plateform
stringclasses
1 value
repo_name
stringlengths
13
113
name
stringlengths
3
74
ext
stringclasses
1 value
path
stringlengths
12
229
size
int64
23
843k
source_encoding
stringclasses
9 values
md5
stringlengths
32
32
text
stringlengths
23
843k
github
singaxiong/SignalGraph-master
abs2wav.m
.m
SignalGraph-master/signal/feature/abs2wav.m
539
utf_8
cea0ad8bb8a4673102d73409fe17b0f8
% len is frame size in samples, % len1 is frame overlap in samples % mag_x is the magnitude (no log) % phase_x is the phase returned by angle(x), where x is the complex Fourier % coefficients. function wav = abs2wav(mag_x, phase_x, len, len1) nFFT = size(mag_x,2)*2-2; img = sqrt(-1); x_phase = mag_x .* ( cos(phase_x...
github
singaxiong/SignalGraph-master
mel_center.m
.m
SignalGraph-master/signal/feature/mel_center.m
233
utf_8
42df32eef0bda360fad0edf370ab6a28
% Compute the center frequency of the mel filterbank function [mel_center_freq] = mel_center(linear_samp, N_mel) for i=1:N_mel tmp_mel = i*linear2mel(linear_samp/2) / (N_mel+1); mel_center_freq(i) = mel2linear(tmp_mel); end
github
singaxiong/SignalGraph-master
fbank2mfcc.m
.m
SignalGraph-master/signal/feature/fbank2mfcc.m
672
utf_8
2dffedf365407ef80cee6e6bcba2e1ff
% Calculate the MFCC from filter bank % Author: Xiao Xiong % Created: 4 Feb 2005 % Last modified: 4 Feb 2005 function [feature] = fbank2mfcc(fbank,logE,DO_BLIND_EQUALIZATION); global bias; bias = zeros(12,1); [N_vector, N_melBank]= size(fbank); % for i=1:N_vector % % calculate the DCT of the vector. The mfcc se...
github
singaxiong/SignalGraph-master
mel_bank_range.m
.m
SignalGraph-master/signal/feature/mel_bank_range.m
1,392
utf_8
d88a751106bcc2f90d500cbc5e7bd889
% Compute the frequency range of the mel filterbank % Inputs: % start_freq: the lowerest linear frequency included in the calculation % linear_samp: sampling frequency % N_mel: number of mel banks used % Output: % bin_lower_edge: the lower edge linear frequency of each mel bank % bin_upper...
github
singaxiong/SignalGraph-master
comp_delta2.m
.m
SignalGraph-master/signal/feature/comp_delta2.m
948
utf_8
0eaf51f57f0d1c064590a2d00bd7f7b6
% this function compute the derivatives of the static coefficients. Its % implementation follows that of the HMM Toolkit 3.2 function delta_coef = comp_delta(static_coef, DELTAWINDOW) [N_vec, N_cep] = size(static_coef); first_vec = static_coef(1,:)'; last_vec = static_coef(N_vec,:)'; static_coef = static_coef'; for...
github
singaxiong/SignalGraph-master
comp_log_Mel.m
.m
SignalGraph-master/signal/feature/comp_log_Mel.m
2,740
utf_8
b0a81240b4ab47bcf09691a0b5b029e8
% comp_log_Mel receives a time domain speech signal and produces its % log Mel filterbank coefficients. It is the Matlab counterpart of the % feature extraction program provided with AURORA2 database. % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last Modified: 28 Jul, 2005 % Inputs: % x 1-D time domain sign...
github
singaxiong/SignalGraph-master
wav2realImag.m
.m
SignalGraph-master/signal/feature/wav2realImag.m
1,254
utf_8
c105cd6fbaa3be19d16a8eee9837b422
% comp_log_Mel receives a time domain speech signal and produces its % log Mel filterbank coefficients. It is the Matlab counterpart of the % feature extraction program provided with AURORA2 database. % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last Modified: 28 Jul, 2005 % Inputs: % x 1-D time domain sign...
github
singaxiong/SignalGraph-master
complexSpec2wav.m
.m
SignalGraph-master/signal/feature/complexSpec2wav.m
329
utf_8
fc69e16214863c9adbd799bef6cb15e8
% len is frame size, len1 is overlap function wav = complexSpec2wav(complex_x, framelen, overlap) nFFT = size(complex_x,2)*2-2; img = sqrt(-1); complex_x(:,nFFT/2+2:nFFT) = conj(complex_x(:,nFFT/2:-1:2)); xi = ifft(complex_x'); xi = real(xi); wav = my_ola(xi, framelen, overlap); %A = [1 -0.97]; %wav = filter(1, A, ...
github
singaxiong/SignalGraph-master
wav2abs_multi.m
.m
SignalGraph-master/signal/feature/wav2abs_multi.m
1,304
utf_8
321cac381148bfda0d77b4aca6e8bb83
% comp_log_Mel receives a time domain speech signal and produces its % log Mel filterbank coefficients. It is the Matlab counterpart of the % feature extraction program provided with AURORA2 database. % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last Modified: 28 Jul, 2005 % Inputs: % x MxN time domain sign...
github
singaxiong/SignalGraph-master
wav2abs_fine.m
.m
SignalGraph-master/signal/feature/wav2abs_fine.m
1,083
utf_8
ebd9d1755f4b796f549985beac9a34d0
% comp_log_Mel receives a time domain speech signal and produces its % log Mel filterbank coefficients. It is the Matlab counterpart of the % feature extraction program provided with AURORA2 database. % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last Modified: 28 Jul, 2005 % Inputs: % x 1-D time domain sign...
github
singaxiong/SignalGraph-master
mel_window_FE.m
.m
SignalGraph-master/signal/feature/mel_window_FE.m
1,567
utf_8
9397249f4155400aca5fd30d97961956
% mel_window generates the mel triangle window when given the number of mel % bins and the number of linear frequency bins % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last modified: 29 Jul, 2005 % Inputs: % N_mel_bin number of mel banks used % N_linear_freq_bin number of fourier transform bins % ...
github
singaxiong/SignalGraph-master
my_gencoswin.m
.m
SignalGraph-master/signal/feature/my_gencoswin.m
3,116
utf_8
8c55ee7cdaee8096567c3d8a4b85a50c
function [w,msg,msgobj] = my_gencoswin(varargin) %GENCOSWIN Returns one of the generalized cosine windows. % GENCOSWIN returns the generalized cosine window specified by the % first string argument. Its inputs can be % Window name - a string, any of 'hamming', 'hann', 'blackman'. % N - le...
