来自cslt Wiki
(差异) ←上一版本 | 最后版本 (差异) | 下一版本→ (差异)
- Do some experiment on "chain" model, and got some conclusion:
- "chain" model is not sensitive with the number of output layer (pdf_num);
- post layer need more context than prior layer
- Start 11000h "chain" model training for sinovoice.