Mercurial > hg > qm-dsp
changeset 462:bfd7d7633e1d
Merge branch 'keydetection_rounding' into chroma-key-tuning-review
| author | Chris Cannam <cannam@all-day-breakfast.com> |
|---|---|
| date | Tue, 28 May 2019 13:45:08 +0100 |
| parents | 38932adb6c02 (current diff) 9414df58fd0e (diff) |
| children | d8ad5893db6f |
| files | |
| diffstat | 3 files changed, 151 insertions(+), 116 deletions(-) [+] |
line wrap: on
line diff
--- a/README.md Tue May 28 13:41:29 2019 +0100 +++ b/README.md Tue May 28 13:45:08 2019 +0100 @@ -3,7 +3,8 @@ ============== This is a C++ library of functions for Digital Signal Processing and -Music Informatics purposes developed at Queen Mary, University of +Music Informatics purposes developed in the [Centre for Digital +Music](http://c4dm.eecs.qmul.ac.uk) at Queen Mary, University of London. It is used by [QM Vamp Plugins](http://isophonics.net/QMVampPlugins)
--- a/dsp/keydetection/GetKeyMode.cpp Tue May 28 13:41:29 2019 +0100 +++ b/dsp/keydetection/GetKeyMode.cpp Tue May 28 13:45:08 2019 +0100 @@ -22,18 +22,20 @@ #include <cstring> #include <cstdlib> +static const int kBinsPerOctave = 36; + // Chords profile -static double MajProfile[36] = -{ 0.0384, 0.0629, 0.0258, 0.0121, 0.0146, 0.0106, 0.0364, 0.0610, 0.0267, - 0.0126, 0.0121, 0.0086, 0.0364, 0.0623, 0.0279, 0.0275, 0.0414, 0.0186, - 0.0173, 0.0248, 0.0145, 0.0364, 0.0631, 0.0262, 0.0129, 0.0150, 0.0098, - 0.0312, 0.0521, 0.0235, 0.0129, 0.0142, 0.0095, 0.0289, 0.0478, 0.0239}; +static double MajProfile[kBinsPerOctave] = { + 0.0384, 0.0629, 0.0258, 0.0121, 0.0146, 0.0106, 0.0364, 0.0610, 0.0267, + 0.0126, 0.0121, 0.0086, 0.0364, 0.0623, 0.0279, 0.0275, 0.0414, 0.0186, + 0.0173, 0.0248, 0.0145, 0.0364, 0.0631, 0.0262, 0.0129, 0.0150, 0.0098, + 0.0312, 0.0521, 0.0235, 0.0129, 0.0142, 0.0095, 0.0289, 0.0478, 0.0239}; -static double MinProfile[36] = -{ 0.0375, 0.0682, 0.0299, 0.0119, 0.0138, 0.0093, 0.0296, 0.0543, 0.0257, - 0.0292, 0.0519, 0.0246, 0.0159, 0.0234, 0.0135, 0.0291, 0.0544, 0.0248, - 0.0137, 0.0176, 0.0104, 0.0352, 0.0670, 0.0302, 0.0222, 0.0349, 0.0164, - 0.0174, 0.0297, 0.0166, 0.0222, 0.0401, 0.0202, 0.0175, 0.0270, 0.0146}; +static double MinProfile[kBinsPerOctave] = { + 0.0375, 0.0682, 0.0299, 0.0119, 0.0138, 0.0093, 0.0296, 0.0543, 0.0257, + 0.0292, 0.0519, 0.0246, 0.0159, 0.0234, 0.0135, 0.0291, 0.0544, 0.0248, + 0.0137, 0.0176, 0.0104, 0.0352, 0.0670, 0.0302, 0.0222, 0.0349, 0.0164, + 0.0174, 0.0297, 0.0166, 0.0222, 0.0401, 0.0202, 0.0175, 0.0270, 0.0146}; // @@ -42,7 +44,7 @@ ////////////////////////////////////////////////////////////////////// GetKeyMode::GetKeyMode( int sampleRate, float tuningFrequency, - double hpcpAverage, double medianAverage ) : + double hpcpAverage, double medianAverage ) : m_hpcpAverage( hpcpAverage ), m_medianAverage( medianAverage ), m_ChrPointer(0), @@ -51,7 +53,6 @@ m_MeanHPCP(0), m_MajCorr(0), m_MinCorr(0), - m_Keys(0), m_MedianFilterBuffer(0), m_SortedBuffer(0), m_keyStrengths(0) @@ -60,17 +61,17 @@ // Chromagram configuration parameters m_ChromaConfig.normalise = MathUtilities::NormaliseUnitMax; - m_ChromaConfig.FS = lrint(sampleRate/(double)m_DecimationFactor); - if (m_ChromaConfig.FS < 1) m_ChromaConfig.FS = 1; + m_ChromaConfig.FS = sampleRate/(double)m_DecimationFactor; + if (m_ChromaConfig.FS < 1) { + m_ChromaConfig.FS = 1; + } - // Set C (= MIDI #12) as our base : + // Set C3 (= MIDI #48) as our base: // This implies that key = 1 => Cmaj, key = 12 => Bmaj, key = 13 => Cmin, etc. - m_ChromaConfig.min = Pitch::getFrequencyForPitch - (48, 0, tuningFrequency); - m_ChromaConfig.max = Pitch::getFrequencyForPitch - (96, 0, tuningFrequency); + m_ChromaConfig.min = Pitch::getFrequencyForPitch( 48, 0, tuningFrequency ); + m_ChromaConfig.max = Pitch::getFrequencyForPitch( 96, 0, tuningFrequency ); - m_ChromaConfig.BPO = 36; + m_ChromaConfig.BPO = kBinsPerOctave; m_ChromaConfig.CQThresh = 0.0054; // Chromagram inst. @@ -80,7 +81,6 @@ m_ChromaFrameSize = m_Chroma->getFrameSize(); // override hopsize for this application m_ChromaHopSize = m_ChromaFrameSize; - m_BPO = m_ChromaConfig.BPO; // std::cerr << "chroma frame size = " << m_ChromaFrameSize << ", decimation factor = " << m_DecimationFactor << " therefore block size = " << getBlockSize() << std::endl; @@ -96,30 +96,38 @@ // Spawn objectc/arrays m_DecimatedBuffer = new double[m_ChromaFrameSize]; - m_ChromaBuffer = new double[m_BPO * m_ChromaBuffersize]; - memset( m_ChromaBuffer, 0, sizeof(double) * m_BPO * m_ChromaBuffersize); + m_ChromaBuffer = new double[kBinsPerOctave * m_ChromaBuffersize]; + memset( m_ChromaBuffer, 0, sizeof(double) * kBinsPerOctave * m_ChromaBuffersize); - m_MeanHPCP = new double[m_BPO]; + m_MeanHPCP = new double[kBinsPerOctave]; - m_MajCorr = new double[m_BPO]; - m_MinCorr = new double[m_BPO]; - m_Keys = new double[2*m_BPO]; + m_MajCorr = new double[kBinsPerOctave]; + m_MinCorr = new double[kBinsPerOctave]; + m_MajProfileNorm = new double[kBinsPerOctave]; + m_MinProfileNorm = new double[kBinsPerOctave]; + + double mMaj = MathUtilities::mean( MajProfile, kBinsPerOctave ); + double mMin = MathUtilities::mean( MinProfile, kBinsPerOctave ); + + for( unsigned int i = 0; i < kBinsPerOctave; i++ ) { + m_MajProfileNorm[i] = MajProfile[i] - mMaj; + m_MinProfileNorm[i] = MinProfile[i] - mMin; + } + m_MedianFilterBuffer = new int[ m_MedianWinsize ]; memset( m_MedianFilterBuffer, 0, sizeof(int)*m_MedianWinsize); m_SortedBuffer = new int[ m_MedianWinsize ]; - memset( m_SortedBuffer, 0, sizeof(int)*m_MedianWinsize); + memset( m_SortedBuffer, 0, sizeof(int)*m_MedianWinsize); - m_Decimator = new Decimator - ( m_ChromaFrameSize*m_DecimationFactor, m_DecimationFactor ); + m_Decimator = new Decimator( m_ChromaFrameSize*m_DecimationFactor, m_DecimationFactor ); m_keyStrengths = new double[24]; } GetKeyMode::~GetKeyMode() { - delete m_Chroma; delete m_Decimator; @@ -128,40 +136,40 @@ delete [] m_MeanHPCP; delete [] m_MajCorr; delete [] m_MinCorr; - delete [] m_Keys; + delete [] m_MajProfileNorm; + delete [] m_MinProfileNorm; delete [] m_MedianFilterBuffer; delete [] m_SortedBuffer; - - delete[] m_keyStrengths; + delete [] m_keyStrengths; } -double GetKeyMode::krumCorr(double *pData1, double *pData2, unsigned int length) +double GetKeyMode::krumCorr( const double *pDataNorm, const double *pProfileNorm, + int shiftProfile, unsigned int length) { double retVal= 0.0; double num = 0; double den = 0; - double mX = MathUtilities::mean( pData1, length ); - double mY = MathUtilities::mean( pData2, length ); - double sum1 = 0; double sum2 = 0; for( unsigned int i = 0; i <length; i++ ) { - num += ( pData1[i] - mX ) * ( pData2[i] - mY ); + int k = (i - shiftProfile + length) % length; - sum1 += ( (pData1[i]-mX) * (pData1[i]-mX) ); - sum2 += ( (pData2[i]-mY) * (pData2[i]-mY) ); + num += pDataNorm[i] * pProfileNorm[k]; + + sum1 += ( pDataNorm[i] * pDataNorm[i] ); + sum2 += ( pProfileNorm[k] * pProfileNorm[k] ); } den = sqrt(sum1 * sum2); - - if( den>0 ) + + if( den>0 ) { retVal = num/den; - else + } else { retVal = 0; - + } return retVal; } @@ -169,118 +177,102 @@ int GetKeyMode::process(double *PCMData) { int key; - unsigned int j,k; ////////////////////////////////////////////// m_Decimator->process( PCMData, m_DecimatedBuffer); - m_ChrPointer = m_Chroma->process( m_DecimatedBuffer ); + m_ChrPointer = m_Chroma->process( m_DecimatedBuffer ); - - // Move bins such that the centre of the base note is in the - // middle of its three bins : - // Added 21.11.07 by Chris Sutton based on debugging with Katy - // Noland + comparison with Matlab equivalent. - MathUtilities::circShift( m_ChrPointer, m_BPO, 1); /* std::cout << "raw chroma: "; - for (int ii = 0; ii < m_BPO; ++ii) { - if (ii % (m_BPO/12) == 0) std::cout << "\n"; + for (int ii = 0; ii < kBinsPerOctave; ++ii) { + if (ii % (kBinsPerOctave/12) == 0) std::cout << "\n"; std::cout << m_ChrPointer[ii] << " "; } std::cout << std::endl; */ // populate hpcp values; int cbidx; - for( j = 0; j < m_BPO; j++ ) - { - cbidx = (m_bufferindex * m_BPO) + j; + for( j = 0; j < kBinsPerOctave; j++ ) { + cbidx = (m_bufferindex * kBinsPerOctave) + j; m_ChromaBuffer[ cbidx ] = m_ChrPointer[j]; } //keep track of input buffers; - if( m_bufferindex++ >= m_ChromaBuffersize - 1) + if( m_bufferindex++ >= m_ChromaBuffersize - 1) { m_bufferindex = 0; + } // track filling of chroma matrix - if( m_ChromaBufferFilling++ >= m_ChromaBuffersize) + if( m_ChromaBufferFilling++ >= m_ChromaBuffersize) { m_ChromaBufferFilling = m_ChromaBuffersize; + } - //calculate mean - for( k = 0; k < m_BPO; k++ ) - { + //calculate mean + for( k = 0; k < kBinsPerOctave; k++ ) { double mnVal = 0.0; - for( j = 0; j < m_ChromaBufferFilling; j++ ) - { - mnVal += m_ChromaBuffer[ k + (j*m_BPO) ]; + for( j = 0; j < m_ChromaBufferFilling; j++ ) { + mnVal += m_ChromaBuffer[ k + (j*kBinsPerOctave) ]; } m_MeanHPCP[k] = mnVal/(double)m_ChromaBufferFilling; } - - for( k = 0; k < m_BPO; k++ ) + // Normalize for zero average + double mHPCP = MathUtilities::mean( m_MeanHPCP, kBinsPerOctave ); + for( k = 0; k < kBinsPerOctave; k++ ) { - m_MajCorr[k] = krumCorr( m_MeanHPCP, MajProfile, m_BPO ); - m_MinCorr[k] = krumCorr( m_MeanHPCP, MinProfile, m_BPO ); - - MathUtilities::circShift( MajProfile, m_BPO, 1 ); - MathUtilities::circShift( MinProfile, m_BPO, 1 ); - } - - for( k = 0; k < m_BPO; k++ ) - { - m_Keys[k] = m_MajCorr[k]; - m_Keys[k+m_BPO] = m_MinCorr[k]; + m_MeanHPCP[k] -= mHPCP; } - for (k = 0; k < 24; ++k) { - m_keyStrengths[k] = 0; - } - for( k = 0; k < m_BPO*2; k++ ) + for( k = 0; k < kBinsPerOctave; k++ ) { - int idx = k / (m_BPO/12); - int rem = k % (m_BPO/12); - if (rem == 0 || m_Keys[k] > m_keyStrengths[idx]) { - m_keyStrengths[idx] = m_Keys[k]; - } - -// m_keyStrengths[k/(m_BPO/12)] += m_Keys[k]; + // The Cromagram has the center of C at bin 0, while the major + // and minor profiles have the center of C at 1. We want to have + // the correlation for C result also at 1. + // To achieve this we have to shift two times: + m_MajCorr[k] = krumCorr( m_MeanHPCP, m_MajProfileNorm, (int)k - 2, kBinsPerOctave ); + m_MinCorr[k] = krumCorr( m_MeanHPCP, m_MinProfileNorm, (int)k - 2, kBinsPerOctave ); } /* std::cout << "raw keys: "; - for (int ii = 0; ii < 2*m_BPO; ++ii) { - if (ii % (m_BPO/12) == 0) std::cout << "\n"; - std::cout << m_Keys[ii] << " "; + for (int ii = 0; ii < kBinsPerOctave; ++ii) { + if (ii % (kBinsPerOctave/12) == 0) std::cout << "\n"; + std::cout << m_MajCorr[ii] << " "; } - std::cout << std::endl; - - std::cout << "key strengths: "; - for (int ii = 0; ii < 24; ++ii) { - if (ii % 6 == 0) std::cout << "\n"; - std::cout << m_keyStrengths[ii] << " "; + for (int ii = 0; ii < kBinsPerOctave; ++ii) { + if (ii % (kBinsPerOctave/12) == 0) std::cout << "\n"; + std::cout << m_MinCorr[ii] << " "; } std::cout << std::endl; */ - double dummy; - // '1 +' because we number keys 1-24, not 0-23. - key = 1 + (int)ceil( (double)MathUtilities::getMax( m_Keys, 2* m_BPO, &dummy )/3 ); + // m_MajCorr[1] is C center 1 / 3 + 1 = 1 + // m_MajCorr[4] is D center 4 / 3 + 1 = 2 + // '+ 1' because we number keys 1-24, not 0-23. + double maxMaj; + int maxMajBin = MathUtilities::getMax( m_MajCorr, kBinsPerOctave, &maxMaj ); + double maxMin; + int maxMinBin = MathUtilities::getMax( m_MinCorr, kBinsPerOctave, &maxMin ); + int maxBin = (maxMaj > maxMin) ? maxMajBin : (maxMinBin + kBinsPerOctave); + key = maxBin / 3 + 1; + +// std::cout << "fractional key pre-sorting: " << (maxBin + 2) / 3.0 << std::endl; // std::cout << "key pre-sorting: " << key << std::endl; //Median filtering // track Median buffer initial filling - if( m_MedianBufferFilling++ >= m_MedianWinsize) + if( m_MedianBufferFilling++ >= m_MedianWinsize) { m_MedianBufferFilling = m_MedianWinsize; - + } + //shift median buffer - for( k = 1; k < m_MedianWinsize; k++ ) - { + for( k = 1; k < m_MedianWinsize; k++ ) { m_MedianFilterBuffer[ k - 1 ] = m_MedianFilterBuffer[ k ]; } @@ -290,8 +282,7 @@ //Copy median into sorting buffer, reversed unsigned int ijx = 0; - for( k = 0; k < m_MedianWinsize; k++ ) - { + for( k = 0; k < m_MedianWinsize; k++ ) { m_SortedBuffer[k] = m_MedianFilterBuffer[m_MedianWinsize-1-ijx]; ijx++; } @@ -310,8 +301,9 @@ // std::cout << "midpoint = " << midpoint << endl; - if( midpoint <= 0 ) + if( midpoint <= 0 ) { midpoint = 1; + } key = m_SortedBuffer[midpoint-1]; @@ -325,3 +317,45 @@ { return (key > 12); } + +unsigned int getChromaSize() +{ + return kBinsPerOctave; +} + +double* GetKeyMode::getKeyStrengths() { + unsigned int k; + + for (k = 0; k < 24; ++k) { + m_keyStrengths[k] = 0; + } + + for( k = 0; k < kBinsPerOctave; k++ ) + { + int idx = k / (kBinsPerOctave/12); + int rem = k % (kBinsPerOctave/12); + if (rem == 0 || m_MajCorr[k] > m_keyStrengths[idx]) { + m_keyStrengths[idx] = m_MajCorr[k]; + } + } + + for( k = 0; k < kBinsPerOctave; k++ ) + { + int idx = (k + kBinsPerOctave) / (kBinsPerOctave/12); + int rem = k % (kBinsPerOctave/12); + if (rem == 0 || m_MinCorr[k] > m_keyStrengths[idx]) { + m_keyStrengths[idx] = m_MinCorr[k]; + } + } + +/* + std::cout << "key strengths: "; + for (int ii = 0; ii < 24; ++ii) { + if (ii % 6 == 0) std::cout << "\n"; + std::cout << m_keyStrengths[ii] << " "; + } + std::cout << std::endl; +*/ + + return m_keyStrengths; +}
--- a/dsp/keydetection/GetKeyMode.h Tue May 28 13:41:29 2019 +0100 +++ b/dsp/keydetection/GetKeyMode.h Tue May 28 13:45:08 2019 +0100 @@ -27,17 +27,18 @@ int process( double* PCMData ); - double krumCorr( double* pData1, double* pData2, unsigned int length ); + double krumCorr( const double *pDataNorm, const double *pProfileNorm, + int shiftProfile, unsigned int length ); unsigned int getBlockSize() { return m_ChromaFrameSize*m_DecimationFactor; } unsigned int getHopSize() { return m_ChromaHopSize*m_DecimationFactor; } double* getChroma() { return m_ChrPointer; } - unsigned int getChromaSize() { return m_BPO; } + unsigned int getChromaSize(); double* getMeanHPCP() { return m_MeanHPCP; } - double *getKeyStrengths() { return m_keyStrengths; } + double* getKeyStrengths(); bool isModeMinor( int key ); @@ -63,8 +64,6 @@ unsigned int m_ChromaFrameSize; //Hop unsigned int m_ChromaHopSize; - //Bins per octave - unsigned int m_BPO; unsigned int m_ChromaBuffersize; @@ -79,9 +78,10 @@ double* m_ChromaBuffer; double* m_MeanHPCP; + double* m_MajProfileNorm; + double* m_MinProfileNorm; double* m_MajCorr; double* m_MinCorr; - double* m_Keys; int* m_MedianFilterBuffer; int* m_SortedBuffer;