github
singaxiong/SignalGraph-master
comp_dynamic_feature2.m
.m
SignalGraph-master/signal/feature/comp_dynamic_feature2.m
305
utf_8
32e76902707ac3a10540ad509a550090
% % compute delta and acceleration features from static features % function feature = comp_dynamic_feature2(feature) D = size(feature,2); feature(:,D+1:2*D) = comp_delta(feature(:,1:D),3); feature(:,2*D+1:3*D) = comp_delta(feature(:,D+1:2*D),2); feature(:,3*D+1:4*D) = comp_delta(feature(:,2*D+1:3*D),2);
github
singaxiong/SignalGraph-master
comp_dynamic_feature.m
.m
SignalGraph-master/signal/feature/comp_dynamic_feature.m
340
utf_8
1359972e2c4c95b2e5fc8c5a68bda81f
% % compute delta and acceleration features from static features % function output = comp_dynamic_feature(feature, delta_order, acc_order) if nargin<3 acc_order = 2; end if nargin<2 delta_order = 3; end static = feature; delta = comp_delta(static,delta_order); acc = comp_delta(delta,delta_order); output = [st...
github
singaxiong/SignalGraph-master
wav2abs.m
.m
SignalGraph-master/signal/feature/wav2abs.m
1,441
utf_8
acc1c6cdf7dae7f52a09cb4cc0be84c0
% comp_log_Mel receives a time domain speech signal and produces its % log Mel filterbank coefficients. It is the Matlab counterpart of the % feature extraction program provided with AURORA2 database. % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last Modified: 28 Jul, 2015 % Inputs: % x 1-D time domain sign...
github
singaxiong/SignalGraph-master
wav2abs_full.m
.m
SignalGraph-master/signal/feature/wav2abs_full.m
956
utf_8
f3538c7ee35e10aa09a766bf6576dafc
% comp_log_Mel receives a time domain speech signal and produces its % log Mel filterbank coefficients. It is the Matlab counterpart of the % feature extraction program provided with AURORA2 database. % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last Modified: 28 Jul, 2005 % Inputs: % x 1-D time domain sign...
github
singaxiong/SignalGraph-master
linear2mel.m
.m
SignalGraph-master/signal/feature/linear2mel.m
175
utf_8
2fc5cddd398f88911a754ca72d7b723a
% convert linear frequency to mel frequency function [mel_freq] = linear2mel(linear_freq) for i=1:length(linear_freq) mel_freq(i) = 2595*log10(1+linear_freq(i)/700); end
github
singaxiong/SignalGraph-master
abs2Mel.m
.m
SignalGraph-master/signal/feature/abs2Mel.m
1,002
utf_8
61ced490e134b7fadcfa7c5f342b359e
% comp_log_Mel receives a time domain speech signal and produces its % log Mel filterbank coefficients. It is the Matlab counterpart of the % feature extraction program provided with AURORA2 database. % Author: Xiao Xiong % Created: 18 Jul, 2005 % Last Modified: 28 Jul, 2005 % Inputs: % x 1-D time domain sign...
github
singaxiong/SignalGraph-master
wav2root_fbank.m
.m
SignalGraph-master/signal/feature/wav2root_fbank.m
351
utf_8
c5a2ed801534d6f84c3d6219522800ee
% Calculate the log mel filter bank from waveform % Author: Xiao Xiong % Created: 4 Feb 2005 % Last modified: 4 Feb 2005 function [fbank] = wav2root_fbank(is_file_name, x, frame_shift); if is_file_name ==1 x = readNIST(x); end if nargin < 3 fbank = (abs2Mel( wav2abs(x) )).^0.1; else fbank = (abs2Mel( wav2...
github
singaxiong/SignalGraph-master
comp_logE.m
.m
SignalGraph-master/signal/feature/comp_logE.m
950
utf_8
542684ab52f338dd2d434e8d55dec4c8
% Compute the log energy of the speech signal from the time domain samples % Note that: to compute LogE, there is no need to do DC_offset removing and % Preemphasis. However, my experiments shows that DC offset removing and % preemphasis in LogE computation improves the recognition accuracy quite a % lot function logE ...
github
singaxiong/SignalGraph-master
mydct.m
.m
SignalGraph-master/signal/feature/mydct.m
918
utf_8
3ba981b441196a5a4f2837da1a051176
% Compute the DCT of the filter bank. The program implements the DCT % computation of the WI007 of AURORA project. % There will usually be 13 DCT coefficients. The first coefficient is % computed and appended after the other two coefficients. so the sequence % of the coefficients is c1c2...c12c0 % Author: Xiao Xiong %...
github
singaxiong/SignalGraph-master
mel_center_FE.m
.m
SignalGraph-master/signal/feature/mel_center_FE.m
801
utf_8
caab013958de7f704612fbdd89605a17
% Compute the center frequency of the mel filterbank % Inputs: % start_freq: the lowerest linear frequency included in the calculation % linear_samp: sampling frequency % N_mel: number of mel banks used % Output: % mel_center_freq: the center linear frequency of each mel bank % bin_upper_e...
github
singaxiong/SignalGraph-master
wav2fbank.m
.m
SignalGraph-master/signal/feature/wav2fbank.m
358
utf_8
0da521214e4a44fb5ee74765d90348aa
% Calculate the log mel filter bank from waveform % Author: Xiao Xiong % Created: 4 Feb 2005 % Last modified: 4 Feb 2005 function [fbank] = wav2fbank(x, fs, frame_shift, nMel) if nargin < 2 fs = 8000; end if nargin < 3 frame_shift = 0.01; end if nargin<4 nMel = 23; end absX = wav2abs(x,fs,frame_shift); f...
github
singaxiong/SignalGraph-master
DC_remove.m
.m
SignalGraph-master/signal/feature/DC_remove.m
143
utf_8
369a94d92ccb9b5428d677ffa8fd5774
% dc_remove remove the DC offset of the time domain signal function [z] = DC_remove(x,a) Num = [a -a]; Dem = [1 -a]; z = filter(Num,Dem,x);
github
singaxiong/SignalGraph-master
mel_window.m
.m
SignalGraph-master/signal/feature/mel_window.m
1,279
utf_8
39b766f4f990ea3f446f91c1131615e8
% Form the mel triangle window when given the number of mel bins and the number of linear % frequency bins function [window] = mel_window( N_mel_bin, N_linear_freq_bin, linear_samp ) % N_linear_freq_bin is the number of frequency bins used by FFT, e.g. 256 FFT_length = N_linear_freq_bin*2; % find the mel bin centers...
