annotate util/SMALL_swipe.m @ 104:e2ce05e21a55

Merge
author vemiya <valentin.emiya@inria.fr>
date Tue, 12 Apr 2011 16:11:41 +0200
parents fc395272d53e
children 8e660fd14774
rev   line source
idamnjanovic@8 1 function [p,s] = SMALL_swipe(X,fs, f, plim,dlog2p,dERBs,woverlap,sTHR)
idamnjanovic@8 2 %
idamnjanovic@8 3 %
idamnjanovic@24 4 % Centre for Digital Music, Queen Mary, University of London.
idamnjanovic@24 5 % This file copyright 2009 Ivan Damnjanovic.
idamnjanovic@8 6 %
idamnjanovic@24 7 % This program is free software; you can redistribute it and/or
idamnjanovic@24 8 % modify it under the terms of the GNU General Public License as
idamnjanovic@24 9 % published by the Free Software Foundation; either version 2 of the
idamnjanovic@24 10 % License, or (at your option) any later version. See the file
idamnjanovic@24 11 % COPYING included with this distribution for more information.
idamnjanovic@24 12 %
idamnjanovic@24 13 %
idamnjanovic@24 14 % This is modified swipep MATLAB code that is working directly in spectral
idamnjanovic@24 15 % domain and uses only one window size. The results are suboptimal
idamnjanovic@24 16 % comparing to original code. It is also converted to SWIPE which uses all
idamnjanovic@24 17 % the harmonics of the signal.
idamnjanovic@24 18 %
idamnjanovic@24 19 % SWIPEP Pitch estimation using SWIPE'.
idamnjanovic@8 20 % P = SWIPEP(X,Fs,[PMIN PMAX],DT,DLOG2P,DERBS,STHR) estimates the pitch
idamnjanovic@8 21 % of the vector signal X every DT seconds. The sampling frequency of
idamnjanovic@8 22 % the signal is Fs (in Hertz). The spectrum is computed using a Hann
idamnjanovic@8 23 % window with an overlap WOVERLAP between 0 and 1. The spectrum is
idamnjanovic@8 24 % sampled uniformly in the ERB scale with a step size of DERBS ERBs. The
idamnjanovic@8 25 % pitch is searched within the range [PMIN PMAX] (in Hertz) with samples
idamnjanovic@8 26 % distributed every DLOG2P units on a base-2 logarithmic scale of Hertz.
idamnjanovic@8 27 % The pitch is fine-tuned using parabolic interpolation with a resolution
idamnjanovic@8 28 % of 1 cent. Pitch estimates with a strength lower than STHR are treated
idamnjanovic@8 29 % as undefined.
idamnjanovic@8 30 %
idamnjanovic@8 31 % [P,T,S] = SWIPEP(X,Fs,[PMIN PMAX],DT,DLOG2P,DERBS,WOVERLAP,STHR)
idamnjanovic@8 32 % returns the times T at which the pitch was estimated and the pitch
idamnjanovic@8 33 % strength S of every pitch estimate.
idamnjanovic@8 34 %
idamnjanovic@8 35 % P = SWIPEP(X,Fs) estimates the pitch using the default settings PMIN =
idamnjanovic@8 36 % 30 Hz, PMAX = 5000 Hz, DT = 0.001 s, DLOG2P = 1/48 (48 steps per
idamnjanovic@8 37 % octave), DERBS = 0.1 ERBs, WOVERLAP = 0.5, and STHR = -Inf.
idamnjanovic@8 38 %
idamnjanovic@8 39 % P = SWIPEP(X,Fs,...,[],...) uses the default setting for the parameter
idamnjanovic@8 40 % replaced with the placeholder [].
