Sinovoice-2014-03-18

来自cslt Wiki
2014年3月19日 (三) 03:24Cslt讨论 | 贡献的版本

跳转至: 导航搜索

Environment setting

  • Raid215 is a bit slow. Move some den-lattice and alignment to Raid212.

Corpora

  • PICC data are done (200h).
  • Huibei telecom data are done (108h).
  • Now totally 1229h (470 + 346 + 105BJ mobile + 200 PICC + 108h) telephone speech is ready.
  • 16k 6000h data: 978h online data from DataTang + 656h online mobile data + 4300h recording data.
  • LM corpus preparation done.

Acoustic modeling

Telephone model training

1000h Training

  • Xent completed. Compiling lattices.
  • Need to test the xEnt performance

PICC dedicated training

  • Need to collect financial text data and retrain the LM
  • Need to comb word list and training text


6000 hour 16k training

Training progress

  • 6000h/CSLT phone set alignment/denlattice completed
  • 6000h/jt phone set alignment/denlattice completed
  • MPE is kicked off


Train Analysis

  • The Qihang model used a subset of the 6k data
  • 2500+950H+tang500h*+20131220, approximately 1700+2400 hours
  • GMM training using this subset achieved 22.47%. Xiaoming's result is 16.1%.
  • Seems the database is still not very consistent
  • Xiaoming will try to reproduce the Qihang training using this subset
  • Tested the 1700h model and 6000h model on the T test sets
  model/testcase  |   ditu |  due1| entity1 | rec1 | shiji | zaixian1 | zaixian2 | kuaisu
  ------------------------------------------------------------------------------------------------
    1700h_mpe       |  12.18 | 12.93 | 5.29   |   3.69     |  21.73  | 25.38   | 19.45   | 12.50
  ------------------------------------------------------------------------------------------------
    6000h_xEnt      |  11.13 | 10.12 | 4.64   |   2.80     |  17.67  | 27.45   | 23.23   | 10.98 
  • 6000h model is general better than the 1700h for careful reading or domain specific recording
  • 6000h with MPE/jt phone set is still on training, but better performance is expected
  • This indicates that we should prepare domain-specific AM (not only 8k/16k). The online test prefers online training data
  • Suggest test the 6000 model on jidong data

Hubei telecom

  • Hubei telecom data (127 h), retrieve 60k sentence by conf thred=0.9, amounting to 50%
xEnt org:  -             wer_15  29.05
MPE iter1:wer_14 29.23;wer_15 29.38
MPE iter2:wer_14 29.05;wer_15 29.11
MPE iter3:wer_14 29.32;wer_15 29.28
MPE iter4:wer_14 29.29;wer_15 29.28
  • retrieve 30k sentences by conf thred=0.95, amounting to 25%, plus the original 770h data
xEnt org:     -             wer_15  29.05
MPE iter1:    -             wer_15: 29.36
  • Incremental training with Hubei telecome based on the model (470+300+BJmobile)
  • MPE4 modeltraining done: org: 27.30, Hubei model: 25.42

Language modeling

  • Training data ready
  • Focus on PICC test set


DNN Decoder

Online decoder adaptation

  • Finished alignment/den-lattice
  • 1st round MPE training on going, 2 days/iteration