github
singaxiong/SignalGraph-master
compute_modulation_stft.m
.m
SignalGraph-master/signal/feature/compute_modulation_stft.m
2,027
utf_8
1daa55ff55f82a9946d091cbb1ef0391
% This function compute the modulation spectrum of speech. The modulation % spectrum can be derived from several types of feature trajectories: 1) % log spectrogram; 2) log Mel filterbanks; 3) MFCC. Note that it is % different from another definition of modulation spectrum that derived % from the evelop of band-passed ...
github
singaxiong/SignalGraph-master
modified_group_delay_feature_original.m
.m
SignalGraph-master/signal/phase/modified_group_delay_feature_original.m
2,785
utf_8
c8e9a230dee83b8e796fa0ee9fdbd9df
function [grp_phase, cep] = modified_group_delay_feature_original(speech, fs, rho, gamma, num_coeff) %input: % file_name: path for the waveform. The waveform should have a header % rho: a parameter to control the shape of modified group delay spectra % gamma: a parameter to control the shape of the modif...
github
singaxiong/SignalGraph-master
BPD2phase.m
.m
SignalGraph-master/signal/phase/BPD2phase.m
1,192
utf_8
c66111afb271d5760afd5d8cefeb3a2c
% This function compute the instantaneous frequency from phase spectrogram % Input: % phase: D x T matrix of phase spectrogram, where D is FFT-bin number and % T is number of frames. % Output: % instantaneous frequency of the size size as phase. % The implementation is based on equation (3) of % Krawcz...
github
singaxiong/SignalGraph-master
get_principal_value.m
.m
SignalGraph-master/signal/phase/get_principal_value.m
160
utf_8
fa050adfd831f1e7cff6bf40615b91be
function phase = get_principal_value(phase) idx = find(phase>pi); phase(idx) = phase(idx) - 2*pi; idx = find(phase<-pi); phase(idx) = phase(idx) + 2*pi; end
github
singaxiong/SignalGraph-master
comp_BPD.m
.m
SignalGraph-master/signal/phase/comp_BPD.m
1,184
utf_8
42aefc3b41815d423a72255c14dd63f4
% This function compute the instantaneous frequency from phase spectrogram % Input: % phase: D x T matrix of phase spectrogram, where D is FFT-bin number and % T is number of frames. % Output: % instantaneous frequency of the size size as phase. % The implementation is based on equation (3) of % Krawcz...
github
singaxiong/SignalGraph-master
comp_group_delay.m
.m
SignalGraph-master/signal/phase/comp_group_delay.m
1,170
utf_8
5c4a06c19f274eff4986894fd82fe1b1
% This function compute the instantaneous frequency from phase spectrogram % Input: % phase: D x T matrix of phase spectrogram, where D is FFT-bin number and % T is number of frames. % Output: % instantaneous frequency of the size size as phase. % The implementation is based on equation (3) of % Krawcz...
github
singaxiong/SignalGraph-master
comp_instan_freq.m
.m
SignalGraph-master/signal/phase/comp_instan_freq.m
1,410
utf_8
1be1e32f7f6dc803138a49689a5f9a4a
% This function compute the instantaneous frequency from phase spectrogram % Input: % phase: D x T matrix of phase spectrogram, where D is FFT-bin number and % T is number of frames. % Output: % instantaneous frequency of the size size as phase. % The implementation is based on equation (3) of % Krawcz...
github
singaxiong/SignalGraph-master
modified_group_delay_raw.m
.m
SignalGraph-master/signal/phase/phase_feature_extraction/modified_group_delay_raw.m
607
utf_8
1667e8c84f970363361814fe3a499f25
function grp_phase = modified_group_delay_raw(sp_complex, sp_delay) %input: % sp_complex: STFT of target waveform x(n) which contain both magnitude and phase information % sp_delay: STFT of modified target waveform n*x(n) % %output: % grp_phase: modifed group delay spectrogram x_spec = sp_complex'; y_s...
github
singaxiong/SignalGraph-master
DC_remove.m
.m
SignalGraph-master/signal/phase/phase_feature_extraction/DC_remove.m
144
utf_8
6c457620184fb04dabaa3ea081625a42
% dc_remove remove the DC offset of the time domain signal function [z] = dc_remove(x,a); Num = [a -a]; Dem = [1 -a]; z = filter(Num,Dem,x);
github
singaxiong/SignalGraph-master
DelaySum_Beamformer_fast.m
.m
SignalGraph-master/signal/array/DelaySum_Beamformer_fast.m
1,145
utf_8
2458ec4efa870d605fa4e01bbc431692
%%%%%%@function: Delay and Sum beamformer%%%%%%%%%%%% function [xout] = DelaySum_Beamformer_fast(sig,tde_est) %%start to process with beamformer nChs = size(sig,2); %%perform Delay and Sum beamforming lfft = 1024; %32ms window size for beamforming ShiftP = 0.25; Wsz = lfft; INC = Wsz*ShiftP; W = hann(Wsz); %Apply the...
github
singaxiong/SignalGraph-master
ApplyConstRirNoise.m
.m
SignalGraph-master/signal/array/ApplyConstRirNoise.m
2,449
utf_8
d77e5e02396b18b03746fc4c07194621
% This function apply RIR to the clean speech signal, and also optionally % add additive noise at a specified SNR. %%%% function [y, rev_y, direct_signal2]=ApplyConstRirNoise(x,fs,RIR,NOISE,SNRdB, useGPU) if nargin<6 useGPU = 0; end % calculate direct+early reflection signal for calculating SNR if useGPU x = g...
github
singaxiong/SignalGraph-master
ComplexSpectrum2SpatialCov.m
.m
SignalGraph-master/signal/array/ComplexSpectrum2SpatialCov.m
2,250
utf_8
95c72cd6ff9ed4dbf11ec449e3cd1234
% this function compute the spatial covariance matrix from complex Fourier % transform of array signals. % Inputs: % X: D x N x T matrix of Fourier transform coefficients. D is the number % of frequency bins, N is the number of microphone channels, and T is the % number of frames % context_size: number of conte...
github
singaxiong/SignalGraph-master
gen_gcc_spec.m
.m
SignalGraph-master/signal/array/gen_gcc_spec.m
2,177
utf_8
bc50df3aa26733ae7229b7892d7464e9
% This function generate spectrogram and GCC for microphone array input function [gcc, magnitude, phase, gcc_interp] = gen_gcc_spec(wav, para) fs = para.fs; % compute the spectrogram magnitude = {}; phase = {}; if para.genSpec specWinSize = para.specWinSize; specShift = para.specShift; for j=1:size(wav,1) ...
