“ASR:2015-03-23”版本间的差异
来自cslt Wiki
（→Text Processing） |
（→Speech Processing） |
||
| 第8行： | 第8行： | ||
==== RNN AM==== | ==== RNN AM==== | ||
* details at http://liuc.cslt.org/pages/rnnam.html | * details at http://liuc.cslt.org/pages/rnnam.html | ||
| − | * tuning parameters on monophone NN | + | * tuning parameters on monophone NN |
| + | |||
==== Mic-Array ==== | ==== Mic-Array ==== | ||
| − | + | * investigate alpha parameter in time domian and frquency domain | |
| − | * investigate alpha parameter in | + | |
====Dropout & Maxout & rectifier ==== | ====Dropout & Maxout & rectifier ==== | ||
| 第46行： | 第46行： | ||
:* Ivector dimention is smaller, performance is better | :* Ivector dimention is smaller, performance is better | ||
:* Augument to hidden layer is better than input layer | :* Augument to hidden layer is better than input layer | ||
| − | + | :* write paper for interspeech -- Xuewei | |
==Text Processing== | ==Text Processing== | ||
2015年3月26日 (四) 01:21的最后版本
Speech Processing
AM development
Environment
- grid-11 often shut down automatically, too slow computation speed.
- GPU has being repired.--Xuewei
RNN AM
- details at http://liuc.cslt.org/pages/rnnam.html
- tuning parameters on monophone NN
Mic-Array
- investigate alpha parameter in time domian and frquency domain
Dropout & Maxout & rectifier
- HOLD
- Need to solve the too small learning-rate problem
- 20h small scale sparse dnn with rectifier. --Mengyuan
- 20h small scale sparse dnn with Maxout/rectifier based on weight-magnitude-pruning. --Mengyuan Zhao
Convolutive network
- HOLD
- CNN + DNN feature fusion
- reproduce experiments -- Yiye
RNN-DAE(Deep based Auto-Encode-RNN)
- HOLD -Zhiyong
- http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=261
Speech rate training
- http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=268
- Technical report HOLD.-- Xiangyu Zeng, Shi Yin
- Paper for NCMMSC done
Neural network visulization
- http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=324
- Technical report done --Mian Wang.
Speaker ID
Ivector based ASR
- http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?step=view_request&cvssid=340
- Ivector dimention is smaller, performance is better
- Augument to hidden layer is better than input layer
- write paper for interspeech -- Xuewei
Text Processing
LM development
Domain specific LM
- LM2.X
- train a large lm using 25w-dict.(hanzhenglong/wxx)
- v2.0c filter the useless Chinese word and add 500 English word. get a little promotion effect.
tag LM
- Tag Lm(JT)
- get new script from mx and test 1 tag lm
- similar word extension in FST
- experiment done
- write the paper
RNN LM
- rnn
- the input and output is word embedding and add some token information like NER..
- map the word to character and train the lm.
- lstm+rnn
- check the lstm-rnnlm code about how to Initialize and update learning rate.(hold)
Word2Vector
W2V based doc classification
- data prepare.(hold)
Knowledge vector
- make a report on Monday
Translation
- v5.0 demo released
- cut the dict and use new segment-tool
Sparse NN in NLP
- prepare the ACL
- check the code to find the problem .
- increase the dimension
- use different test set.
QA
online learning
- data is ready.prepare the ACL paper
- prepare sougouQ data and test it using current online learning method
framework
- extract the module
- composite module
- fix the bug
leftover problem
- new inter will install SEMPRE
