http://cslt.org/mediawiki/index.php?action=history&feed=atom&title=2014-05-23 2026-04-15T01:04:36Z 本wiki的该页面的版本历史 MediaWiki 1.23.3 http://cslt.org/mediawiki/index.php?title=2014-05-23&diff=9993&oldid=prev 2014-05-23T01:55:40Z

<p>以内容“==Resoruce Building== * Release management has been started * Blaster 0.1 &amp; vivian 0.0 system release == Leftover questions== * Asymmetric window: Great improvement on...”创建新页面</p> <p><b>新页面</b></p><div>==Resoruce Building==<br /> * Release management has been started<br /> * Blaster 0.1 &amp; vivian 0.0 system release<br /> <br /> == Leftover questions==<br /> * Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. Overfitting? <br /> * Multi GPU training: Error encountered<br /> * Multilanguage training<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 /> * GA-based block sparsity (++++)<br /> <br /> ===Noise training===<br /> :* All experiments completed. Combing experiments. <br /> <br /> ===GFbank===<br /> <br /> * WSJ clean condition done. Obtained the same performance as the time domain implementation<br /> * Should experiment with the Tencent training set. <br /> <br /> ===Multilingual ASR===<br /> <br /> * Multilingual LM decoding<br /> * Fixing the non-tag bug<br /> <br /> ===English model===<br /> <br /> &lt;pre&gt;<br /> RESULTS:<br /> (state-gauss = 10000 100000)<br /> <br /> 1. Shujutang 100h chi-eng 16k:<br /> <br /> LM/AM | xEnt | mpe_1 | mpe_2 | mpe_3 | mpe_4 |<br /> --------- --------- --------- --------- --------- ---------<br /> wsj | 23.86 | 20.95 | 20.90 | 20.84 | 20.81 |<br /> <br /> <br /> 2. Shujutang 100H chi-eng 8k:<br /> <br /> LM/AM | xEnt | mpe_1 | mpe_2 | mpe_3 | mpe_4 |<br /> --------- --------- --------- --------- --------- ---------<br /> wsj | 26.27 | 23.63 | 23.14 | 22.93 | 23.00 |<br /> <br /> <br /> 3. voxforge pure eng 16k:<br /> <br /> LM/AM | xEnt | mpe_1 | mpe_2 | mpe_3 | mpe_4 |<br /> --------- --------- --------- --------- --------- ---------<br /> wsj | 21.38 | 24.89 | 24.50 | 23.31 | 23.13 |<br /> <br /> <br /> 4. fisher pure eng 8k:<br /> Not finish yet.<br /> LM/AM | xEnt | mpe_1 | mpe_2 | mpe_3 | mpe_4 |<br /> --------- --------- --------- --------- --------- ---------<br /> wsj | 40.65 | - | - | - | - |<br /> <br /> &lt;/pre&gt;<br /> <br /> * Need to experiment with Gigabyptes LM<br /> * Need to check the AM settings and LM used in the Kaldi egs/fisher<br /> <br /> ===Denoising &amp; Farfield ASR===<br /> <br /> * Baseline: close-talk model decode far-field speech: 92.65<br /> * Will investigate DAE model<br /> <br /> ===Kaiser Window ===<br /> * Test on different numbers of Fbanks: no significant difference between 23/30/40 (both 8k/16k) [#223]<br /> * Test on Kaiser &amp; Povey window: no significant difference obtained for both 8k/16k [#224, #225]<br /> <br /> ===VAD===<br /> <br /> * DNN-based VAD (7.49) showers much better performance than energy based VAD (45.74)<br /> * [http://cslt.riit.tsinghua.edu.cn/mediawiki/images/1/1d/Dnn_vad_VS_energy_vad.pdf click here]<br /> * Need to test small scale network <br /> :* 600-800 network<br /> :* 100 X 4 + 2<br /> <br /> ===Scoring===<br /> <br /> * Fixing bug for the stream mode<br /> <br /> ==LM development==<br /> <br /> ===Domain specific LM===<br /> * English lexicon done, build HCLG<br /> * Re-build LM with the new lexicon<br /> * Tested on Dianxin dev set<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 /> <br /> * Investigate MS RNN LM training<br /> <br /> ==QA==<br /> <br /> ===FST-based matching===<br /> :* Word-based FST 1-2 seconds with 1600 patterns. Huilan's implementation &lt;1 second.<br /> :* THRAX toolkit for grammar to FST<br /> <br /> * Investigate determinization of G embedding <br /> :* Refer to Kaldi new code</div>

Cslt