annotate dsp/onsets/DetectionFunction.cpp @ 280:9c403afdd9e9

* Various fixes related to the bar estimator code
author Chris Cannam <c.cannam@qmul.ac.uk>
date Tue, 10 Feb 2009 16:37:11 +0000
parents c8908cdc8c32
children 6cb2b3cd5356
rev   line source
c@225 1 /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*- vi:set ts=8 sts=4 sw=4: */
c@225 2
c@225 3 /*
c@225 4 QM DSP Library
c@225 5
c@225 6 Centre for Digital Music, Queen Mary, University of London.
c@225 7 This file copyright 2005-2006 Christian Landone.
c@225 8 All rights reserved.
c@225 9 */
c@225 10
c@225 11 #include "DetectionFunction.h"
c@272 12 #include <cstring>
c@225 13
c@225 14 //////////////////////////////////////////////////////////////////////
c@225 15 // Construction/Destruction
c@225 16 //////////////////////////////////////////////////////////////////////
c@225 17
c@225 18 DetectionFunction::DetectionFunction( DFConfig Config ) :
c@225 19 m_window(0)
c@225 20 {
c@227 21 m_magHistory = NULL;
c@227 22 m_phaseHistory = NULL;
c@227 23 m_phaseHistoryOld = NULL;
c@239 24 m_magPeaks = NULL;
c@225 25
c@225 26 initialise( Config );
c@225 27 }
c@225 28
c@225 29 DetectionFunction::~DetectionFunction()
c@225 30 {
c@225 31 deInitialise();
c@225 32 }
c@225 33
c@225 34
c@225 35 void DetectionFunction::initialise( DFConfig Config )
c@225 36 {
c@225 37 m_dataLength = Config.frameLength;
c@225 38 m_halfLength = m_dataLength/2;
c@239 39
c@225 40 m_DFType = Config.DFType;
c@238 41 m_stepSize = Config.stepSize;
c@225 42
c@239 43 m_whiten = Config.adaptiveWhitening;
c@239 44 m_whitenRelaxCoeff = Config.whiteningRelaxCoeff;
c@239 45 m_whitenFloor = Config.whiteningFloor;
c@239 46 if (m_whitenRelaxCoeff < 0) m_whitenRelaxCoeff = 0.9997;
c@239 47 if (m_whitenFloor < 0) m_whitenFloor = 0.01;
c@239 48
c@227 49 m_magHistory = new double[ m_halfLength ];
c@227 50 memset(m_magHistory,0, m_halfLength*sizeof(double));
c@225 51
c@227 52 m_phaseHistory = new double[ m_halfLength ];
c@227 53 memset(m_phaseHistory,0, m_halfLength*sizeof(double));
c@225 54
c@227 55 m_phaseHistoryOld = new double[ m_halfLength ];
c@227 56 memset(m_phaseHistoryOld,0, m_halfLength*sizeof(double));
c@225 57
c@239 58 m_magPeaks = new double[ m_halfLength ];
c@239 59 memset(m_magPeaks,0, m_halfLength*sizeof(double));
c@239 60
c@225 61 m_phaseVoc = new PhaseVocoder;
c@225 62
c@225 63 m_DFWindowedFrame = new double[ m_dataLength ];
c@225 64 m_magnitude = new double[ m_halfLength ];
c@225 65 m_thetaAngle = new double[ m_halfLength ];
c@225 66
c@225 67 m_window = new Window<double>(HanningWindow, m_dataLength);
c@225 68 }
c@225 69
c@225 70 void DetectionFunction::deInitialise()
c@225 71 {
c@227 72 delete [] m_magHistory ;
c@227 73 delete [] m_phaseHistory ;
c@227 74 delete [] m_phaseHistoryOld ;
c@239 75 delete [] m_magPeaks ;
c@225 76
c@225 77 delete m_phaseVoc;
c@225 78
c@225 79 delete [] m_DFWindowedFrame;
c@225 80 delete [] m_magnitude;
c@225 81 delete [] m_thetaAngle;
c@225 82
c@225 83 delete m_window;
c@225 84 }
c@225 85
c@280 86 double DetectionFunction::process( const double *TDomain )
c@225 87 {
c@225 88 m_window->cut( TDomain, m_DFWindowedFrame );
c@280 89
c@280 90 // Our own FFT implementation supports power-of-two sizes only.
