max@1: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
max@1: 
max@1: /*
max@1:     QM Vamp Plugin Set
max@1: 
max@1:     Centre for Digital Music, Queen Mary, University of London.
max@1: 
max@1:     This program is free software; you can redistribute it and/or
max@1:     modify it under the terms of the GNU General Public License as
max@1:     published by the Free Software Foundation; either version 2 of the
max@1:     License, or (at your option) any later version.  See the file
max@1:     COPYING included with this distribution for more information.
max@1: */
max@1: 
max@1: #include "SongParts.h"
max@1: 
max@1: #include <base/Window.h>
max@1: #include <dsp/onsets/DetectionFunction.h>
max@1: #include <dsp/onsets/PeakPicking.h>
max@1: #include <dsp/transforms/FFT.h>
max@1: #include <dsp/tempotracking/TempoTrackV2.h>
max@1: #include <dsp/tempotracking/DownBeat.h>
max@1: #include <chromamethods.h>
max@1: #include <maths/MathUtilities.h>
max@1: #include <boost/numeric/ublas/matrix.hpp>
max@1: #include <boost/numeric/ublas/io.hpp>
max@1: #include <boost/math/distributions/normal.hpp>
max@1: #include "armadillo"
max@1: #include <fstream>
max@1: #include <sstream>
max@1: #include <cmath>
max@1: #include <vector>
max@1: 
max@1: #include <vamp-sdk/Plugin.h>
max@1: 
max@1: using namespace boost::numeric;
max@1: using namespace arma;
max@1: using std::string;
max@1: using std::vector;
max@1: using std::cerr;
max@1: using std::cout;
max@1: using std::endl;
max@1: 
max@1: 
max@1: #ifndef __GNUC__
max@1: #include <alloca.h>
max@1: #endif
max@1: 
max@1: 
max@1: // Result Struct
max@1: typedef struct Part {
max@1:     int n;
max@1:     vector<unsigned> indices;
max@1:     string letter;
max@1:     unsigned value;
max@1:     int level;
max@1:     int nInd;
max@1: }Part;
max@1: 
max@1: 
max@1: /* ------------------------------------ */
max@1: /* ----- BEAT DETECTOR CLASS ---------- */
max@1: /* ------------------------------------ */
max@1: 
max@1: class BeatTrackerData
max@1: {
max@1:     /* --- ATTRIBUTES --- */
max@1: public:
max@1:     DFConfig dfConfig;
max@1:     DetectionFunction *df;
max@1:     DownBeat *downBeat;
max@1:     vector<double> dfOutput;
max@1:     Vamp::RealTime origin;
max@1:     
max@1:     
max@1:     /* --- METHODS --- */
max@1:     
max@1:     /* --- Constructor --- */
max@1: public:
max@1:     BeatTrackerData(float rate, const DFConfig &config) : dfConfig(config) {
max@1: 	
max@1:         df = new DetectionFunction(config);
max@1:         // decimation factor aims at resampling to c. 3KHz; must be power of 2
max@1:         int factor = MathUtilities::nextPowerOfTwo(rate / 3000);
max@1:         // std::cerr << "BeatTrackerData: factor = " << factor << std::endl;
max@1:         downBeat = new DownBeat(rate, factor, config.stepSize);
max@1:     }
max@1:     
max@1:     /* --- Desctructor --- */
max@1:     ~BeatTrackerData() {
max@1: 	delete df;
max@1:         delete downBeat;
max@1:     }
max@1:     
max@1:     void reset() {
max@1:         delete df;
max@1:         df = new DetectionFunction(dfConfig);
max@1:         dfOutput.clear();
max@1:         downBeat->resetAudioBuffer();
max@1:         origin = Vamp::RealTime::zeroTime;
max@1:     }
max@1: };
max@1: 
max@1: 
max@1: /* --------------------------------------- */
max@1: /* ----- CHROMA EXTRACTOR CLASS ---------- */
max@1: /* --------------------------------------- */
max@1: 
max@1: class ChromaData
max@1: {
max@1:     
max@1:      /* --- ATTRIBUTES --- */
max@1:     
max@1: public:
max@1:     int frameCount;
max@1:     int nBPS;
max@1:     Vamp::Plugin::FeatureList logSpectrum;