github
singaxiong/SignalGraph-master
getCorrelationVector_fast2.m
.m
SignalGraph-master/signal/array/getCorrelationVector_fast2.m
2,336
utf_8
502f4d8cbd3189b15571414071422eb7
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %%%@function: Calculate the correlation vectors for all the permutations of the given input channels. %%%The method uses inverse Fourier transform of the cross-power spectrum to %%%obtain the cross-correlation based on GCC-PHAT (C Knapp...
github
singaxiong/SignalGraph-master
CHiME4_setup.m
.m
SignalGraph-master/signal/array/imageRIR/CHiME4_setup.m
1,979
utf_8
a4ae23b8b7655aefd581e4abbdbd58e1
% This function set the parameters for simulating RIRs for CHiME3/4 % microphone array. % See Lehmann/ISM_setup.m for how to set the parameters for image method. % Chenglin Xu / Xiong Xiao, 22 Jun 2016 % function [SetupStruc] = CHiME4_setup(t60, room) SetupStruc.Fs = 16000; % sampling frequency in Hz S...
github
singaxiong/SignalGraph-master
ISM_RoomResp_GPU.m
.m
SignalGraph-master/signal/array/imageRIR/ISM_RoomResp_GPU.m
13,299
utf_8
d183251541364838bfdbf97d702b8705
function [RIRvec] = ISM_RoomResp_GPU(Fs,beta,rt_type,rt_val,X_src,X_rcv,room,varargin) %ISM_RoomResp RIR based on Lehmann & Johansson's image-source method % % RIR = ISM_RoomResp(Fs,BETA,RT_TYPE,RT_VAL,SOURCE,SENSOR,ROOM) % RIR = ISM_RoomResp( ... ,'arg1',val1,'arg2',val2,...) % % This function generates the room imp...
github
singaxiong/SignalGraph-master
ISM_RoomResp.m
.m
SignalGraph-master/signal/array/imageRIR/Lehmann/ISM_RoomResp.m
11,228
utf_8
27ca356f0d951e2e6aea01d8e430af95
function [RIRvec] = ISM_RoomResp(Fs,beta,rt_type,rt_val,X_src,X_rcv,room,varargin) %ISM_RoomResp RIR based on Lehmann & Johansson's image-source method % % RIR = ISM_RoomResp(Fs,BETA,RT_TYPE,RT_VAL,SOURCE,SENSOR,ROOM) % RIR = ISM_RoomResp( ... ,'arg1',val1,'arg2',val2,...) % % This function generates the room impulse...
github
singaxiong/SignalGraph-master
F_frame_select.m
.m
SignalGraph-master/graph/F_frame_select.m
1,270
utf_8
375fa81bac26883df9c669853c3013f0
% select frames from the input stream % function [output, validFrameMask] = F_frame_select(input_layer, curr_layer) input = input_layer.a; [D,T,N] = size(input); words = ExtractWordsFromString_v2(curr_layer.frameSelect); selectionType = words{1}; switch selectionType case 'last' % only select the last N fram...
github
singaxiong/SignalGraph-master
DNN_Cost10.m
.m
SignalGraph-master/graph/DNN_Cost10.m
27,098
utf_8
e45bdcdcefcdcd4a47b31d777d5b7b66
% This function does both the forward pass and backward pass of DNN. % The forward pass is like a multilayer nonlinear transform of the input, % while the backward pass computes the gradients. % The inputs of the function are: % theta - a 1-D array which contains all the parameers of the DNN % visible - the inpu...
github
singaxiong/SignalGraph-master
DetermineLayerParent.m
.m
SignalGraph-master/graph/DetermineLayerParent.m
432
utf_8
bc0ad65104dd9973d0b35a2f040b23b5
% find the parent of every layer function parent = DetermineLayerParent(layer) for i=1:length(layer) switch lower(layer{i}.name) case {'weight2activation', 'input'} parent{i} = []; otherwise immediateParent = i+layer{i}.prev; parent{i} = []; for ...
github
singaxiong/SignalGraph-master
F_LSTM_back.m
.m
SignalGraph-master/graph/F_LSTM_back.m
3,537
utf_8
e2f734bf4ab09f0a6c1aedd998937dcc
% Implement the forward pass of an LSTM layer. % Author: Xiong Xiao, Temasek Labs, NTU, Singapore. % Last modified: 13 Oct 2015 % function [LSTM_layer] = F_LSTM(input, LSTM_layer) % % The weight matrix of LSTM is organized as follows: % W = [ W_cf W_hf W_xf; % W_cc W_hc W_xc; % W_ci W_hi W_xi; % W_c...
github
singaxiong/SignalGraph-master
PadShortTrajectory.m
.m
SignalGraph-master/graph/PadShortTrajectory.m
2,210
utf_8
a52cedb0e0190356cef740a8d3bce207
% If a trajectory is shorter than others, as specified by the mask, we set % its first dimension a defined number. There are two usage of this % function: % 1. pad the output trajectory of a layer with a big negative number, e.g. -1e10 to notice % other layers the trajectory is shorter than others in the minibatch....
github
singaxiong/SignalGraph-master
F_matrix_multiply.m
.m
SignalGraph-master/graph/F_matrix_multiply.m
1,346
utf_8
1595e74865cd357bdf5a2a038d79f180
% compute matrix multiply X * Y % function [output, validFrameMask] = F_matrix_multiply(input_layers, curr_layer) X = input_layers{1}.a; Y = input_layers{2}.a; [Dx,Tx,Nx] = size(X); [Dy,Ty,Ny] = size(Y); if Tx ~=Dy fprintf('Error: matrix size not match each other\n'); end validFrameMask = []; if Nx==1 && Ny==1 ...
github
singaxiong/SignalGraph-master
F_jointCost.m
.m
SignalGraph-master/graph/F_jointCost.m
672
utf_8
d3af0dc03f17fa3dc67a52bde221fb7c
% compute the joint cost between neighboring data points % function cost = F_jointCost(input_layer, curr_layer) input = input_layer.a; [D,T,N] = size(input); dimension = curr_layer.dimension; % dimension defines along which dimension the smoothness are defined. cost = 0; if IsInGPU(input) cost = gpuArray.zeros(...
github
singaxiong/SignalGraph-master
F_complex2realImag.m
.m
SignalGraph-master/graph/F_complex2realImag.m
365
utf_8
5776ca4c1fa2703c1abe17b8b365ca64
% assume the input is a DxTxN tensor. The first D/2 dimensions are store the % real parts, and the last D/2 dimensions store the imaginary parts. The % function returns the corresponding complex numbers % function [output] = F_complex2realImag(input_layer) input = input_layer.a; realpart = real(input); imagpart = imag(...
