ASR:2015-05-11

Speech Processing

AM development

Environment

• grid-15 often does not work

RNN AM

• details at http://liuc.cslt.org/pages/rnnam.html
• Test monophone on RNN using dark-knowledge --Chao Liu
• run using wsj,MPE --Chao Liu
• run bi-directon --Chao Liu
• modify code --Zhiyuan

Mic-Array

• 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 --Zhiyong Zhang
• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=261

Speaker ID

• DNN-based sid --Yiye Lin
• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zhangzy&step=view_request&cvssid=327

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.
• 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
• train on wsj(testbase dev93+evl92)

Dark knowledge

• Ensemble using 100h dataset to construct diffrernt structures -- Mengyuan

• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zxw&step=view_request&cvssid=264 --Zhiyong Zhang

• adaptation for chinglish under investigation --Mengyuan Zhao

• Try to improve the chinglish performance extremly

• unsupervised training with wsj contributes to aurora4 model --Xiangyu Zeng

• test large database with AMIDA

bilingual recognition

• http://cslt.riit.tsinghua.edu.cn/cgi-bin/cvss/cvss_request.pl?account=zxw&step=view_request&cvssid=359 --Zhiyuan Tang and Mengyuan

Text Processing

RNN LM

• rnn

• test the ppl and code the character-lm(hold)

• lstm+rnn

• check the lstm-rnnlm code about how to Initialize and update learning rate.(hold)

W2V based document classification

• make a technical report about document classification using CNN --yiqiao
• CNN adapt to resolve the low resource problem

Translation

• similar-pair method in English word using translation model.

• result:wer:70%-50% on top1.
• change the AM model

Order representation

• modify the objective function
• sup-sampling method to solve the low frequence word

Stochastic ListNet

• using sampling method and test

relation classifier

• test the bidirectional neural network(B-RNN) and get a little improvement