 700h dataset. iter1 still not finish, but the performance is already better than best DNN MPE model.