http://cslt.org/mediawiki/index.php?action=history&feed=atom&title=2014-08-222014-08-22 - 版本历史2026-04-15T10:02:36Z本wiki的该页面的版本历史MediaWiki 1.23.3http://cslt.org/mediawiki/index.php?title=2014-08-22&diff=10721&oldid=prevCslt：以内容“==Resoruce Building== == Leftover questions== * Investigating LOUDS FST. * CLG embedded decoder plus online compiler. * DNN-GMM co-training * NN LM == AM developmen...”创建新页面2014-08-22T02:13:24Z<p>以内容“==Resoruce Building== == Leftover questions== * Investigating LOUDS FST. * CLG embedded decoder plus online compiler. * DNN-GMM co-training * NN LM == AM developmen...”创建新页面</p>
<p><b>新页面</b></p><div>==Resoruce Building==<br />
<br />
== Leftover questions==<br />
<br />
* Investigating LOUDS FST. <br />
* CLG embedded decoder plus online compiler.<br />
* DNN-GMM co-training<br />
* NN LM<br />
<br />
== AM development ==<br />
<br />
=== Sparse DNN ===<br />
* WJS sparse DNN does not obtain further improvement<br />
<br />
===Noise training===<br />
<br />
:* Noisy training journal paper almost done.<br />
<br />
==Drop out & Rectification & convolutive network==<br />
<br />
* Change learning to 0.001, the training process can be started: <br />
*# check the drop probability <br />
*# check learning rate<br />
*# continuous training<br />
<br />
* Rectification<br />
# Rectification itself failed with large weights.<br />
# Including L1 penalty enables the training but got very bad performance.<br />
# Try to set the maximum value with rectifier<br />
<br />
* Convolutive network<br />
# Test more configurations <br />
<br />
<br />
===Denoising & Farfield ASR===<br />
<br />
* Lasso-based dereverberation obtained reasonable results<br />
:# spectrum based lasso outperforms fbank based lasso.<br />
:# temporal-frequency based lasso outperforms just temporal based lasso.<br />
:# using 200 frame to estimate utterance-based lasso coefficients is possible, with marginal performance degradation.<br />
:# using lasso can solve the problem of dynamic reverberation.<br />
:# Need to investigate static reverberation. <br />
:# The 1/3 paper has been checked in to cvs. <br />
<br />
===VAD===<br />
<br />
* Found some problems in Puqiang's speech data. Some files are labelled incorrectly.<br />
<br />
<br />
===Speech rate training===<br />
<br />
* Append an additional dimension to the feature vector, indicating the rate of speech <br />
* The ROS is computed as words per second<br />
<br />
===Scoring===<br />
<br />
* Refine the acoustic model with AMIDA database. problem solved by involving both wsj and AMIDA.<br />
<br />
===Confidence===<br />
<br />
* Knowledge prepared<br />
* First experiment with combining lattice-based confidence and DNN confidence. <br />
* Further step will add ROS.<br />
<br />
<br />
===Embedded decoder===<br />
<br />
* Chatting LM released (80k)<br />
* Train two smaller network: 500x4+600, 400x4+500: on going<br />
* Build a new graph with MPE3 am and chatting LM.<br />
<br />
==LM development==<br />
<br />
===Domain specific LM===<br />
<br />
h2. G determinization problem solved<br />
<br />
h2. NUM tag LM:<br />
<br />
27h JS test: 20.16 vs 20.19<br />
2h JS test: 17.48 vs 17.49<br />
<br />
h2. Analyze the property of the tag LM: (1) random NUM should obtain better perfomance; (2) other words are not seriously impacted.<br />
<br />
<br />
==Word2Vector==<br />
<br />
===W2V based doc classification===<br />
<br />
* Initial results variable Bayesian GMM obtained. Performance is not as good as the conventional GMM.<br />
* Interest group setup, reading scheduled every Thusday<br />
* Non-linear inter-language transform: English-Spanish-Czch: wv model training done, transform model on investigation<br />
<br />
<br />
==RNN LM==<br />
<br />
* New toolkit from Thomas obtained<br />
* Need more investigation on the toolkit<br />
<br />
<br />
==Speaker ID==<br />
<br />
* Second model done<br />
<br />
<br />
==Translation==<br />
* Failed due to out of memory <br />
* Re-train the model with limitation on iteration number. Goes to 8th iteration</div>Cslt