http://cslt.org/mediawiki/index.php?action=history&feed=atom&title=2014-01-03 2026-04-15T05:13:01Z 本wiki的该页面的版本历史 MediaWiki 1.23.3 http://cslt.org/mediawiki/index.php?title=2014-01-03&diff=8934&oldid=prev 2014-01-03T02:20:27Z

<p>以内容“== AM development == === Sparse DNN === * Optimal Brain Damage(OBD). # Online OBD held. # OBD + L1 norm start to investigation. * Efficient computing # Conducti...”创建新页面</p> <p><b>新页面</b></p><div>== AM development ==<br /> <br /> === Sparse DNN ===<br /> <br /> * Optimal Brain Damage(OBD). <br /> <br /> # Online OBD held. <br /> # OBD + L1 norm start to investigation. <br /> <br /> * Efficient computing<br /> <br /> # Conducting rearrangement the matrix structure and compose zero blocks by some smart approaches, leading to better computing speed. <br /> <br /> <br /> === Efficient DNN training ===<br /> <br /> # L1-L2 grid checking: L1/L2(&lt; 1e-6) seems good for record1900 but worse for other test sets.<br /> <br /> [http://192.168.0.50:3000/series/?action=view&amp;series=199,199.0,199.1,199.2,199.3,199.4,199.5,199.6,199.7,199.8,199.9 link here]<br /> <br /> # Asymmetric window: Great improvement on training set(WER 34% to 24%), however the improvement is lost on test. Overfitting? <br /> # Frame-skipping. Skipping 1 frame speeds up decoding in a consistent way while retaining the accuracy largely. Skipping more frames lead to unacceptable performance degradation.<br /> # Interpolation does not provide performance gain.<br /> <br /> [http://192.168.0.50:3000/series/?action=view&amp;series=199,199.5,199.6,199.7,199.8,199.9,198,198.0,198.1,198.2,198.3,198.4,198.5,198.6 link here]<br /> <br /> === Optimal phoneset===<br /> <br /> * Analyze Tencent English phone set. Found some errors in CH/EN phone sharing.<br /> * Develop a new sharing scheme, start training the new system. <br /> * Start training for all-separated phones<br /> * Start training mixed system with Chinglish data.<br /> <br /> <br /> ===Engine optimization===<br /> <br /> * Investigating LOUDS FST. On progress. <br /> <br /> <br /> ==LM development==<br /> <br /> ===NN LM===<br /> <br /> * Collecting a bigger lexicon: 40k words related to music, 56k words from an official dictionary.<br /> * Working on NN LM based on word2vector.<br /> <br /> ==Embedded development==<br /> <br /> * Liuchao's cellphone, Qualcomm Snapdragon Krait MSM8960 @ 1.5GHz, using 1 core<br /> small nnet 100/600/600/600/600/1264 with MFCC input<br /> <br /> * 4500 words:<br /> <br /> :* construct LG: 0.41s<br /> :* compose HCLG with det: 13.70s, 5.318 MB<br /> :* compose HCLG without det: 6.61s, 5.488 MB<br /> <br /> * 950 words:<br /> <br /> :* construct LG: 0.15s<br /> :* compose HCLG with det: 2.63s, 0.947 MB, decode RT 0.649<br /> :* compose HCLG without det: 1.74s, 0.998 MB, decode RT 0.548<br /> <br /> * For word list or simple grammars, determinization leads to small RT increase, but can improve HCLG compiling dramatically. This is particularly the case for embedded devices. <br /> * The accuracy does not change with/without determinization.<br /> <br /> <br /> ==Speech QA==<br /> <br /> <br /> * Use N-best to expand match in QA. Better performance were obtained.<br /> :* 1-best matches 96/121 <br /> :* 10-best matches 102/121<br /> <br /> * Use N-best to recover errors in entity check. Working on.<br /> * Use Pinyin to recover errors in entity check. Future work.</div>

Cslt