Sinovoice-2014-12-10

来自cslt Wiki
2014年2月10日 (一) 06:41Cslt讨论 | 贡献的版本

跳转至: 导航搜索

DNN training

Environment setting

  • Another 3 3T disks are ready for RAID-0.
  • Another GPU machine was purchased. Add 4 3T disks to construct RAID-0.

Corpora

  • Scripts for confidence generation is ready for auto transcription.
  • 300h telephone speech data (Sinovoice recording) were done.
  • Adaptation data 900 sentences ready.

470 hour 8k training

  • 300h incremental training (IT) done
Model CE MPE1 MPE2 MPE3 MPE4
4k states 23.27/22.85 21.35/18.87 21.18/18.76 21.07/18.54 20.93/18.32
8k states 22.16/22.22 20.55/18.03 20.36/17.94 20.32/17.78 20.29/17.80
8k states + IT - 20.04/17.38 20.01/17.32 20.07/17.44 19.94/17.65

6000 hour 16k training

  • Ran CE DNN to iteration 5 (8400 states, 80000 pdf)
  • Testing results go down to 13.77% WER (Sinovoice results: 11.78).
Model WER RT
small LM, it 4, -5/-9 15.80 -
large LM, it 4, -5/-9 15.30 -
large LM, it 4, -6/-9 15.36 -
large LM, it 4, -7/-9 15.25 -
large LM, it 5, -5/-9 14.17 -
large LM, it 5, -5/-10 13.77 -

Adaptation

  • Code ready for direct adaptation, insertion adaptation and KL-regularized adaptatoin
  • 50 sentences for adaptation, 834 sentences for testing
  • WER from 14.56 to 11.13
  • Hidden layer adaptation is better than input and output adaptation
  • Before Linear adaptation is better than after-linear adaptation
  • Results are here

DNN Decoder

  • Comparison between CLG and HCLG decoder
  • CLG decoder uses less memory in decoding
  • HCLG is faster and more accurate than HCLG, and more amiable to beam control here
  • std::exp/std::log result in very slow computation in train203. Solved the problem by replacing to standard exp() and log().