http://cslt.org/mediawiki/index.php?action=history&feed=atom&title=2014-04-11 2026-04-14T11:24:10Z 本wiki的该页面的版本历史 MediaWiki 1.23.3 http://cslt.org/mediawiki/index.php?title=2014-04-11&diff=9658&oldid=prev 2014-04-11T02:13:59Z

<p>以内容“==Resoruce Building== * Current text resource has been re-arranged and listed == Leftover questions== * Asymmetric window: Great improvement on training set(WER 34% to...”创建新页面</p> <p><b>新页面</b></p><div>==Resoruce Building==<br /> * Current text resource has been re-arranged and listed<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 /> :* Found a paper in 2000 with similar ideas. <br /> :* Try to get a student working on high performance computing to do the optimization<br /> <br /> <br /> ===Noise training===<br /> :* More experiments with no-noise<br /> :* More experiments with additional noise types<br /> <br /> <br /> ===AMR compression re-training===<br /> <br /> * 1700h MPE adaptation done<br /> * 1700h stream mode adaptation runs into MPE1<br /> <br /> ===GFbank===<br /> * Significant improvement found with GFBank <br /> * Significant improvement found with FBank + GFBank<br /> <br /> ===Denoising &amp; Farfield ASR===<br /> * Recording done<br /> * Prepare to construct the baseline<br /> <br /> <br /> ===VAD===<br /> <br /> * Code ready, need to figure out speech/no-speech smooth<br /> <br /> ===Farfield recognition===<br /> <br /> ===Scoring===<br /> <br /> * g-score based on MLP is done<br /> * t-score based on linear regression improves the performance<br /> <br /> <br /> ==Word to Vector==<br /> <br /> * LDA baseline (sogou 1700*9 training set) done<br /> * Wordvector classification is much better than the LDA system<br /> &lt;pre&gt;<br /> word vector: <br /> general: dict - 15w; train_data - ren_ming_ri_bao(5g); windows-5<br /> 1. size - 50 time=30m 12thread<br /> 2. size - 10 time=10m 12thread<br /> <br /> data: class_num=9 document_num=9*2000<br /> train_num =9*1600<br /> test_num =9*200<br /> dev_num =9*200<br /> <br /> train_set:<br /> C000008 C000010 C000013 C000014 C000016 C000020 C000022 C000023 C000024 total <br /> 财经 IT 健康 体育 旅游 教育 招聘 文化 军事 <br /> lda_inf 0.845 0.2756 0.698 0.9502 0.63499 0.32 0.8080 0.3505 0.864 0.6385<br /> lda_inf_10 0.8149 0.0887 0.628 0.9641 0.5739 0.105 0.707363 0.2334 0.8628 0.553167<br /> w2v_filter_filer 0.7463 0.713 0.657 0.9106 0.68659 0.54 0.74638 0.692 0.84518 0.72638<br /> w2v_filter_filer_10 0.7608 0.4323 0.57394 0.865 0.549 0.335 0.577 0.6129 0.78099 0.609769<br /> <br /> test_set:<br /> C000008 C000010 C000013 C000014 C000016 C000020 C000022 C000023 C000024 total <br /> 财经 IT 健康 体育 旅游 教育 招聘 文化 军事 <br /> w2v_filter_filter 0.6865 0.7263 0.6716 0.84577 0.7462 0.46268 0.6567 0.7114 0.8905 0.71088<br /> w2v_filter_filter_10 0.791 0.4079 0.56218 0.74129 0.62189 0.22885 0.562 0.6766 0.84079 0.603648<br /> lda_inf 0.8706 0.26368 0.6965 0.8009 0.582 0.2537 0.72139 0.3184 0.82587 0.59259<br /> lda_inf_10 0.776 0.1044 0.6467 0.9054 0.62189 0.1144 0.56218 0.24378 0.796 0.530127<br /> <br /> note:w2v_filter--remove the stop word in traing word vector<br /> note:w2v_filter_filter -- remove the stop word in traing word vector and remove the documnet stop words<br /> &lt;/pre&gt;<br /> <br /> ==LM development==<br /> <br /> ===NN LM===<br /> <br /> * Character-based NNLM (6700 chars, 7gram), 500M data training done.<br /> :* Non-boundary char LM is better than boundary char LM<br /> <br /> * Investigate MS RNN LM training<br /> <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 /> :* Char-FST Implementation is done. <br /> <br /> <br /> ===Speech QA===<br /> * Investigate determinization of G embedding</div>

Cslt