2014-05-23

Resoruce Building

• Release management has been started
• Blaster 0.1 & vivian 0.0 system release

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. Combing experiments.

GFbank

• WSJ clean condition done. Obtained the same performance as the time domain implementation
• Should experiment with the Tencent training set.

Multilingual ASR

• Multilingual LM decoding
• Fixing the non-tag bug

English model

RESULTS:
(state-gauss = 10000 100000)

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  |


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  |


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  |


4. fisher pure eng 8k:
Not finish yet.
  LM/AM  |  xEnt   |  mpe_1  |  mpe_2  |  mpe_3  |  mpe_4  |
--------- --------- --------- --------- --------- ---------
   wsj   |  40.65  |    -    |    -    |    -    |    -    |

• Need to experiment with Gigabyptes LM
• Need to check the AM settings and LM used in the Kaldi egs/fisher

Denoising & Farfield ASR

• Baseline: close-talk model decode far-field speech: 92.65
• Will investigate DAE model

Kaiser Window

• Test on different numbers of Fbanks: no significant difference between 23/30/40 (both 8k/16k) [#223]
• Test on Kaiser & Povey window: no significant difference obtained for both 8k/16k [#224, #225]

VAD

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

• 600-800 network
• 100 X 4 + 2

Scoring

• Fixing bug for the stream mode

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