idamnjanovic@8 41 %
idamnjanovic@8 42 % REMARKS: (1) For better results, make DLOG2P and DERBS as small as
idamnjanovic@8 43 % possible and WOVERLAP as large as possible. However, take into account
idamnjanovic@8 44 % that the computational complexity of the algorithm is inversely
idamnjanovic@8 45 % proportional to DLOG2P, DERBS and 1-WOVERLAP, and that the default
idamnjanovic@8 46 % values have been found empirically to produce good results. Consider
idamnjanovic@8 47 % also that the computational complexity is directly proportional to the
idamnjanovic@8 48 % number of octaves in the pitch search range, and therefore , it is
idamnjanovic@8 49 % recommendable to restrict the search range to the expected range of
idamnjanovic@8 50 % pitch, if any. (2) This code implements SWIPE', which uses only the
idamnjanovic@8 51 % first and prime harmonics of the signal. To convert it into SWIPE,
idamnjanovic@8 52 % which uses all the harmonics of the signal, replace the word
idamnjanovic@8 53 % PRIMES with a colon (it is located almost at the end of the code).
idamnjanovic@8 54 % However, this may not be recommendable since SWIPE' is reported to
idamnjanovic@8 55 % produce on average better results than SWIPE (Camacho and Harris,
idamnjanovic@8 56 % 2008).
idamnjanovic@8 57 %
idamnjanovic@8 58 % EXAMPLE: Estimate the pitch of the signal X every 10 ms within the
idamnjanovic@8 59 % range 75-500 Hz using the default resolution (i.e., 48 steps per
idamnjanovic@8 60 % octave), sampling the spectrum every 1/20th of ERB, using a window
idamnjanovic@8 61 % overlap factor of 50%, and discarding samples with pitch strength
idamnjanovic@8 62 % lower than 0.2. Plot the pitch trace.
idamnjanovic@8 63 % [x,Fs] = wavread(filename);
idamnjanovic@8 64 % [p,t,s] = swipep(x,Fs,[75 500],0.01,[],1/20,0.5,0.2);
idamnjanovic@8 65 % plot(1000*t,p)
idamnjanovic@8 66 % xlabel('Time (ms)')
idamnjanovic@8 67 % ylabel('Pitch (Hz)')
idamnjanovic@8 68 %
idamnjanovic@8 69 % REFERENCES: Camacho, A., Harris, J.G, (2008) "A sawtooth waveform
idamnjanovic@8 70 % inspired pitch estimator for speech and music," J. Acoust. Soc. Am.
idamnjanovic@8 71 % 124, 1638-1652.
idamnjanovic@24 72
idamnjanovic@8 73 if ~ exist( 'plim', 'var' ) || isempty(plim), plim = [30 5000]; end
idamnjanovic@8 74 %if ~ exist( 'dt', 'var' ) || isempty(dt), dt = 0.001; end
idamnjanovic@8 75 if ~ exist( 'dlog2p', 'var' ) || isempty(dlog2p), dlog2p = 1/48; end
idamnjanovic@8 76 if ~ exist( 'dERBs', 'var' ) || isempty(dERBs), dERBs = 0.05; end
idamnjanovic@8 77 % if ~ exist( 'woverlap', 'var' ) || isempty(woverlap)
idamnjanovic@8 78 % woverlap = 0.5;
idamnjanovic@8 79 % elseif woverlap>1 || woverlap<0
idamnjanovic@8 80 % error('Window overlap must be between 0 and 1.')
idamnjanovic@8 81 % end
idamnjanovic@8 82 if ~ exist( 'sTHR', 'var' ) || isempty(sTHR), sTHR = -Inf; end
idamnjanovic@8 83 %t = [ 0: dt: length(x)/fs ]'; % Times
idamnjanovic@8 84 % Define pitch candidates
idamnjanovic@8 85 log2pc = [ log2(plim(1)): dlog2p: log2(plim(2)) ]';