github
singaxiong/SignalGraph-master
GetLastValidFrameIndex.m
.m
SignalGraph-master/graph/GetLastValidFrameIndex.m
295
utf_8
efb210983e12c300579a3373ec033c17
% given a mask of valid frames in sentences, return the an array of index of last % valid frames in each sentence. % function last_idx = GetLastValidFrameIndex(mask) T = size(mask,1); delta = mask(2:end,:) - mask(1:end-1,:); [max_delta, last_idx] = max(delta); last_idx(max_delta==0) = T; end
github
singaxiong/SignalGraph-master
F_log.m
.m
SignalGraph-master/graph/F_log.m
608
utf_8
0882ccca05b4d1586dfd3e7a1cdce88f
function [output,validFrameMask] = F_log(input_layer, const) input = input_layer.a; [D,T,N] = size(input); if strcmpi(class(gather(input(1))), 'single') const = single(const); end if N==1 output = log(input+const); validFrameMask = []; else [validFrameMask, variableLength] = getValidFrameMask(input_l...
github
singaxiong/SignalGraph-master
F_copyVec2Mat.m
.m
SignalGraph-master/graph/F_copyVec2Mat.m
537
utf_8
923dfe5e09c345809da06b84e8eb3878
% put a vector in selected positions of a matrix. % function [output] = F_copyVec2Mat(input_layer, curr_layer) input = input_layer.a; [D,T,N] = size(input); D1 = curr_layer.targetDims(1); D2 = curr_layer.targetDims(2); index2copy = curr_layer.index2copy; if IsInGPU(input(1)) output = gpuArray.zeros(D1, D2, T, N)...
github
singaxiong/SignalGraph-master
F_frame_shift.m
.m
SignalGraph-master/graph/F_frame_shift.m
846
utf_8
26b4b78616cd43b99a894fd1a3de79ba
% shift the frames of the input stream function [output, validFrameMask] = F_frame_shift(input_layer, curr_layer) input = input_layer.a; [D,T,N] = size(input); if N>1; [mask, variableLength] = GetValidFrameMask(input_layer); else variableLength = 0; end delay = curr_layer.delay; if delay>0 % positive delay means ...
github
singaxiong/SignalGraph-master
F_ll_gaussian.m
.m
SignalGraph-master/graph/F_ll_gaussian.m
781
utf_8
4f4dc09d56ce3eed34b6f0e517a97d64
% take the covariance matrix of input trajectories function [output, validFrameMask] = F_ll_gaussian(prev_layers, curr_layer) mu = prev_layers{1}.a; variance = prev_layers{2}.a; input = prev_layers{3}.a; LogLikelihood = -(input-mu).^2 ./ variance /2; LogLikelihood = LogLikelihood - 0.5 * log(2*variance*pi); LogLi...
github
singaxiong/SignalGraph-master
CheckTrajectoryLength.m
.m
SignalGraph-master/graph/CheckTrajectoryLength.m
361
utf_8
b25dc4508c27b9a043847c0a1769900e
% check whether the sequences are of the same length. % The sequences are aligned from begining. If some sequence is shorter, it % will be padded with a big negative number, i.e. -1e10, in the first dimension. % function [mask, variableLength] = CheckTrajectoryLength(data) mask = permute(data(1,:,:), [2 3 1]) == -1...
github
singaxiong/SignalGraph-master
F_SpatialCovMask.m
.m
SignalGraph-master/graph/F_SpatialCovMask.m
2,619
utf_8
ab900e03f0fc54f6115f4cb3fe6d00f4
% Estimate spatial covariance matrix for sentences using a mask. The mask % specifies speech presense probability at all time frequency locations, % with a 1 means speech present and 0 means speech absent. % function output = F_SpatialCovMask(prev_layers, curr_layer) mask = prev_layers{1}.a; data = prev_layers{2}.a; ...
github
singaxiong/SignalGraph-master
F_cov.m
.m
SignalGraph-master/graph/F_cov.m
210
utf_8
06c9bc99b65c2c8f170b7640fc18ca65
% take the covariance matrix of input trajectories function output = F_cov(input) [D,M,N] = size(input); if N==1 %output = input*input' / M; output = cov(input'); else % to be implemented end end
github
singaxiong/SignalGraph-master
prepareCostEvaluation2.m
.m
SignalGraph-master/graph/prepareCostEvaluation2.m
2,991
utf_8
fade4494db48d03f5ee85c5178d805e8
% This function is intended to be called from mean square erorr or cross % entropy cost function layers. It decides which input stream is system % output and which is desired target and scale of frames. % % The function also applies following operations on the output and target: % 3) scale: assign weights to diffe...
github
singaxiong/SignalGraph-master
B_LSTM.m
.m
SignalGraph-master/graph/B_LSTM.m
5,159
utf_8
e383ab26f0ce045ed187a46c29103dfd
% Implement the back propagation of an LSTM layer. % Author: Xiong Xiao, Temasek Labs, NTU, Singapore. % Last modified: 13 Oct 2015 % function [grad, grad_W, grad_b] = B_LSTM(input_layer, LSTM_layer, future_layers) input = input_layer.a; W = LSTM_layer.W; if strcmpi(class(input), 'gpuArray'); useGPU=1; else useGPU = 0;...
github
singaxiong/SignalGraph-master
F_cmn.m
.m
SignalGraph-master/graph/F_cmn.m
812
utf_8
32b31f850a5e4974f643dd42068b88aa
function [output,validFrameMask] = F_cmn(input_layer) input = input_layer.a; [D,T,N] = size(input); if N==1 output = CMN(input')'; validFrameMask = []; else [validFrameMask, variableLength] = getValidFrameMask(input_layer); if variableLength input2 = ExtractVariableLengthTrajectory(input, vali...
github
singaxiong/SignalGraph-master
F_enframe.m
.m
SignalGraph-master/graph/F_enframe.m
247
utf_8
097d431913b1609386da4fb50b3f93df
function output = F_enframe(input, frame_len, frame_shift) useGPU = strcmpi(class(input), 'gpuArray'); % do not use GPU for enframe output = my_enframe(gather(input), frame_len, frame_shift); if useGPU output = gpuArray(output); end end
github
singaxiong/SignalGraph-master
B_log.m
.m
SignalGraph-master/graph/B_log.m
167
utf_8
fb87c979868fa4829cdc19e364667f16
function grad = B_log(future_layers, input,curr_layer) future_grad = GetFutureGrad(future_layers, curr_layer); grad = 1./(input+curr_layer.const).*future_grad; end
github
singaxiong/SignalGraph-master
ExpandContext_v2.m
.m
SignalGraph-master/graph/ExpandContext_v2.m
1,445
utf_8
8497d7de8b32d3ae0e73599b657cc880
% Concatenate the neighbouring frames vectors to higher dimensional vectors % Input is Number of dimension x number of frames function [y] = ExpandContext_v2(x, context, window_type) [dim, nFr, nSeg] = size(x); if nargin<3 window_type = 'null'; end context_size = length(context); if context_size > 1 if 0 ...