max@1:     size_t blockSize;
max@1:     int lengthOfNoteIndex;
max@1:     vector<float> meanTunings;
max@1:     vector<float> localTunings;
max@1:     float whitening;
max@1:     float preset;
max@1:     float useNNLS;
max@1:     vector<float> localTuning;
max@1:     vector<float> kernelValue;
max@1:     vector<int> kernelFftIndex;
max@1:     vector<int> kernelNoteIndex;
max@1:     float *dict;
max@1:     bool tuneLocal;
max@1:     float doNormalizeChroma;
max@1:     float rollon;
max@1:     float s;
max@1:     vector<float> hw;
max@1:     vector<float> sinvalues;
max@1:     vector<float> cosvalues;
max@1:     Window<float> window;
max@1:     FFTReal fft;
max@1:     size_t inputSampleRate;
max@1:     
max@1:     /* --- METHODS --- */
max@1:     
max@1:     /* --- Constructor --- */
max@1:     
max@1: public:
max@1:     ChromaData(float inputSampleRate, size_t block_size) :
max@1:         frameCount(0),
max@1:         nBPS(3),
max@1:         logSpectrum(0),
max@1:         blockSize(0),
max@1:         lengthOfNoteIndex(0),
max@1:         meanTunings(0),
max@1:         localTunings(0),
max@1:         whitening(1.0),
max@1:         preset(0.0),
max@1:         useNNLS(1.0),
max@1:         localTuning(0.0),
max@1:         kernelValue(0),
max@1:         kernelFftIndex(0),
max@1:         kernelNoteIndex(0),
max@1:         dict(0),
max@1:         tuneLocal(0.0),
max@1:         doNormalizeChroma(0),
max@1:         rollon(0.0),
max@1:     	s(0.7),
max@1:     	sinvalues(0),
max@1:     	cosvalues(0),
max@1:     	window(HanningWindow, block_size),
max@1:     	fft(block_size),
max@1:     	inputSampleRate(inputSampleRate)
max@1:     {
max@1:         // make the *note* dictionary matrix
max@1:         dict = new float[nNote * 84];
max@1:         for (int i = 0; i < nNote * 84; ++i) dict[i] = 0.0;
max@1:         blockSize = block_size;
max@1:     }
max@1:     
max@1:     
max@1:     /* --- Desctructor --- */
max@1:     
max@1:     ~ChromaData() {
max@1:         delete [] dict;
max@1:     }
max@1:     
max@1:     /* --- Public Methods --- */
max@1:     
max@1:     void reset() {
max@1:         frameCount = 0;
max@1:         logSpectrum.clear();
max@1:         for (int iBPS = 0; iBPS < 3; ++iBPS) {
max@1:             meanTunings[iBPS] = 0;
max@1:             localTunings[iBPS] = 0;
max@1:         }
max@1:         localTuning.clear();
max@1:     }
max@1:     
max@1:     void baseProcess(float *inputBuffers, Vamp::RealTime timestamp)
max@1:     {   
max@1:         	
max@1:         frameCount++;   
max@1:         float *magnitude = new float[blockSize/2];
max@1:         double *fftReal = new double[blockSize];
max@1:         double *fftImag = new double[blockSize];
max@1: 
max@1:         // FFTReal wants doubles, so we need to make a local copy of inputBuffers
max@1:         double *inputBuffersDouble = new double[blockSize];
max@1:         for (size_t i = 0; i < blockSize; i++) inputBuffersDouble[i] = inputBuffers[i];
max@1:         
max@1:         fft.process(false, inputBuffersDouble, fftReal, fftImag);
max@1:         
max@1:         float energysum = 0;
max@1:         // make magnitude
max@1:         float maxmag = -10000;
max@1:         for (int iBin = 0; iBin < static_cast<int>(blockSize/2); iBin++) {
max@1:             magnitude[iBin] = sqrt(fftReal[iBin] * fftReal[iBin] + 
max@1:                                    fftImag[iBin] * fftImag[iBin]);
max@1:             if (magnitude[iBin]>blockSize*1.0) magnitude[iBin] = blockSize; 
max@1:             // a valid audio signal (between -1 and 1) should not be limited here.
