2014-06-03

Resoruce Building

• Release management has been started

Leftover questions

• Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. Overfitting?
• Multi GPU training: Error encountered
• Multilanguage training
• Investigating LOUDS FST.
• CLG embedded decoder plus online compiler.
• DNN-GMM co-training

AM development

Sparse DNN

• GA-based block sparsity (+++++)

Noise training

• All experiments completed.
• Paper writing will be started this week

GFbank

• Test on Tencent database is done. Better performance observed than Fbank
• Equal-loudness pre-filter added, slightly better performance was obtained
• Running into Sinovoice 8k 1400 + 100 mixture training. 9 xEnt iteration completed.

Multilingual ASR

• Multilingual LM decoding
• Fixing the non-tag bug ???

English model

(state-gauss = 10000 100000, various LM, beam 13)

1. Shujutang 100h chi-eng 16k:

  LM/AM  |  xEnt   |  mpe_1  |  mpe_2  |  mpe_3  |  mpe_4  |
--------- --------- --------- --------- --------- ---------
   wsj   |  23.86  |  20.95  |  20.90  |  20.84  |  20.81  |
   cmu   |  22.22  |    -    |    -    |    -    |  18.83  |
   giga  |  21.77  |    -    |    -    |    -    |  18.61  |
  armid  |  20.45  |    -    |    -    |    -    |    -    |


2. Shujutang 100H chi-eng 8k:

  LM/AM  |  xEnt   |  mpe_1  |  mpe_2  |  mpe_3  |  mpe_4  |
--------- --------- --------- --------- --------- ---------
   wsj   |  26.27  |  23.63  |  23.14  |  22.93  |  23.00  |
   cmu   |  24.11  |    -    |    -    |    -    |  20.36  |
   giga  |  23.11  |    -    |    -    |    -    |  20.11  |
  armid  |    -    |    -    |    -    |    -    |    -    |


3. voxforge pure eng 16k:

  LM/AM  |  xEnt   |  mpe_1  |  mpe_2  |  mpe_3  |  mpe_4  |
--------- --------- --------- --------- --------- ---------
   wsj   |  21.38  |  24.89  |  24.50  |  23.31  |  23.13  |
   cmu   |  24.00  |    -    |    -    |    -    |  21.33  |
   giga  |  18.75  |    -    |    -    |    -    |  22.45  |
  armid  |    -    |    -    |    -    |    -    |    -    |

4. fisher pure eng 8k:
Not finish yet.
  LM/AM  |  xEnt   |  mpe_1  |  mpe_2  |  mpe_3  |  mpe_4  |
--------- --------- --------- --------- --------- ---------
   wsj   |  40.65  |  36.16  |  35.94  |  35.88  |  35.80  |
   cmu   |  35.07  |    -    |    -    |    -    |  31.16  |
   giga  |  41.18  |    -    |    -    |    -    |  36.23  |
  armid  |    -    |    -    |    -    |    -    |    -    |

Denoising & Farfield ASR

• Investigating DAE model
• Kaldi-based MSE obj training toolkit preparation

VAD

• DNN-based VAD (7.49) showers much better performance than energy based VAD (45.74)
• Need to test small scale network (+)

• 600-800 network
• 100 X 4 + 2

Scoring

• Bug for the stream mode fixed

Embedded decoder

• word list graph test passed
• wlist2LG toolkit checked in
• Prepare to deliver Android compiler options (.mk)
• Interface design should be completed in one day
• Prepare HCLG for 20k LM, decoding on progress.

LM development

Domain specific LM

• English lexicon done, build HCLG
• Re-build LM with the new lexicon
• Tested on Dianxin dev set

NN LM

• Character-based NNLM (6700 chars, 7gram), 500M data training done.

• Inconsistent pattern in WER were found on Tenent test sets
• probably need to use another test set to do investigation.

• Investigate MS RNN LM training

QA

FST-based matching

• Word-based FST 1-2 seconds with 1600 patterns. Huilan's implementation <1 second.
• THRAX toolkit for grammar to FST

• Investigate determinization of G embedding

• Refer to Kaldi new code