function C = en_beatchroma(Segs, Tsegs, Tbeats)
% C = en_beatchroma(Segs,Tsegs,Tbeats) or C = en_beatchrom(ENAstruct)
%   Return beat-synchronous chroma features derived from 
%   the Echo Nest analyze features.  Segs is the 12xN 
%   array of chroma features from EN Analyze, and Tsegs
%   is the vector of segment times.  Tbeats is a vector 
%   of beat times; the segments overlapping each beat 
%   interval are averaged and returned in C.
%   Alternatively, pass in the entire EN Analyze structure as 
%   returned by en_analyze.m to read the data from there.
% 2010-04-08 Dan Ellis dpwe@ee.columbia.edu

nchr = 12;

if isstruct(Segs)
  ENAstruct = Segs;
  % read chroma out of structure; apply per-segment loudness
  Segs = ENAstruct.pitches .* repmat(10.^(ENAstruct.segmentloudness/20),nchr,1);
  % read times too
  Tsegs = ENAstruct.segment;
  Tbeats = ENAstruct.beat;
end

%%%% Old way
%  Tres = 0.01;
%  % We do this the quick and dirty way by upsampling to Tres 
%  % resolution, then averaging down again
%
%  % Set duration as one segment beyond the last beat time
%  duration = 2*Tbeats(end) - Tbeats(end-1);
%
%  % Y = samplemx(D.pitches*diag(idB(D.segmentloudness)./sqrt(sum(D.pitches))),D.segmentstart, [0:0.01:D.duration]);
%  SegsUpsamp = samplemx(Segs, Tsegs, 0:Tres:duration);
%
%  % then downsample down again
%  C = beatavg(SegsUpsamp, Tbeats/Tres);

% Now properly figure time overlaps & weight
C = resample_mx(Segs, Tsegs, Tbeats);
  
% and renormalize columns
n = max(C);
C = C.*repmat(1./n,size(C,1),1);