“Zhiyong Zhang”版本间的差异

来自cslt Wiki
跳转至: 导航搜索
Technical Report To Write
 
(相同用户的6个中间修订版本未显示)
第1行: 第1行:
  
=Task To Do=
 
* 1, RNN speech recognition (Tied-context-dependent-state and End-to-End)
 
* 2, Real environment noise cancellation(DNN-DAE/CNN-DAE/RNN-DAE: echo or reverberation)
 
* 3, Integrate the class information to HCLG fst for speech recognition
 
* 4, Multi-Mode features based VAD
 
* 5, DNN based Language identification and Speaker identification
 
* 6, Distant speech recognition(Reverberation, Mutli-microphones)
 
* 7, Voice conversation
 
* 8, Unbound activation function(Rectifier/Maxout/Pnorm) go-through searching method.
 
* 9, Sparse DNN
 
* 10, Neural network visulization
 
 
=Technical Report To Write=
 
* 1, DNN-DAE based noise cancellation
 
* 2, Speech Rate DNN speech recognition
 
* 3, CNN+fbank feature combination
 
* 4, Uyghur low-resource acoustic model enhancement
 
* 5, Uyghur 20h database release
 
  
 
=Papers To Read =
 
=Papers To Read =
第40行: 第22行:
 
* Testing on 100h-Ch+100h-En, better performance observed.
 
* Testing on 100h-Ch+100h-En, better performance observed.
 
* Now testing the source code on 1400h_8k data, but stange decoding results got.Need to further investigate.
 
* Now testing the source code on 1400h_8k data, but stange decoding results got.Need to further investigate.
 +
 +
=Reading Lists=
 +
*[[媒体文件:Efficient_mini-batch_training_for_stochastic_optimization.pdf |苏圣 2015-10-29 Efficient_mini-batch_training_for_stochastic_optimization ]]
 +
*[[媒体文件:2015_Fitnets-Hints for thin deep nets.pdf |张之勇 2015-10-29 2015_Fitnets-Hints for thin deep nets ]]
 +
*http://www.cs.cmu.edu/~muli/file/minibatch_sgd.pdf

2015年10月29日 (四) 07:14的最后版本


Papers To Read

  • 1, Learned-Norm pooling for deep feedforward and recurrent neural networks


Task schedules

Summary

   --------------------------------------------------------------------------------------------------------
    Priority | Tasks name                    |      Status          |     Notions
   --------------------------------------------------------------------------------------------------------    
        1    | Bi-Softmax                    | ■■■□□□□□□□ | 1400h am training and problem fixing
   --------------------------------------------------------------------------------------------------------
        2    | RNN+DAE                       | □□□□□□□□□□ |
   --------------------------------------------------------------------------------------------------------

Speech Recognition

Multi-lingual Am training

Bi-Softmax

  • Using two distinct softmax for English and Chinese data.
  • Testing on 100h-Ch+100h-En, better performance observed.
  • Now testing the source code on 1400h_8k data, but stange decoding results got.Need to further investigate.

Reading Lists