Task previous

来自cslt Wiki
2016年10月16日 (日) 12:17Tangzy讨论 | 贡献的版本

(差异) ←上一版本 | 最后版本 (差异) | 下一版本→ (差异)
跳转至: 导航搜索

Task To Do

Speech Recognition

CTC expanded

  • Voice activity detection
  • LSTM+CTC
  • TDNN+CTC
  • BLANK as silence, others as speech
  • Keyword detection
  • Character/Word-level, external key-word fst
  • G-fst need to be signal word?
  • Emotion recognition
  • LSTM-CTC

Network architecture test

  • chain model:
  • tdnn + simple-lstm
  • only keep forget gate
  • ctc+mpe
  • similar to chain training

Spiral Joint Training of SPEECH and SPEAKER

  • ASR & SID parallel training and benefit mutual

Small data-set and Big model

  • Investigate the efficiency of pre-training on small/big-model using dark-knowledge

Low-resource language improvement

  • SID
  • How to improve low-resource speaker

End-to-End speech recognition

  • Discriminative-Learning code implementation
  • Zhiyuan Tang

Multi-task

  • Fusion of speech-recognition and speech-rate
  • Xiangyu Zeng
  • Self-informed neural network structure learning
  • Mengyuan Zhao

Integrate the class information to HCLG fst for speech recognition

  • zhiyuan

Distant speech recognition

  • RNN-DAE: echo or reverberation
  • Xuewei Zhang/Zhiyuan Tang/Mengyuan Zhao/Zhiyong Zhang
  • Reverberation
  • Mutli-microphones
  • (Lasso),Xuewei Zhang

Voice conversation

  • hold

Sparse DNN

  • Zhiyuan Tang

Correlation based SENONE cluster

NN Multi-GPU parallel traing

  • Multi-GPU using data parallelization
  • Sheng Su
  • nnet3 mpe
  • Xuewei Zhang

Audio Embedding

  • Ke Ning

RNN training accelerating

Data selection

  • Zhiyong Zhang
  • Sub-modular data selection
  • Objective-function loss training self-adaptation

Decoder

  • Confidence output for task-required

Speaker Verification

binary code

  • Lantian Li

RNN-ivector

  • Lantian Li

DNN clustering

  • Lantian Li

Task DONE

Multi-Mode features based VAD

  • Shi Yin

DNN based Language identification and Speaker identification

  • Xuewei Zhang/Zhiyuan Tang

Neural network visulization

  • Mian Wang,DONE

Dark knowledge

  • Mengyuan Zhao, Xiangyu Zeng, Zhiyong Zhang, Chao Liu

Normal RNN speech recognition

  • Mengyuan Zhao

Monmentum-like Hessien-Free acceleration

  • Nestrov/Adagrad/AdaDelta/AdaM
  • Zhiyong Zhang/Xiangyu Zeng

Activation value normalization through time --Batch Normalization

  • Zhiyong Zhang

Mix-training Balance decision tree

  • Zhiyong Zhang

20-h Chinese data-set release

  • Xuewei Zhang

Unbound activation function(Rectifier/Maxout/Pnorm) go-through searching method

  • nne3 test --Xuewei Zhang

Technical Report To Write

1, DNN-DAE based noise cancellation -- Xiangyu Zeng / Mengyuan Zhao / Zhiyong Zhang  --DONE
2, Speech Rate DNN speech recognition --Shi Yin/Xiangyu Zeng --DONE
3, CNN+fbank feature combination --Mian Wang /Yiye Lin /Mengyuan Zhao /Shi Yin
4, Uyghur low-resource acoustic model enhancement -- Shi Yin / Mengyuan Zhao / Zhiyong Zhang --DONE
5, Uyghur 20h database release --Kaer /Shi Yin --DONE
6,Dark-Knowledge Transfer
   *: Xiangyu Zeng/ Mengyuan Zhao / Zhiyong Zhang

Paper to Write

Patent done

  • A method of new word enhancement for speech recognition --Yue Zhang

Project

  • Xiaomi TV
  • Mengyuan Zhao/Zhiyong Zhang
  • TAG-lm & Domain-specific general lm
  • Chinese-English mix-training