“ASR:2015-06-08”版本间的差异
来自cslt Wiki
（以“==Speech Processing == === AM development === ==== Environment ==== * grid-15 and grid-11 run slowly or do not work ==== RNN AM==== * details at http://liuc.cslt.o...”为内容创建页面） |
（→Speech Processing） |
||
| （1位用户的3个中间修订版本未显示） | |||
| 第3行： | 第3行： | ||
==== Environment ==== | ==== Environment ==== | ||
| − | * | + | * |
| − | + | ||
==== RNN AM==== | ==== RNN AM==== | ||
| − | * | + | *morpheme RNN-zhiyuan |
| − | + | ||
| − | + | ||
| − | + | ||
| − | + | ||
| − | + | ||
| − | ==== Mic-Array ==== | + | * ==== Mic-Array ==== |
* hold | * hold | ||
* Change the prediction from fbank to spectrum features | * Change the prediction from fbank to spectrum features | ||
| 第47行： | 第41行： | ||
* test large database with AMIDA | * test large database with AMIDA | ||
* test hidden layer knowledge transfer--xuewei | * test hidden layer knowledge transfer--xuewei | ||
| + | * test random last output layer when train MPE--zhiyuan | ||
===bilingual recognition=== | ===bilingual recognition=== | ||
| 第62行： | 第57行： | ||
====W2V based document classification==== | ====W2V based document classification==== | ||
| − | * | + | * APSIPA paper |
* CNN adapt to resolve the low resource problem | * CNN adapt to resolve the low resource problem | ||
| − | === | + | ===Pair-wise LM=== |
| − | * draft paper of journal | + | * draft paper of journal |
===Order representation === | ===Order representation === | ||
2015年6月10日 (三) 07:59的最后版本
Speech Processing
AM development
Environment
RNN AM
- morpheme RNN-zhiyuan
- ==== Mic-Array ====
- hold
- Change the prediction from fbank to spectrum features
- investigate alpha parameter in time domian and frquency domain
- ALPHA>=0, using data generated by reverber toolkit
- consider theta
- compute EER with kaldi
RNN-DAE(Deep based Auto-Encode-RNN)
- hold
- deliver to mengyuan
Speaker ID
- DNN-based sid --Tian Lan
Ivector&Dvector based ASR
- hold --Tian Lan
- Cluster the speakers to speaker-classes, then using the distance or the posterior-probability as the metric
- Direct using the dark-knowledge strategy to do the ivector training.
- Ivector dimention is smaller, performance is better
- Augument to hidden layer is better than input layer
- train on wsj(testbase dev93+evl92)
Dark knowledge
- Ensemble using 100h dataset to construct diffrernt structures -- Mengyuan
- adaptation English and Chinglish
- Try to improve the chinglish performance extremly
- unsupervised training with wsj contributes to aurora4 model --Xiangyu Zeng
- test large database with AMIDA
- test hidden layer knowledge transfer--xuewei
- test random last output layer when train MPE--zhiyuan
bilingual recognition
- hold
- http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zxw&step=view_request&cvssid=359 --Zhiyuan Tang and Mengyuan
language vector
- train DNN with language vector--xuewei
Text Processing
RNN LM
- character-lm rnn(hold)
- lstm+rnn
- check the lstm-rnnlm code about how to Initialize and update learning rate.(hold)
W2V based document classification
- APSIPA paper
- CNN adapt to resolve the low resource problem
Pair-wise LM
- draft paper of journal
Order representation
- modify the objective function(hold)
- sup-sampling method to solve the low frequence word(hold)
- journal paper
binary vector
- nips paper
Stochastic ListNet
- done
relation classifier
- done
plan to do
- combine LDA with neural network
