2014-09-05

Leftover questions

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

AM development

Sparse DNN

• Investigating layer-based DNN training

Noise training

• Noisy training journal paper almost done.

Drop out & Rectification & convolutive network

• Drop out

• No performance improvement found yet.
• [1]

• Rectification

• Dropout NA problem was caused by large magnitude of weights

• Convolutive network

1. Test more configurations

Denoising & Farfield ASR

• Lasso-based dereverberation obtained reasonable results

• optimize the training parameters by the development set
• Found similar alpha for both near and far recordings. Need more investigation.

VAD

• Noise model training stuck by local minimal.
• Some discrepancy between CSLT results & Puqiang results

[2]

• 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)

Speech rate training

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

[3]

• 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, correcting some errors.
• Finished extraction of two features: DNN posterior + lattice posterior

LM development

Domain specific LM

h2. G determinization problem solved.

h2. NUM tag LM:

• Seems OK with the tag LM.

[4]

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

• Investigate more iterations to obtain a better more
• Checking the discrepancy between the matlab nnet tool & sklearn.

RNN LM

• Prepare WSJ database
• Trained model 10000 x 4 + 320 + 10000
• Start to test on n-best rescore

Speaker ID

• Second model done

Emotion detection

• delivered to Sinovoice

Translation

• v2.0 demo ready

QA

• Labeled 1000 utterances as the evaluation
• 35% 11-class accuracy
• EA not done yet