github
singaxiong/SignalGraph-master
B_dynamic_feat.m
.m
SignalGraph-master/graph/B_dynamic_feat.m
1,569
utf_8
66daf2dc67027079438eb1d24766be48
function grad = B_dynamic_feat(curr_layer, future_layers) output = curr_layer.a; [dim,nFr,nSeg] = size(output); dimS = dim/3; precision = class(gather(output(1))); D = genDeltaTransform(nFr, 2, 1, precision); A = D*D; D = full(D); A = full(A); fgrad = GetFutureGrad(future_layers, curr_layer); if nSeg==1 grad = f...
github
singaxiong/SignalGraph-master
F_realImag2complex.m
.m
SignalGraph-master/graph/F_realImag2complex.m
419
utf_8
d08b1be2e58f183294691a85e5998a27
% assume the input is a DxTxN tensor. The first D/2 dimensions are store the % real parts, and the last D/2 dimensions store the imaginary parts. The % function returns the corresponding complex numbers % function [output] = F_realImag2complex(input_layer) input = input_layer.a; [D, T, N] = size(input); j = sqrt(-1); ...
github
singaxiong/SignalGraph-master
phoneID2posterior.m
.m
SignalGraph-master/graph/phoneID2posterior.m
666
utf_8
4a3cb76d5c887a260f21f381606b9236
function posterior = phoneID2posterior(phoneID, nClass, classID) if nargin<3 classID = 1:nClass; end nSample = length(phoneID); posterior = single(zeros(nClass, nSample)); if 1 c = unique(phoneID); for i = 1:length(c) pos = find(classID==c(i)); if length(pos)==0 i end ...
github
singaxiong/SignalGraph-master
F_word2vec.m
.m
SignalGraph-master/graph/F_word2vec.m
1,068
utf_8
8380e169188ed9cbad7ca1c0c07d379e
% input is of dimension dim x nFr, where dim is |V| * context size and |V| % is the vocabulary size. function output = F_word2vec(input, W, singlePrecision) [dim, nFr, nSeg] = size(input); if nSeg>1 input = reshape(input, dim, nFr*nSeg); end [m,n] = size(W); context = dim/n; nVec = nFr*nSeg; if 1 curr_inpu...
github
singaxiong/SignalGraph-master
phoneID2classID.m
.m
SignalGraph-master/graph/phoneID2classID.m
220
utf_8
76ca89aa9a3d2e7d01c3f72607dfceb1
function classID = phoneID2classID(phoneID, vocab) classID = single(zeros(size(phoneID))); c = unique(phoneID); for i = 1:length(c) pos = find(vocab==c(i)); idx = phoneID==c(i); classID(idx) = pos; end end
github
singaxiong/SignalGraph-master
SumContext.m
.m
SignalGraph-master/graph/SumContext.m
922
utf_8
e2b8462f7f362ce0e9f6b16f6fe763b9
% Concatenate the neighbouring frames vectors to higher dimensional vectors % Input is Number of dimension x number of frames function [y] = SumContext(x, context, window_type) [dim nFr] = size(x); if nargin<3 window_type = 'null'; end if context>1 half_context = (context-1)/2; if 0 % this implementati...
github
singaxiong/SignalGraph-master
F_exp.m
.m
SignalGraph-master/graph/F_exp.m
88
utf_8
83683ed80f04c6e38909ddf5f389d470
function [output] = F_exp(input_layer) input = input_layer.a; output = exp(input); end
github
singaxiong/SignalGraph-master
F_real_imag2BFweight.m
.m
SignalGraph-master/graph/F_real_imag2BFweight.m
1,307
utf_8
fba7243acb561d96c37c53c3d2072438
function [output,validFrameMask] = F_real_imag2BFweight(input_layer, freq_bin, online) % assume input is an array of time delay of C microphone channels. % freq_bin is an array of center frequencies of N FFT bins. input = input_layer.a; [D, T, nSent] = size(input); N = length(freq_bin); nCh = D/N/2; j = sqrt(-1); ...
github
singaxiong/SignalGraph-master
B_MVDR_spatialCov.m
.m
SignalGraph-master/graph/B_MVDR_spatialCov.m
4,790
utf_8
6ffae6f4cef18a3c0f4b8ef5341ea252
function grad = B_MVDR_spatialCov(X, curr_layer, beamform_layer, after_power_layer) % X is the multichannel complex spectrum inputs [D,C,T,N] = size(X); % weight is the beamforming weight weight = reshape(curr_layer.a, D,C,N); lambda = curr_layer.lambda; phi_s = curr_layer.phi_s; phi_n = curr_layer.phi_n; if isfield(...
github
singaxiong/SignalGraph-master
B_jointCost.m
.m
SignalGraph-master/graph/B_jointCost.m
998
utf_8
69d012c2efc9427e388fcefe59945c04
% compute the joint cost between neighboring data points % function grad = B_jointCost(input_layer, curr_layer) input = input_layer.a; [D,T,N] = size(input); dimension = curr_layer.dimension; % dimension defines along which dimension the smoothness are defined. precision = class(gather(input(1))); if IsInGPU(input...
github
singaxiong/SignalGraph-master
F_ll_gmm.m
.m
SignalGraph-master/graph/F_ll_gmm.m
2,902
utf_8
f5838b67ad31a2b93f829c1dd0f4aedb
% take the covariance matrix of input trajectories function curr_layer = F_ll_gmm(input, curr_layer) useGPU = strcmpi(class(input(1)), 'gpuArray'); prior = curr_layer.prior; mu = curr_layer.mu; invCov = curr_layer.invCov; [dim,nFr] = size(input); nGaussian = length(prior); [d1,d2,d3] = size(invCov); if nGaussian==1 ...
github
singaxiong/SignalGraph-master
prepareCostEvaluation.m
.m
SignalGraph-master/graph/prepareCostEvaluation.m
6,111
utf_8
805f4e644938c911967b16bccc780d49
% This function is intended to be called from mean square erorr or cross % entropy cost function layers. It decides which input stream is system % output and which is desired target and scale of frames. % % The function also applies following operations on the output and target: % 1) labelDelay: delay the target f...
