File size: 5,434 Bytes
796ff25 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 |
function [path,TDOA]=localize(X,chanlist)
% LOCALIZE Tracks the speaker spatial position over time and computes the
% corresponding TDOA using SRP-PHAT and the Viterbi algorithm
%
% [path,TDOA]=localize(X,chanlist)
%
% Inputs:
% Y: nbin x nfram x nchan STFT of the inpu signal
% chanlist: list of input channels (from 1 to 6)
%
% Output:
% path: 3 x nfram position of the speaker over time in centimeters
% TDOA: nchan x nfram corresponding TDOAs between the speaker position and
% the microphone positions
%
% Note: for computational efficiency, the position on the z-axis is assumed
% to be constant over time.
%
% If you use this software in a publication, please cite:
%
% Jon Barker, Ricard Marxer, Emmanuel Vincent, and Shinji Watanabe, The
% third 'CHiME' Speech Separation and Recognition Challenge: Dataset,
% task and baselines, submitted to IEEE 2015 Automatic Speech Recognition
% and Understanding Workshop (ASRU), 2015.
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Copyright 2015-2016 University of Sheffield (Jon Barker, Ricard Marxer)
% Inria (Emmanuel Vincent)
% Mitsubishi Electric Research Labs (Shinji Watanabe)
% This software is distributed under the terms of the GNU Public License
% version 3 (http://www.gnu.org/licenses/gpl.txt)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
if nargin < 2,
chanlist=[1 3:6];
end
% Define hyper-parameters
pow_thresh=-20; % threshold in dB below which a microphone is considered to fail
center_factor=0.05; % weight given to the prior that the speaker's horizontal position is close to the center
smoothing_factor=3; % weight given to the transition probabilities
% Remove zero frequency
X = X(2:end,:,:);
[nbin,nfram,nchan] = size(X);
wlen=2*nbin;
f=16000/wlen*(1:nbin).';
% Compute relative channel power
if length(chanlist) > 2,
xpow=shiftdim(sum(sum(abs(X).^2,2),1));
xpow=10*log10(xpow/max(xpow));
else
xpow=zeros(1,2);
end
% Define microphone positions in centimeters
xmic=[-10 0 10 -10 0 10]; % left to right axis
ymic=[9.5 9.5 9.5 -9.5 -9.5 -9.5]; % bottom to top axis
zmic=[0 -2 0 0 0 0]; % back to front axis
xmic=xmic(chanlist);
ymic=ymic(chanlist);
zmic=zmic(chanlist);
% Define grid of possible speaker positions in centimeters
xres=46;
xpos=linspace(-45,45,xres);
yres=46;
ypos=linspace(-45,45,yres);
zres=4;
zpos=linspace(15,45,zres);
ngrid=xres*yres*zres;
% Compute horizontal distances between grid points
xvect=reshape(repmat(xpos.',[1 yres]),xres*yres,1);
yvect=reshape(repmat(ypos,[xres 1]),xres*yres,1);
pair_dist=sqrt((repmat(xvect,[1 xres*yres])-repmat(xvect.',[xres*yres 1])).^2+(repmat(yvect,[1 xres*yres])-repmat(yvect.',[xres*yres 1])).^2);
% Compute horizontal distances to the center
center_dist=sqrt((xvect-mean(xpos)).^2+(yvect-mean(ypos)).^2);
% Compute theoretical TDOAs between front pairs
d_grid=zeros(nchan,xres,yres,zres); % speaker-to-microphone distances
for c=1:nchan,
d_grid(c,:,:,:)=sqrt(repmat((xpos.'-xmic(c)).^2,[1 yres zres])+repmat((ypos-ymic(c)).^2,[xres 1 zres])+repmat((permute(zpos,[3 1 2])-zmic(c)).^2,[xres yres 1]));
end
d_grid=reshape(d_grid,nchan,ngrid);
pairs=[];
for c=1:nchan,
pairs=[pairs [c*ones(1,nchan-c); c+1:nchan]]; % microphone pairs
end
npairs=size(pairs,2);
tau_grid=zeros(npairs,ngrid); % TDOAs
for p=1:npairs,
c1=pairs(1,p);
c2=pairs(2,p);
tau_grid(p,:)=(d_grid(c2,:)-d_grid(c1,:))/343/100;
end
% Compute the SRP-PHAT pseudo-spectrum
srp=zeros(nfram,ngrid);
for p=1:npairs, % Loop over front pairs
c1=pairs(1,p);
c2=pairs(2,p);
d=sqrt((xmic(c1)-xmic(c2))^2+(ymic(c1)-ymic(c2))^2+(zmic(c1)-zmic(c2))^2);
alpha=10*343/(d*16000);
lin_grid=linspace(min(tau_grid(p,:)),max(tau_grid(p,:)),100);
lin_spec=zeros(nbin,nfram,100); % GCC-PHAT pseudo-spectrum over a uniform interval
if (xpow(c1)>pow_thresh) && (xpow(c2)>pow_thresh), % discard channels with low power (microphone failure)
P=X(:,:,c1).*conj(X(:,:,c2));
P=P./abs(P);
for ind=1:100,
EXP=repmat(exp(-2*1i*pi*lin_grid(ind)*f),1,nfram);
lin_spec(:,:,ind)=ones(nbin,nfram)-tanh(alpha*real(sqrt(2-2*real(P.*EXP))));
end
end
lin_spec=shiftdim(sum(lin_spec,1));
tau_spec=zeros(nfram,ngrid); % GCC-PHAT pseudo-spectrum over the whole grid
for t=1:nfram,
tau_spec(t,:)=interp1(lin_grid,lin_spec(t,:),tau_grid(p,:));
end
srp=srp+tau_spec; % sum over the microphone pairs
end
% Loop over possible z-axis positions
path=zeros(zres,nfram);
logpost=zeros(zres,1);
xpath=zeros(zres,nfram);
ypath=zeros(zres,nfram);
zpath=zeros(zres,nfram);
srp=reshape(srp,nfram,xres*yres,zres);
for zind=1:zres,
% Weight by distance to the center
weighted_srp=srp(:,:,zind)-center_factor*repmat(center_dist.',[nfram 1]);
% Track the source position over time
[path(zind,:),logpost(zind)]=viterbi(weighted_srp.',zeros(xres*yres,1),zeros(xres*yres,1),-smoothing_factor*pair_dist);
for t=1:nfram,
[xpath(zind,t),ypath(zind,t)]=ind2sub([xres yres],path(zind,t));
zpath(zind,t)=zind;
end
end
% Select the best z-axis position
[~,zind]=max(logpost);
path=(zind-1)*xres*yres+path(zind,:);
xpath=xpos(xpath(zind,:));
ypath=ypos(ypath(zind,:));
zpath=zpos(zpath(zind,:));
% Derive TDOA
d_path=d_grid(:,path);
TDOA=d_path/343/100;
path=[xpath; ypath; zpath];
return
|