“2014-08-01”版本间的差异

2014年8月1日 (五) 01:53的最后版本

Resoruce Building

Leftover questions

• Investigating LOUDS FST.
• CLG embedded decoder plus online compiler.
• DNN-GMM co-training
• NN LM

AM development

Sparse DNN

• WJS sparse DNN shows a slightly better than non-sparse cases when the network is in a large scale
• Pre-training does work for DNN training (for both 4/5/6 layers)

Noise training

• Journal paper writing on going

Multilingual ASR

• Native English speaker + Chinglish speaker obtained better performance.

Drop out & convolutional network

• Change learning to 0.001, the training process can be started.
• Frame Accuracy goes to : (with/without drop probability normalization)

Denoising & Farfield ASR

• By tuning parameters of late-response lag & response time, obtained performance improvement with Lasso.

Simulation results: Baseline:
--------------------------------------------------------------------------
                model/test              |  far_evl92   |   near_evl92
--------------------------------------------------------------------------
                   clean_ce             |     59.38    |    19.25
                 mpe_clean_ce           |     40.46    |    12.94
--------------------------------------------------------------------------

Lasso with optimal parameters(lambda=0.05, delta=5, N=10)
--------------------------------------------------------------------------
                model/test              |  far_evl92   |   near_evl92
--------------------------------------------------------------------------
                clean_ce                |     54.63    |    15.75
--------------------------------------------------------------------------
              mpe_clean_ce              |     36.58    |    11.64
--------------------------------------------------------------------------

Real data results:
--------------------------------------------------------------------------
                   model/test              |  far_evl92   |   near_evl92
 --------------------------------------------------------------------------
                   clean_ce                |     94.86    |    63.48
--------------------------------------------------------------------------
                 mpe_clean_ce              |     92.29    |    58.37
--------------------------------------------------------------------------
dereverberated recording :

--------------------------------------------------------------------------
                   model/test              |  far_evl92   |   near_evl92
 --------------------------------------------------------------------------
                   clean_ce                |     94.91    |    61.03
--------------------------------------------------------------------------
                 mpe_clean_ce              |     91.28    |    54.16
--------------------------------------------------------------------------

• Adaptation under running

VAD

• Waiting for testing results

Scoring

• Refine the acoustic model with AMIDA database. problem solved by involving both wsj and AMIDA.

Confidence

• Be familiar with Kaldi
• Need to extract lattice and DNN features

Embedded decoder

• Chatting LM released (80k)
• Train two smaller network: 500x4+600, 400x4+500: on going
• Need to upload the new client code onto git (+)
• Build a new graph with MPE3 am and chatting LM.

LM development

Domain specific LM

h2. Domain specific LM construction

h3. TAG LM

• TAG obtained better performance

h3. Chatting LM

• First version released (80k lexicon)
• Prepare 2nd released (120k lexicon)
• Test on Xiaotang long

Word2Vector

W2V based doc classification

• Initial results variable Bayesian GMM obtained. Performance is not as good as the conventional GMM.

Speaker ID

• Full-data SRE trial goes into the final stage
• results will be ready soon

Translation

• collecting more data (Xinhua parallel text, bible, name entity) for the second version
• check possible parameters to control phrase pair lexicon