github
singaxiong/SignalGraph-master
B_SpatialCovSplitMask.m
.m
SignalGraph-master/graph/B_SpatialCovSplitMask.m
3,924
utf_8
c6e34d6826e87593fdb307ceb72fc1a0
function grad = B_SpatialCovSplitMask(future_layers, prev_layers, curr_layer) maskSpeech = prev_layers{1}.a; maskNoise = prev_layers{2}.a; data = prev_layers{3}.a; [D,T,N] = size(maskSpeech); [D2,T,N] = size(data); nCh = D2/D; spatCov = curr_layer.a; future_grad = GetFutureGrad(future_layers, curr_layer); % data = ab...
github
singaxiong/SignalGraph-master
B_splice_single_sentence.m
.m
SignalGraph-master/graph/B_splice_single_sentence.m
735
utf_8
cffbaa95f01ac2328651334eabd9a305
function grad = B_splice_single_sentence(grad, future_grad, context, mask) [dim,nFr,nSeg] = size(future_grad); dim = dim/context; half_ctx = (context-1)/2; tmp1 = []; tmp2 = []; for i=-half_ctx:half_ctx curr_future_grad = future_grad( (i+half_ctx)*dim+1 : (i+half_ctx+1)*dim, :,:); if i<0 grad(:,1:end+...
github
singaxiong/SignalGraph-master
AddSpMatMat_sparseonly.m
.m
SignalGraph-master/graph/AddSpMatMat_sparseonly.m
1,254
utf_8
629f59b157df7444875142289fdcabf0
% This function add a full matrix with a sparse matrix function out = AddSpMatMat_sparseonly(w1,spMat, w2, Mat) [m,n] = size(spMat); if w1==-1 spMat = -spMat; elseif w1~=1 spMat = spMat * w1; end if 0 if w2==-1 Mat = -Mat; elseif w2~=1 Mat = Mat * w2; end idx = find(spMat); ...
github
singaxiong/SignalGraph-master
PadGradientVariableLength.m
.m
SignalGraph-master/graph/PadGradientVariableLength.m
746
utf_8
636f132edcdcf6076d300da30beab7c4
% this function store gradient in the format of the output function gradOut = PadGradientVariableLength(grad, mask) [D,T] = size(grad); [nFr, nSeg] = size(mask); if nFr == 1 nFrActual = ones(nFr, nSeg); else nFrActual = gather(sum(mask==0)); end if sum(nFrActual)~=T fprintf('Error: the number of valid frame...
github
singaxiong/SignalGraph-master
F_tconv.m
.m
SignalGraph-master/graph/F_tconv.m
1,292
utf_8
465f00257ea10294d995f7222ba84a85
% this function perform convolution on the temporal direction function [output,X2] = F_tconv(input_layers, curr_layer) if length(input_layers)==1 input = input_layers{1}.a; W = curr_layer.W; b = curr_layer.b; elseif length(input_layers)==2 W = input_layers{1}.a; b = zeros(size(W,1),1); input = ...
github
singaxiong/SignalGraph-master
B_copyVec2Mat.m
.m
SignalGraph-master/graph/B_copyVec2Mat.m
549
utf_8
f0cc1affa02b8df82f643ae0d940ad28
% put a vector in selected positions of a matrix. % function [grad] = B_copyVec2Mat(input_layer, curr_layer, future_layers) input = input_layer.a; [D,T,N] = size(input); D1 = curr_layer.targetDims(1); D2 = curr_layer.targetDims(2); index2copy = curr_layer.index2copy; future_grad = GetFutureGrad(future_layers, curr_l...
github
singaxiong/SignalGraph-master
B_transpose.m
.m
SignalGraph-master/graph/B_transpose.m
161
utf_8
a8b586b4379501ef83e0b6fbaecd0372
% repeat a matrix % function [grad] = B_transpose(future_layers, curr_layer) future_grad = GetFutureGrad(future_layers, curr_layer); grad = future_grad'; end
github
singaxiong/SignalGraph-master
ExtractVariableLengthTrajectory.m
.m
SignalGraph-master/graph/ExtractVariableLengthTrajectory.m
579
utf_8
838da619111253a6623884429add9f4e
% check whether the sequences are of the same length. % The sequences are aligned from begining. If some sequence is shorter, it % will be padded with a big negative number, i.e. -1e10, in the first dimension. % function [data2, mask, variableLength] = ExtractVariableLengthTrajectory(data, mask) if nargin<2 mask...
github
singaxiong/SignalGraph-master
F_power_spectrum.m
.m
SignalGraph-master/graph/F_power_spectrum.m
348
utf_8
2c4d414d8e6ccd501e560584133c4fed
function [output] = F_power_spectrum(input_layer) % assume the input is a DxTxN matrix of complex spectrum, where D is the dimension of the feature vector, % T is the number of frames in the minibatch or utterance, and N is the % number of sentences ComplexSpectrum = input_layer.a; output = abs( ComplexSpectrum .* co...
github
singaxiong/SignalGraph-master
GetFutureGrad.m
.m
SignalGraph-master/graph/GetFutureGrad.m
1,914
utf_8
be962d90465c9238fa58ec5ba372b036
% If a future layer has multiple inputs, it's grad will be a cell array, % each cell is for one input. A cell is empty if the corresponding input % does not need gradient to be propagated. % This function is used to handle the various scenarios when the future % layer has multiple inputs. function future_grad = GetFut...
github
singaxiong/SignalGraph-master
F_sparse_affine_transform.m
.m
SignalGraph-master/graph/F_sparse_affine_transform.m
522
utf_8
f9d44815f075f79d1c080990328b6ca1
function output = F_sparse_affine_transform(input, transform, bias, singlePrecision) [D,M,N] = size(input); if N>1 input = reshape(input, D,M*N); end visible_nonzero_idx = find(sum(abs(input),2)>0); visible_nonzero = full(input(visible_nonzero_idx,:)); if singlePrecision==1 visible_nonzero = single(visible_n...
github
singaxiong/SignalGraph-master
F_reshape.m
.m
SignalGraph-master/graph/F_reshape.m
581
utf_8
cd9a32d11bbcf5e3ef29c88412f23839
% reshape the first N dimsions of the data % sourceDims: an array [D1 D2 D3,...] that specifies the dimensions to be % involved in reshaping. % targetDims: an array [M1 M2 M3,...] that specifies the dimensions after % reshaping. prod(sourceDims) need to be eqal to prod(targetDims). % function [output] = F_reshape(input...