c@280 91 // If we have to use this implementation (as opposed to the
c@280 92 // version of process() below that operates on frequency domain
c@280 93 // data directly), we will have to use the next smallest power of
c@280 94 // two from the block size. Results may vary accordingly!
c@280 95
c@280 96 int actualLength = MathUtilities::previousPowerOfTwo(m_dataLength);
c@280 97
c@280 98 if (actualLength != m_dataLength) {
c@280 99 // Pre-fill mag and phase vectors with zero, as the FFT output
c@280 100 // will not fill the arrays
c@280 101 for (int i = actualLength/2; i < m_dataLength/2; ++i) {
c@280 102 m_magnitude[i] = 0;
c@280 103 m_thetaAngle[0] = 0;
c@280 104 }
c@280 105 }
c@280 106
c@280 107 m_phaseVoc->process(actualLength, m_DFWindowedFrame, m_magnitude, m_thetaAngle);
c@225 108
c@239 109 if (m_whiten) whiten();
c@239 110
c@227 111 return runDF();
c@227 112 }
c@227 113
c@280 114 double DetectionFunction::process( const double *magnitudes, const double *phases )
c@227 115 {
c@227 116 for (size_t i = 0; i < m_halfLength; ++i) {
c@227 117 m_magnitude[i] = magnitudes[i];
c@227 118 m_thetaAngle[i] = phases[i];
c@227 119 }
c@227 120
c@239 121 if (m_whiten) whiten();
c@239 122
c@227 123 return runDF();
c@227 124 }
c@227 125
c@239 126 void DetectionFunction::whiten()
c@239 127 {
c@239 128 for (unsigned int i = 0; i < m_halfLength; ++i) {
c@239 129 double m = m_magnitude[i];
c@239 130 if (m < m_magPeaks[i]) {
c@239 131 m = m + (m_magPeaks[i] - m) * m_whitenRelaxCoeff;
c@239 132 }
c@239 133 if (m < m_whitenFloor) m = m_whitenFloor;
c@239 134 m_magPeaks[i] = m;
c@239 135 m_magnitude[i] /= m;
c@239 136 }
c@239 137 }
c@239 138
c@227 139 double DetectionFunction::runDF()
c@227 140 {
c@227 141 double retVal = 0;
c@227 142
c@225 143 switch( m_DFType )
c@225 144 {
c@225 145 case DF_HFC:
c@225 146 retVal = HFC( m_halfLength, m_magnitude);
c@225 147 break;
c@225 148
c@238 149 case DF_SPECDIFF:
c@225 150 retVal = specDiff( m_halfLength, m_magnitude);
c@225 151 break;
c@225 152
c@225 153 case DF_PHASEDEV:
c@239 154 retVal = phaseDev( m_halfLength, m_thetaAngle);
c@225 155 break;
c@225 156
c@225 157 case DF_COMPLEXSD:
c@225 158 retVal = complexSD( m_halfLength, m_magnitude, m_thetaAngle);
c@225 159 break;
c@237 160
c@237 161 case DF_BROADBAND:
c@239 162 retVal = broadband( m_halfLength, m_magnitude);
c@239 163 break;
c@225 164 }
c@225 165
c@225 166 return retVal;
c@225 167 }
c@225 168
c@225 169 double DetectionFunction::HFC(unsigned int length, double *src)
c@225 170 {
c@225 171 unsigned int i;
c@225 172 double val = 0;
c@225 173
c@225 174 for( i = 0; i < length; i++)
c@225 175 {
c@225 176 val += src[ i ] * ( i + 1);
c@225 177 }
c@225 178 return val;
c@225 179 }
c@225 180
c@225 181 double DetectionFunction::specDiff(unsigned int length, double *src)
c@225 182 {
c@225 183 unsigned int i;
c@225 184 double val = 0.0;
c@225 185 double temp = 0.0;
c@225 186 double diff = 0.0;
c@225 187
c@225 188 for( i = 0; i < length; i++)
c@225 189 {
c@227 190 temp = fabs( (src[ i ] * src[ i ]) - (m_magHistory[ i ] * m_magHistory[ i ]) );
c@225 191
c@225 192 diff= sqrt(temp);
c@225 193
c@238 194 // (See note in phaseDev below.)
