《双周报告.ppt》由会员分享,可在线阅读,更多相关《双周报告.ppt(24页珍藏版)》请在得力文库 - 分享文档赚钱的网站上搜索。
1、双周报告 石颖 2016年7月5号PART 1:THE PROCESS OF ASRPART 2:NEURAL NETWORKPART 3:EXPERIMENTPART1:THE PROCESS OF ASRPROCESS OF ASR(SPEECHTEXT)DNNSpeech(vector)MFCC/FbankAcoustic feature(matrix frame by frame)textwordphonetriphonestateGMM-HMMINPUTOUTPUT(TRAIN)DNNINPUTOUTPUTDECODETEXTPDFPART2:NEURAL NETWORKBP(BAC
2、K PROPAGATION)Traditional network can not deal with linearlynon-separable directly,but itseasy for BP.The Architecture is:Input hiddenoutput But there is a shortcoming:Gradient disappearance RBM(RESTRICTED BOLTZMANN MACHINES)Give energy function E(v,h)between Visible and Hidden,p(v,h),p(v|h),p(h|v)P
3、(v|h)p(h|v)DBN(DEEP BELIEVE NERWORK)1-NThe process of trian DBN1:Unsupervise learning RBM initalize the wight(pre-training)2:Supervise learning BP fine tune the weightBecause of the pre-train the DBN can prevent the disappeard of gradient RNN AND LSTMGood at deal with sequence dataRNN is good at dea
4、l with the sequence data but the gradient disappearance on time dimension is a shotrcomingLSTMCNN AND TDNN1.Sparse Connectivity2.Shared Weights3.convolution-sub-samplingTDNNSUMMARIZESNNThe main task of all kind of neural network is to find:F(x)yActually it is:F(x)=yPART 3:EXPERIMENTTHERE ARE FOUR EX
5、PERIMENTS1.process of ASR2.tdnn3.lstm4.simplyfying lstm(revise the config document and implement by myself)PROCESS OF ASRLSTMCONFIGPicture from config%WER 78.66 63827/81139,472 ins,15198 del,48157 sub exp/nnet3/lstm_S_ld5/decode_test_word/wer_9_0.0Simplyfying LSTMLSTM%WER 24.41 19803/81139,224 ins,1144 del,18435 sub exp/nnet3/lstm_1_ld5/decode_test_word/wer_9_0.0SUMMARIZESShellRead code How to handle the bug of programHow to read configHow to creat a network by revise the config