github
singaxiong/SignalGraph-master
DetermineGradientPass.m
.m
SignalGraph-master/graph/DetermineGradientPass.m
642
utf_8
c81af40adebf68a1ee95fd8ab3710d66
% For each layer, determine whether we need to pass gradient back function layer = DetermineGradientPass(layer) % find the parent of every layer parent = DetermineLayerParent(layer); skipBP = zeros(length(layer),1); for i=1:length(layer) if isfield(layer{i}, 'skipBP') skipBP(i) = layer{i}.skipBP; end ...
github
singaxiong/SignalGraph-master
F_absmax_norm.m
.m
SignalGraph-master/graph/F_absmax_norm.m
870
utf_8
22bbbbdf72396aa93e99b2057f7d70f2
function [output,validFrameMask] = F_absmax_norm(input_layer, curr_layer) input = input_layer.a; [D,T,N] = size(input); if isfield(curr_layer, 'minmax') absmax = curr_layer.max; else absmax = 1; end if N==1 output = AbsMaxNorm(input,absmax); validFrameMask = []; else [validFrameMask, variableLeng...
github
singaxiong/SignalGraph-master
B_matrix_multiply.m
.m
SignalGraph-master/graph/B_matrix_multiply.m
1,703
utf_8
cef35ce68c6a73a13676fa240d63a990
% compute matrix multiply X * Y % function [grad] = B_matrix_multiply(input_layers, curr_layer, future_layers) X = input_layers{1}.a; Y = input_layers{2}.a; % input1 and input2 [Dx,Tx,Nx] = size(X); [Dy,Ty,Ny] = size(Y); if Tx ~=Dy fprintf('Error: matrix size not match each other\n'); end future_grad = GetFuture...
github
singaxiong/SignalGraph-master
B_splice_single_sentence2.m
.m
SignalGraph-master/graph/B_splice_single_sentence2.m
1,659
utf_8
4aad68f8c3c6d7d5006f6dc064768e92
function grad = B_splice_single_sentence2(future_grad, context, mask) [dim,nFr,nSeg] = size(future_grad); dim = dim/context; half_ctx = (context-1)/2; future_grad2 = reshape(future_grad, dim, context, nFr, nSeg); future_grad2(:,:,end+1,:) = 0; for i=-half_ctx:half_ctx all_idx{i+half_ctx+1} = min(nFr,max(1, (1:nF...
github
singaxiong/SignalGraph-master
F_logdet.m
.m
SignalGraph-master/graph/F_logdet.m
390
utf_8
7abd5e79f3ee6b32785e8917738ab346
% take the covariance matrix of input trajectories function output = F_logdet(input) precision = class(gather(input(1))); if ~strcmpi(precision, 'double') % we need to use double precision input = double(input); end [D,M,N] = size(input); if N==1 output = log(det(input)); else % to be implemented end ...
github
singaxiong/SignalGraph-master
F_SpatialCovSplitMask.m
.m
SignalGraph-master/graph/F_SpatialCovSplitMask.m
1,857
utf_8
7cdd4e50196d7253d9bdb1ef796de30d
% Estimate spatial covariance matrix for sentences using a mask. The mask % specifies speech presense probability at all time frequency locations, % with a 1 means speech present and 0 means speech absent. % function output = F_SpatialCovSplitMask(prev_layers, curr_layer) maskSpeech = prev_layers{1}.a; maskNoise = pre...
github
singaxiong/SignalGraph-master
ExpandContext.m
.m
SignalGraph-master/graph/ExpandContext.m
1,280
utf_8
9de3e7b1510c9612b45818dce0ffbfbb
% Concatenate the neighbouring frames vectors to higher dimensional vectors % Input is Number of dimension x number of frames function [y] = ExpandContext(x, context, window_type) [dim nFr] = size(x); if nargin<3 window_type = 'null'; end if context>1 half_context = (context-1)/2; idx = [ones(1,half_conte...
github
singaxiong/SignalGraph-master
B_relu.m
.m
SignalGraph-master/graph/B_relu.m
420
utf_8
a64ffccba7462e217ed5642c22ad4e06
function grad = B_relu(future_layers, curr_layer) output = curr_layer.a; if isfield(curr_layer, 'threshold') threshold = curr_layer.threshold; else threshold = 0; end mask = output>threshold; if strcmpi(class(output), 'gpuArray') grad = gpuArray.zeros(size(output)); else grad = zeros(size(output)); en...
github
singaxiong/SignalGraph-master
F_beamforming.m
.m
SignalGraph-master/graph/F_beamforming.m
320
utf_8
9f31df275d1477fd10b6abe2386dbfbd
function [output] = F_beamforming(input_layers, curr_layer) weight = input_layers{1}.a; nBin = length(curr_layer.freqBin); input = input_layers{2}.a; [Di,Ti,Ni] = size(input); output = bsxfun(@times, input, conj(weight)); output = reshape(output, nBin,Di/nBin,Ti,Ni); output = squeeze(sum(output,2)); end
github
singaxiong/SignalGraph-master
B_LSTM_back_Aug24_2017_9am.m
.m
SignalGraph-master/graph/B_LSTM_back_Aug24_2017_9am.m
8,283
utf_8
cf82934a9d4560dd128bb7be57dac381
% Implement the back propagation of an LSTM layer. % Author: Xiong Xiao, Temasek Labs, NTU, Singapore. % Last modified: 13 Oct 2015 % function [grad, grad_W, grad_b] = B_LSTM(input_layer, LSTM_layer, future_layers) input = input_layer.a; W = LSTM_layer.W; if strcmpi(class(input), 'gpuArray'); useGPU=1; else useGPU = 0;...
github
singaxiong/SignalGraph-master
B_MVDR_spatialCov_old.m
.m
SignalGraph-master/graph/B_MVDR_spatialCov_old.m
2,444
utf_8
6b5e7c679dc856abc812f83aed9bc66f
function grad = B_MVDR_spatialCov(X, curr_layer, beamform_layer, after_power_layer) % X is the multichannel complex spectrum inputs [D,C,T,N] = size(X); % weight is the beamforming weight weight = reshape(curr_layer.a, D,C,N); lambda = curr_layer.lambda; phi_s = curr_layer.phi_s; phi_n = curr_layer.phi_n; if isfield(...
github
singaxiong/SignalGraph-master
B_LSTM_back.m
.m
SignalGraph-master/graph/B_LSTM_back.m
5,879
utf_8
756db0d0783bbd903adead13b7448a3e
% Implement the back propagation of an LSTM layer. % Author: Xiong Xiao, Temasek Labs, NTU, Singapore. % Last modified: 13 Oct 2015 % function [grad, grad_W, grad_b] = B_LSTM(input, LSTM_layer, future_layers) W = LSTM_layer.W; if strcmpi(class(input), 'gpuArray'); useGPU=1; else useGPU = 0; end precision = class(gather...