idamnjanovic@8 86 pc = 2 .^ log2pc;
idamnjanovic@8 87 S = zeros( length(pc), 1 ); % Pitch strength matrix
idamnjanovic@8 88 % Determine P2-WSs
idamnjanovic@8 89 %logWs = round( log2( 8*fs ./ plim ) );
idamnjanovic@8 90 ws = [2822];%2.^[ logWs(1): -1: logWs(2) ]; % P2-WSs
idamnjanovic@8 91 pO = 8 * fs ./ ws; % Optimal pitches for P2-WSs
idamnjanovic@8 92 % Determine window sizes used by each pitch candidate
idamnjanovic@8 93 d = 1 + log2pc - log2( 8*fs./ws(1) );
idamnjanovic@8 94 % Create ERB-scale uniformly-spaced frequencies (in Hertz)
idamnjanovic@8 95 fERBs = erbs2hz([ hz2erbs(min(pc)/4): dERBs: hz2erbs(fs/2) ]');
idamnjanovic@8 96 for i = 1 : length(ws)
idamnjanovic@8 97 %dn = max( 1, round( 8*(1-woverlap) * fs / pO(i) ) ); % Hop size
idamnjanovic@8 98 % Zero pad signal
idamnjanovic@8 99 %xzp = [ zeros( ws(i)/2, 1 ); x(:); zeros( dn + ws(i)/2, 1 ) ];
idamnjanovic@8 100 % Compute spectrum
idamnjanovic@8 101 %w = hanning( ws(i) ); % Hann window
idamnjanovic@8 102 %o = max( 0, round( ws(i) - dn ) ); % Window overlap
idamnjanovic@8 103 %[ X, f, ti ] = specgram( xzp, ws(i), fs, w, o );
idamnjanovic@8 104 % Select candidates that use this window size
idamnjanovic@8 105 if length(ws) == 1
idamnjanovic@8 106 j=[1:size(pc)]'; k = [];
idamnjanovic@8 107 elseif i == length(ws)
idamnjanovic@8 108 j=find(d-i>-1); k=find(d(j)-i<0);
idamnjanovic@8 109 elseif i==1
idamnjanovic@8 110 j=find(d-i<1); k=find(d(j)-i>0);
idamnjanovic@8 111 else
idamnjanovic@8 112 j=find(abs(d-i)<1); k=1:length(j);
idamnjanovic@8 113 end
idamnjanovic@8 114 % Compute loudness at ERBs uniformly-spaced frequencies
idamnjanovic@8 115 fERBs = fERBs( find( fERBs > pc(1)/4, 1, 'first' ) : end );
idamnjanovic@8 116 L = sqrt( max( 0, interp1( f, X, fERBs, 'spline', 0) ) );
idamnjanovic@8 117 % Compute pitch strength
idamnjanovic@8 118 Si = pitchStrengthAllCandidates( fERBs, L, pc );
idamnjanovic@8 119 % Interpolate pitch strength at desired times
idamnjanovic@8 120 % if size(Si,2) > 1
idamnjanovic@8 121 % warning off MATLAB:interp1:NaNinY
idamnjanovic@8 122 % Si = interp1( ti, Si', t, 'linear', NaN )';
idamnjanovic@8 123 % warning on MATLAB:interp1:NaNinY
idamnjanovic@8 124 % else
idamnjanovic@8 125 % Si = repmat( NaN, length(Si),1 );
idamnjanovic@8 126 % end
idamnjanovic@8 127 % Add pitch strength to combination
idamnjanovic@8 128 % lambda = d( j(k) ) - i;
idamnjanovic@8 129 mu = ones( size(j) );
idamnjanovic@8 130 % mu(k) = 1 - abs( lambda );
idamnjanovic@8 131 S(j,:) = S(j,:) + repmat(mu,1,size(Si,2)) .* Si;
idamnjanovic@8 132 end
idamnjanovic@8 133 % Fine tune pitch using parabolic interpolation
idamnjanovic@8 134 p = repmat( NaN, size(S,2), 1 );
idamnjanovic@8 135 s = repmat( NaN, size(S,2), 1 );
idamnjanovic@8 136 for j = 1 : size(S,2)
idamnjanovic@8 137 [ s(j), i ] = max( S(:,j), [], 1 );
idamnjanovic@8 138 if s(j) < sTHR, continue, end
idamnjanovic@8 139 if i == 1 || i == length(pc)
idamnjanovic@8 140 p(j) = pc(i);
idamnjanovic@8 141 else
