2014-06-06

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.
• 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

• Paper writing will be started this week

GFbank

• Running into Sinovoice 8k 1400 + 100 mixture training.
• GFbank 14 xEnt iteration completed:

                                  Huawei disanpi     BJ mobile   8k English data

FBank non-stream (17 iteration) 22.01% 26.63% - GFbank stream (14 iteration) 22.47%; 27.52% -

Multilingual ASR

                                  Huawei disanpi     BJ mobile   8k English data

FBank non-stream - - -

• Multilingual LM decoding
• TAG-based decoding still problematic. Decoding goes into subgraph, however the decoding results are incorrect.
• Investigate with free-loop grammar.
• Non-tag test should be conducted on both Baidu & micro blob data
• Should test the 8k shujutang data on the mixture model.

Denoising & Farfield ASR

• Add artificial reverberant with various energy decay & time delay. Draw a plot decay vs WER, delay vs WER.
• Use more training data to do adaptation.
• Record the wave with a single speaker & near-field microphone and do test again.

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 test
• 100 X 4 + 2 network training

Scoring

• Collect more data with human scoring to train discriminative models

Embedded decoder

1200 X 4 + 10k AM:

       150k       20k     10k      5k 

WER 42.23 43.45 44.54 46.07 RT 1h31 48m 44m 43m

LM development

Domain specific LM

• Retrieve both Baidu & microblog
• Need to check into gitLab(+).

Word2Vector

• Design network spider
• Design semantic related word tree

• First version based on pattern match done
• Filter with query log
• Further refinement with Baidu Baike hierarchy

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