http://www.cslt.org/mediawiki/index.php?action=history&feed=atom&title=2014-07-252014-07-25 - 版本历史2026-04-03T21:25:45Z本wiki的该页面的版本历史MediaWiki 1.23.3http://www.cslt.org/mediawiki/index.php?title=2014-07-25&diff=10525&oldid=prevCslt：以内容“==Resoruce Building== == Leftover questions== * Investigating LOUDS FST. * CLG embedded decoder plus online compiler. * DNN-GMM co-training == AM development == ==...”创建新页面2014-07-25T02:13:03Z<p>以内容“==Resoruce Building== == Leftover questions== * Investigating LOUDS FST. * CLG embedded decoder plus online compiler. * DNN-GMM co-training == AM development == ==...”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 />
<br />
== AM development ==<br />
<br />
=== Sparse DNN ===<br />
* WJS sparse DNN shows a slightly better than non-sparse cases when the network is in a large scale<br />
* Pre-training does work for DNN training (for both 4/5/6 layers)<br />
<br />
===Noise training===<br />
:* Journal paper writing on going<br />
<br />
<br />
===Multilingual ASR===<br />
<br />
* With multlingual training, performance is largely retained with most of known test sets; <br />
* However for unknown accents, performance is not stable<br />
<br />
==Drop out & convolutional network==<br />
<br />
* Zhiyong will study drop out<br />
* Zhiyong & Mengyuan will study convolutional network<br />
<br />
===Denoising & Farfield ASR===<br />
<br />
* Use an reverberation tool to generate a new set of datasets<br />
<br />
* xEnt results(eval 92):<br />
before adaptation after adaptation<br />
clean: - -<br />
near: 19.25 12.94<br />
far: 59.38 40.46<br />
<br />
* Lasso-based reverberation cancellation got initial clean data<br />
<br />
===VAD===<br />
<br />
* Waiting for engineering work<br />
<br />
===Scoring===<br />
<br />
* Refine the acoustic model with AMIDA database. problem solved by involving both wsj and AMIDA.<br />
<br />
===Embedded decoder===<br />
<br />
* Chatting LM release<br />
* Train two smaller network: 500x4+600, 400x4+500: on going<br />
* Need to upload the new client code onto git<br />
* Build a new graph with MPE3 am and chatting LM.<br />
<br />
<br />
==LM development==<br />
<br />
===Domain specific LM===<br />
<br />
h2. Domain specific LM construction<br />
<br />
h3. TAG LM<br />
<br />
* TAG still problematic with all-to-number tag<br />
* check the randomness of the number tag.<br />
<br />
h3. Chatting LM<br />
* Building chatting lexicon<br />
* First version released (80k lexicon)<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 />
<br />
<br />
==Semantic word tree==<br />
<br />
:* Version v2.0 released (filter with query log)<br />
:* Please deliver to /nfs/disk/perm/data/corpora/semanticTree (Xingchao)<br />
:* Version v3.0 under going. Further refinement with Baidu Baike hierarchy<br />
<br />
<br />
==NN LM==<br />
<br />
* Character-based NNLM (6700 chars, 7gram), 500M data training done.<br />
:* Inconsistent pattern in WER were found on Tenent test sets<br />
:* probably need to use another test set to do investigation. <br />
* Investigate MS RNN LM training<br />
<br />
==Speaker ID==<br />
<br />
* reading materials <br />
* prepare to run sre08<br />
<br />
<br />
==Translation==<br />
* collecting more data (Xinhua parallel text, bible, name entity) for the second version<br />
* work into text alignment<br />
* Will release v2.0 today</div>Cslt