2014-08-22

Leftover questions

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

AM development

Sparse DNN

• WJS sparse DNN does not obtain further improvement

Noise training

• Noisy training journal paper almost done.

Drop out & Rectification & convolutive network

• Change learning to 0.001, the training process can be started:
  1. check the drop probability
  2. check learning rate
  3. continuous training

• Rectification

1. Rectification itself failed with large weights.
2. Including L1 penalty enables the training but got very bad performance.
3. Try to set the maximum value with rectifier

• Convolutive network

1. Test more configurations

Denoising & Farfield ASR

• Lasso-based dereverberation obtained reasonable results

1. spectrum based lasso outperforms fbank based lasso.
2. temporal-frequency based lasso outperforms just temporal based lasso.
3. using 200 frame to estimate utterance-based lasso coefficients is possible, with marginal performance degradation.
4. using lasso can solve the problem of dynamic reverberation.
5. Need to investigate static reverberation.
6. The 1/3 paper has been checked in to cvs.

VAD

• Found some problems in Puqiang's speech data. Some files are labelled incorrectly.

Speech rate training

• Append an additional dimension to the feature vector, indicating the rate of speech
• The ROS is computed as words per second

Scoring

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

Confidence

• Knowledge prepared
• First experiment with combining lattice-based confidence and DNN confidence.
• Further step will add ROS.

Embedded decoder

• Chatting LM released (80k)
• Train two smaller network: 500x4+600, 400x4+500: on going
• Build a new graph with MPE3 am and chatting LM.

LM development

Domain specific LM

h2. G determinization problem solved

h2. NUM tag LM:

27h JS test: 20.16 vs 20.19 2h JS test: 17.48 vs 17.49

h2. Analyze the property of the tag LM: (1) random NUM should obtain better perfomance; (2) other words are not seriously impacted.

Word2Vector

W2V based doc classification

• Initial results variable Bayesian GMM obtained. Performance is not as good as the conventional GMM.
• Interest group setup, reading scheduled every Thusday
• Non-linear inter-language transform: English-Spanish-Czch: wv model training done, transform model on investigation

RNN LM

• New toolkit from Thomas obtained
• Need more investigation on the toolkit

Speaker ID

• Second model done

Translation

• Failed due to out of memory
• Re-train the model with limitation on iteration number. Goes to 8th iteration