2014-08-29

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

• Error found in data setting. Re-run the test with gamma=20,30
• Re-run test with gamma=1,0.1
• Noisy training journal paper almost done.

Drop out & Rectification & convolutive network

• Change learning to 0.001, the training process can be started:
  1. change the drop probability from 0.5 to 0.2. Frame accuracy is improved. WER seems problematic.
  2. Experiment learning rate 1 and 8, NA

• 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. Found some specious problems with frequency-dependent Lasso.
2. Proposed full frequency Lasso & full frequency-temporal Lasso.
3. good performance was obtained with F-dependent Lasso

• Near data: 10.79 -> 10.35 (lamdba=0.05)
• Far data : 40.53 -> 35.65 (lambda=0.15)

VAD

• Some discrepancy between CSLT results & Puqiang results

[1]

• check if the label is really problematic
• check if short-time spike noise is the major problem (can be solved by spike filtering)
• check if low-energy babble noise caused mismatch (can be solved by global energy detection)
• test noise data trained model

Speech rate training

• Some interesting results with the simple speech rate change algorithm was obtained on the WSJ db

[2]

• Seems ROS model is superior to the normal one with faster speech
• Need to check distribution of ROS on WSJ
• Suggest to extract speech data of different ROS, construct a new test set
• Suggest to use Tencent training data
• Suggest to remove silence when compute ROS

Scoring

• hold

Confidence

• Implement a tool for data labeling
• Finished extraction of two features: DNN posterior + lattice posterior

LM development

Domain specific LM

h2. G determinization problem re-open.

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 performance; (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.
• Prepare WSJ database, re-test RNN.

Speaker ID

• Second model done

Emotion detection

• initial performance obtained

[3]

Translation

• Failed due to out of memory
• Re-start the training due to some errors in grid