idamnjanovic@8 142 I = i-1 : i+1;
idamnjanovic@8 143 tc = 1 ./ pc(I);
idamnjanovic@8 144 ntc = ( tc/tc(2) - 1 ) * 2*pi;
idamnjanovic@8 145 c = polyfit( ntc, S(I,j), 2 );
idamnjanovic@8 146 ftc = 1 ./ 2.^[ log2(pc(I(1))): 1/12/100: log2(pc(I(3))) ];
idamnjanovic@8 147 nftc = ( ftc/tc(2) - 1 ) * 2*pi;
idamnjanovic@8 148 [s(j) k] = max( polyval( c, nftc ) );
idamnjanovic@8 149 p(j) = 2 ^ ( log2(pc(I(1))) + (k-1)/12/100 );
idamnjanovic@8 150 % if (p(j)-pc(I(1)))<0.75*abs(p(j)-pc(I(2)))
idamnjanovic@8 151 % p(j)=pc(I(1));
idamnjanovic@8 152 % elseif (pc(I(3))-p(j))<0.75*abs(p(j)-pc(I(2)))
idamnjanovic@8 153 % p(j)=pc(I(3));
idamnjanovic@8 154 % else
idamnjanovic@8 155 p(j)=pc(I(2));
idamnjanovic@8 156 % end
idamnjanovic@8 157 end
idamnjanovic@8 158 end
idamnjanovic@8 159
idamnjanovic@8 160 function S = pitchStrengthAllCandidates( f, L, pc )
idamnjanovic@8 161 % Create pitch strength matrix
idamnjanovic@8 162 S = zeros( length(pc), size(L,2) );
idamnjanovic@8 163 % Define integration regions
idamnjanovic@8 164 k = ones( 1, length(pc)+1 );
idamnjanovic@8 165 for j = 1 : length(k)-1
idamnjanovic@8 166 k(j+1) = k(j) - 1 + find( f(k(j):end) > pc(j)/4, 1, 'first' );
idamnjanovic@8 167 end
idamnjanovic@8 168 k = k(2:end);
idamnjanovic@8 169 % Create loudness normalization matrix
idamnjanovic@8 170 N = sqrt( flipud( cumsum( flipud(L.*L) ) ) );
idamnjanovic@8 171 for j = 1 : length(pc)
idamnjanovic@8 172 % Normalize loudness
idamnjanovic@8 173 warning off MATLAB:divideByZero
idamnjanovic@8 174 NL = L(k(j):end,:) ./ repmat( N(k(j),:), size(L,1)-k(j)+1, 1);
idamnjanovic@8 175 warning on MATLAB:divideByZero
idamnjanovic@8 176 % Compute pitch strength
idamnjanovic@8 177
idamnjanovic@8 178 S(j,:) = pitchStrengthOneCandidate( f(k(j):end), NL, pc(j) );
idamnjanovic@8 179 end
idamnjanovic@8 180
idamnjanovic@8 181 function S = pitchStrengthOneCandidate( f, NL, pc )
idamnjanovic@8 182 n = fix( f(end)/pc - 0.75 ); % Number of harmonics
idamnjanovic@8 183 if n==0, S=NaN; return, end
idamnjanovic@8 184 k = zeros( size(f) ); % Kernel
idamnjanovic@8 185 % Normalize frequency w.r.t. candidate
idamnjanovic@8 186 q = f / pc;
idamnjanovic@8 187 % Create kernel
idamnjanovic@8 188 for i = [ 1:n] % primes(n) ]
idamnjanovic@8 189 a = abs( q - i );
idamnjanovic@8 190 % Peak's weigth
idamnjanovic@8 191 p = a < .25;
idamnjanovic@8 192 k(p) = cos( 2*pi * q(p) );
idamnjanovic@8 193 % Valleys' weights
idamnjanovic@8 194 v = .25 < a & a < .75;
idamnjanovic@8 195 k(v) = k(v) + cos( 2*pi * q(v) ) / 2;
idamnjanovic@8 196 end
idamnjanovic@8 197 % Apply envelope
idamnjanovic@8 198 k = k .* sqrt( 1./f );
idamnjanovic@8 199 % K+-normalize kernel
idamnjanovic@8 200 k = k / norm( k(k>0) );
idamnjanovic@8 201 % Compute pitch strength
idamnjanovic@8 202 S = k' * NL;
idamnjanovic@8 203
idamnjanovic@8 204 function erbs = hz2erbs(hz)
idamnjanovic@8 205 erbs = 21.4 * log10( 1 + hz/229 );
idamnjanovic@8 206
idamnjanovic@8 207 function hz = erbs2hz(erbs)
idamnjanovic@8 208 hz = ( 10 .^ (erbs./21.4) - 1 ) * 229;