“140603 Mengyuan Zhao”版本间的差异
来自cslt Wiki
（以内容“Weekly report: *gamma-tone-filter bank (GTFBANK) test on Tencent data. results on cvss. *Sinovoice EN training: xEnt & MPE training ,done.test them on cmu-lm and giga-l...”创建新页面） |
|||
| 第2行： | 第2行： | ||
*gamma-tone-filter bank (GTFBANK) test on Tencent data. results on cvss. | *gamma-tone-filter bank (GTFBANK) test on Tencent data. results on cvss. | ||
*Sinovoice EN training: | *Sinovoice EN training: | ||
| − | xEnt & MPE training ,done.test them on cmu-lm and giga-lm. | + | #xEnt & MPE training ,done.test them on cmu-lm and giga-lm. |
*Sinovoice multi-language training based on 1400h_Dianhua + 100h_Shujutang (8K) | *Sinovoice multi-language training based on 1400h_Dianhua + 100h_Shujutang (8K) | ||
#Fbank, xEnt 16 iters done. | #Fbank, xEnt 16 iters done. | ||
#Start GTFBANK_stream training. xEnt 9 iters done. | #Start GTFBANK_stream training. xEnt 9 iters done. | ||
*Test frame-skipping decode method on Tencent fbank model. result shows that WER have a little increament (<1%). | *Test frame-skipping decode method on Tencent fbank model. result shows that WER have a little increament (<1%). | ||
2014年6月3日 (二) 04:47的最后版本
Weekly report:
- gamma-tone-filter bank (GTFBANK) test on Tencent data. results on cvss.
- Sinovoice EN training:
- xEnt & MPE training ,done.test them on cmu-lm and giga-lm.
- Sinovoice multi-language training based on 1400h_Dianhua + 100h_Shujutang (8K)
- Fbank, xEnt 16 iters done.
- Start GTFBANK_stream training. xEnt 9 iters done.
- Test frame-skipping decode method on Tencent fbank model. result shows that WER have a little increament (<1%).
