http://index.cslt.org/mediawiki/index.php?action=history&feed=atom&title=Task_previous 2026-04-16T22:40:37Z 本wiki的该页面的版本历史 MediaWiki 1.23.3 http://index.cslt.org/mediawiki/index.php?title=Task_previous&diff=23059&oldid=prev 2016-10-16T12:17:42Z

<p>以“ =Task To Do= ==Speech Recognition== ===CTC expanded=== *Voice activity detection :*LSTM+CTC :*TDNN+CTC ::* BLANK as silence, others as speech *Keyword detection :...”为内容创建页面</p> <p><b>新页面</b></p><div><br /> =Task To Do=<br /> ==Speech Recognition==<br /> <br /> ===CTC expanded===<br /> *Voice activity detection<br /> :*LSTM+CTC<br /> :*TDNN+CTC<br /> ::* BLANK as silence, others as speech<br /> <br /> *Keyword detection<br /> :* Character/Word-level, external key-word fst<br /> :* G-fst need to be signal word?<br /> <br /> *Emotion recognition<br /> :* LSTM-CTC<br /> <br /> ===Network architecture test===<br /> *chain model:<br /> :*tdnn + simple-lstm<br /> ::* only keep forget gate<br /> <br /> *ctc+mpe<br /> :* similar to chain training<br /> <br /> ===Spiral Joint Training of SPEECH and SPEAKER===<br /> *ASR &amp; SID parallel training and benefit mutual<br /> <br /> ===Small data-set and Big model===<br /> *Investigate the efficiency of pre-training on small/big-model using dark-knowledge<br /> <br /> ===Low-resource language improvement===<br /> *SID<br /> :* How to improve low-resource speaker<br /> <br /> ===End-to-End speech recognition===<br /> * Discriminative-Learning code implementation<br /> :* Zhiyuan Tang<br /> <br /> ===Multi-task===<br /> * Fusion of speech-recognition and speech-rate<br /> :* Xiangyu Zeng<br /> * Self-informed neural network structure learning<br /> :* Mengyuan Zhao<br /> <br /> ===Integrate the class information to HCLG fst for speech recognition===<br /> *zhiyuan<br /> <br /> ===Distant speech recognition===<br /> *RNN-DAE: echo or reverberation<br /> :*Xuewei Zhang/Zhiyuan Tang/Mengyuan Zhao/Zhiyong Zhang<br /> *Reverberation<br /> :*Mutli-microphones<br /> :*(Lasso),Xuewei Zhang<br /> <br /> ===Voice conversation===<br /> *hold<br /> <br /> ===Sparse DNN===<br /> *Zhiyuan Tang<br /> <br /> ===Correlation based SENONE cluster===<br /> <br /> ===NN Multi-GPU parallel traing===<br /> *Multi-GPU using data parallelization<br /> :*Sheng Su<br /> * nnet3 mpe<br /> :* Xuewei Zhang<br /> <br /> ===Audio Embedding===<br /> *Ke Ning<br /> <br /> ===RNN training accelerating===<br /> <br /> ===Data selection===<br /> *Zhiyong Zhang<br /> *Sub-modular data selection<br /> *Objective-function loss training self-adaptation<br /> <br /> ===Decoder===<br /> *Confidence output for task-required<br /> <br /> ==Speaker Verification==<br /> ===binary code===<br /> *Lantian Li<br /> <br /> ===RNN-ivector===<br /> *Lantian Li<br /> <br /> ===DNN clustering===<br /> *Lantian Li<br /> <br /> =Task DONE=<br /> ==Multi-Mode features based VAD==<br /> * Shi Yin<br /> <br /> ==DNN based Language identification and Speaker identification==<br /> * Xuewei Zhang/Zhiyuan Tang<br /> <br /> ==Neural network visulization==<br /> * Mian Wang,DONE<br /> <br /> ==Dark knowledge==<br /> * Mengyuan Zhao, Xiangyu Zeng, Zhiyong Zhang, Chao Liu<br /> <br /> ==Normal RNN speech recognition==<br /> * Mengyuan Zhao<br /> <br /> ==Monmentum-like Hessien-Free acceleration==<br /> * Nestrov/Adagrad/AdaDelta/AdaM<br /> * Zhiyong Zhang/Xiangyu Zeng<br /> <br /> ==Activation value normalization through time --Batch Normalization==<br /> * Zhiyong Zhang<br /> <br /> ==Mix-training Balance decision tree==<br /> * Zhiyong Zhang<br /> <br /> ==20-h Chinese data-set release==<br /> * Xuewei Zhang<br /> <br /> ==Unbound activation function(Rectifier/Maxout/Pnorm) go-through searching method==<br /> * nne3 test --Xuewei Zhang<br /> <br /> =Technical Report To Write=<br /> 1, DNN-DAE based noise cancellation -- Xiangyu Zeng / Mengyuan Zhao / Zhiyong Zhang --DONE<br /> 2, Speech Rate DNN speech recognition --Shi Yin/Xiangyu Zeng --DONE<br /> 3, CNN+fbank feature combination --Mian Wang /Yiye Lin /Mengyuan Zhao /Shi Yin<br /> 4, Uyghur low-resource acoustic model enhancement -- Shi Yin / Mengyuan Zhao / Zhiyong Zhang --DONE<br /> 5, Uyghur 20h database release --Kaer /Shi Yin --DONE<br /> 6，Dark-Knowledge Transfer<br /> *: Xiangyu Zeng/ Mengyuan Zhao / Zhiyong Zhang<br /> <br /> =Paper to Write=<br /> <br /> =Patent done=<br /> * A method of new word enhancement for speech recognition --Yue Zhang<br /> <br /> =Project=<br /> * Xiaomi TV<br /> :*Mengyuan Zhao/Zhiyong Zhang<br /> :*TAG-lm &amp; Domain-specific general lm<br /> *Chinese-English mix-training</div>

Tangzy