140519 - Mengyuan Zhao
From cslt Wiki
- Sinovoice gfbank model:
- 6000H-stream-mpe, still running, 1 iter done, but result is bad. trying to find error.
- Multi-lingual training based on 100H Tencent data & 26h-mobile-eng data:
- do some phone-shared experiments
- do some state-gauss = 4200 40000 experiments
- EN training:
- ShuJuTang chi-eng 16k_wav training, done.
- ShuJuTang chi-eng 8k_wav training, wav convert and data preparation done, dnn running.
- voxforge 16k Pure En training, done.
- fisher 8k Pure En, data preparation done.