http://www.cslt.org/mediawiki/index.php?action=history&feed=atom&title=Sinovoice-2016-4-14 2026-04-04T04:34:21Z 本wiki的该页面的版本历史 MediaWiki 1.23.3 http://www.cslt.org/mediawiki/index.php?title=Sinovoice-2016-4-14&diff=19757&oldid=prev 2016-04-14T06:32:35Z

<p>以“==Data== *16K LingYun :* 2000h data ready :* 4300h real-env data to label * YueYu :* Total 250h(190h-YueYu + 60h-English) :* Add 60h YueYu :* CER: 75%-&gt;76% * WeiY...”为内容创建页面</p> <p><b>新页面</b></p><div>==Data==<br /> <br /> *16K LingYun<br /> :* 2000h data ready<br /> :* 4300h real-env data to label<br /> <br /> * YueYu<br /> :* Total 250h(190h-YueYu + 60h-English)<br /> :* Add 60h YueYu<br /> :* CER: 75%-&gt;76%<br /> <br /> * WeiYu<br /> :* 50h for training<br /> :* 120h labeled ready<br /> <br /> ==Model training==<br /> ===Big-Model Training===<br /> * 7*2048-10000h net weight-matrix factoring, to improve the decoding speed --SVD<br /> :* SVD looks OK, but fine-tuning still didn't work.<br /> Base WER:<br /> relu_2000_mpe_1000H: 17.72<br /> relu_1200_mpe_1000H: 18.60<br /> <br /> |layer / nodes retaind| 200 | 400 | 600 | 800 | 1000 | 1200 | 1400 | 1600 |<br /> | hidden 2 | | | 22.53 | 20.30 | 19.01 | | | |<br /> | hidden 7 | | 18.92 | 18.30 | 17.92 | | | | | <br /> | final | | | 18.32 | 18.00 | 17.83 | | | | <br /> * 7*1024 cross-entropy total train, then mpe, 0.2 improvment<br /> * 7*1024 svd factoring, speed the decoding<br /> <br /> * 8k<br /> <br /> ===Embedding===<br /> * 10000h-chain 5*400+800 DONE.<br /> :* Beam affect the performance of chain model significantly, need more investigation.<br /> * 5*576-2400 TDNN model<br /> <br /> ===SinSong Robot===<br /> * Test based on 10000h(7*2048-xent) model<br /> ------------------------------------------------<br /> condition | clean | replay(0.5m) | real-env<br /> ------------------------------------------------<br /> wer | 3 | 18(mpe-14) | too-bad<br /> ------------------------------------------------<br /> <br /> * Plan to record in restaurant on April 10.<br /> <br /> ===Character LM===<br /> *Except Sogou-2T, 9-gram has been done.<br /> *Worse than word-lm(9%-&gt;6%)<br /> *Add word boundary tag to Character-LM trainig<br /> *Merge Character-LM &amp; word-LM<br /> :* Union<br /> :* Compose, success.<br /> * 2-step decoding: first, character-based LM. Then, word-based LM.<br /> <br /> ===Project===<br /> * Pingan &amp; Yueyu Deletion error too more<br /> :* TDNN deletion error rate &gt; DNN deletion error rate<br /> :* TDNN Silence scale is too sensitive for different test cases.<br /> <br /> ==SID==<br /> ===Digit===<br /> * Same Channel test EER: 100% <br /> :* Speaker confirm<br /> :* phone channel<br /> <br /> * Cross Channel<br /> :* Mic-wav PLDA adaptation EER from 9% to 7% (20-30 persons)</div>

Zhangzy