Sinovoice-2014-01-06

来自cslt Wiki
2014年1月6日 (一) 10:42Cslt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

Project negotiation

DNN training

Environment setting

Corpora

470hour training

1. System environment completed in sinovoice. Prepare a document for the building process.

600 hour training

2. 470h training started in Sinovoice server. Reach DNN 11 iterations. Training Acc 48., CV Acc 47.15. Higher than 170h results.

  470h training just started in CSLT cluster, In the monophone step.
  470 training with 8400 states, running into the first iteration.

3. Prepare 6k hour data.

4. Zhiyong & Xiaoming work on training recipe and configurations. 5. Xiaoming work on test set preparation.

Decoder

6. CLG decoder. LiuChao need to handle some code change for (1) Kaldi tree loading (2) bigLM compose (3) DNN feature computing