c@238 195
c@238 196 val += diff;
c@225 197
c@227 198 m_magHistory[ i ] = src[ i ];
c@225 199 }
c@225 200
c@225 201 return val;
c@225 202 }
c@225 203
c@225 204
c@239 205 double DetectionFunction::phaseDev(unsigned int length, double *srcPhase)
c@225 206 {
c@225 207 unsigned int i;
c@225 208 double tmpPhase = 0;
c@225 209 double tmpVal = 0;
c@225 210 double val = 0;
c@225 211
c@225 212 double dev = 0;
c@225 213
c@225 214 for( i = 0; i < length; i++)
c@225 215 {
c@227 216 tmpPhase = (srcPhase[ i ]- 2*m_phaseHistory[ i ]+m_phaseHistoryOld[ i ]);
c@225 217 dev = MathUtilities::princarg( tmpPhase );
c@238 218
c@238 219 // A previous version of this code only counted the value here
c@238 220 // if the magnitude exceeded 0.1. My impression is that
c@238 221 // doesn't greatly improve the results for "loud" music (so
c@238 222 // long as the peak picker is reasonably sophisticated), but
c@238 223 // does significantly damage its ability to work with quieter
c@238 224 // music, so I'm removing it and counting the result always.
c@238 225 // Same goes for the spectral difference measure above.
c@225 226
c@238 227 tmpVal = fabs(dev);
c@238 228 val += tmpVal ;
c@225 229
c@227 230 m_phaseHistoryOld[ i ] = m_phaseHistory[ i ] ;
c@227 231 m_phaseHistory[ i ] = srcPhase[ i ];
c@225 232 }
c@225 233
c@225 234
c@225 235 return val;
c@225 236 }
c@225 237
c@225 238
c@225 239 double DetectionFunction::complexSD(unsigned int length, double *srcMagnitude, double *srcPhase)
c@225 240 {
c@225 241 unsigned int i;
c@225 242 double val = 0;
c@225 243 double tmpPhase = 0;
c@225 244 double tmpReal = 0;
c@225 245 double tmpImag = 0;
c@225 246
c@225 247 double dev = 0;
c@225 248 ComplexData meas = ComplexData( 0, 0 );
c@227 249 ComplexData j = ComplexData( 0, 1 );
c@225 250
c@225 251 for( i = 0; i < length; i++)
c@225 252 {
c@227 253 tmpPhase = (srcPhase[ i ]- 2*m_phaseHistory[ i ]+m_phaseHistoryOld[ i ]);
c@225 254 dev= MathUtilities::princarg( tmpPhase );
c@225 255
c@227 256 meas = m_magHistory[i] - ( srcMagnitude[ i ] * exp( j * dev) );
c@225 257
c@225 258 tmpReal = real( meas );
c@225 259 tmpImag = imag( meas );
c@225 260
c@225 261 val += sqrt( (tmpReal * tmpReal) + (tmpImag * tmpImag) );
c@225 262
c@227 263 m_phaseHistoryOld[ i ] = m_phaseHistory[ i ] ;
c@227 264 m_phaseHistory[ i ] = srcPhase[ i ];
c@227 265 m_magHistory[ i ] = srcMagnitude[ i ];
c@225 266 }
c@225 267
c@225 268 return val;
c@225 269 }
c@225 270
c@239 271 double DetectionFunction::broadband(unsigned int length, double *src)
c@237 272 {
c@237 273 double val = 0;
c@237 274 for (unsigned int i = 0; i < length; ++i) {
c@239 275 double sqrmag = src[i] * src[i];
c@237 276 if (m_magHistory[i] > 0.0) {
c@237 277 double diff = 10.0 * log10(sqrmag / m_magHistory[i]);
c@237 278 if (diff > m_dbRise) val = val + 1;
c@237 279 }
c@237 280 m_magHistory[i] = sqrmag;
c@237 281 }
c@237 282 return val;
c@237 283 }
c@237 284
c@225 285 double* DetectionFunction::getSpectrumMagnitude()
c@225 286 {
c@225 287 return m_magnitude;
c@225 288 }
c@225 289