max@1:             if (maxmag < magnitude[iBin]) maxmag = magnitude[iBin];
max@1:             if (rollon > 0) {
max@1:                 energysum += pow(magnitude[iBin],2);
max@1:             }
max@1:         }
max@1:     
max@1:         float cumenergy = 0;
max@1:         if (rollon > 0) {
max@1:             for (int iBin = 2; iBin < static_cast<int>(blockSize/2); iBin++) {
max@1:                 cumenergy +=  pow(magnitude[iBin],2);
max@1:                 if (cumenergy < energysum * rollon / 100) magnitude[iBin-2] = 0;
max@1:                 else break;
max@1:             }
max@1:         }
max@1:     
max@1:         if (maxmag < 2) {
max@1:             // cerr << "timestamp " << timestamp << ": very low magnitude, setting magnitude to all zeros" << endl;
max@1:             for (int iBin = 0; iBin < static_cast<int>(blockSize/2); iBin++) {
max@1:                 magnitude[iBin] = 0;
max@1:             }
max@1:         }
max@1:         
max@1:         // cerr << magnitude[200] << endl;
max@1:         
max@1:         // note magnitude mapping using pre-calculated matrix
max@1:         float *nm  = new float[nNote]; // note magnitude
max@1:         for (int iNote = 0; iNote < nNote; iNote++) {
max@1:             nm[iNote] = 0; // initialise as 0
max@1:         }
max@1:         int binCount = 0;
max@1:         for (vector<float>::iterator it = kernelValue.begin(); it != kernelValue.end(); ++it) {
max@1:             nm[kernelNoteIndex[binCount]] += magnitude[kernelFftIndex[binCount]] * kernelValue[binCount];
max@1:             binCount++;  
max@1:         }
max@1:     
max@1:         float one_over_N = 1.0/frameCount;
max@1:         // update means of complex tuning variables
max@1:         for (int iBPS = 0; iBPS < nBPS; ++iBPS) meanTunings[iBPS] *= float(frameCount-1)*one_over_N;
max@1:     
max@1:         for (int iTone = 0; iTone < round(nNote*0.62/nBPS)*nBPS+1; iTone = iTone + nBPS) {
max@1:             for (int iBPS = 0; iBPS < nBPS; ++iBPS) meanTunings[iBPS] += nm[iTone + iBPS]*one_over_N;
max@1:             float ratioOld = 0.997;
max@1:             for (int iBPS = 0; iBPS < nBPS; ++iBPS) {
max@1:                 localTunings[iBPS] *= ratioOld; 
max@1:                 localTunings[iBPS] += nm[iTone + iBPS] * (1 - ratioOld);
max@1:             }
max@1:         }
max@1:     
max@1:         float localTuningImag = 0;
max@1:         float localTuningReal = 0;
max@1:         for (int iBPS = 0; iBPS < nBPS; ++iBPS) {
max@1:             localTuningReal += localTunings[iBPS] * cosvalues[iBPS];
max@1:             localTuningImag += localTunings[iBPS] * sinvalues[iBPS];
max@1:         }
max@1:     
max@1:         float normalisedtuning = atan2(localTuningImag, localTuningReal)/(2*M_PI);
max@1:         localTuning.push_back(normalisedtuning);
max@1:     
max@1:         Vamp::Plugin::Feature f1; // logfreqspec
max@1:         f1.hasTimestamp = true;
max@1:         f1.timestamp = timestamp;
max@1:         for (int iNote = 0; iNote < nNote; iNote++) {
max@1:             f1.values.push_back(nm[iNote]);
max@1:         }
max@1:     
max@1:         // deletes
max@1:         delete[] inputBuffersDouble;
max@1:         delete[] magnitude;
max@1:         delete[] fftReal;
max@1:         delete[] fftImag;
max@1:         delete[] nm;
max@1:     
max@1:         logSpectrum.push_back(f1); // remember note magnitude
max@1:     }
max@1:     
max@1:     bool initialise()
max@1:     {
max@1:         dictionaryMatrix(dict, s);
max@1: 	
max@1:     	// make things for tuning estimation
max@1:     	for (int iBPS = 0; iBPS < nBPS; ++iBPS) {
max@1:             sinvalues.push_back(sin(2*M_PI*(iBPS*1.0/nBPS)));
max@1:             cosvalues.push_back(cos(2*M_PI*(iBPS*1.0/nBPS)));
max@1:         }
max@1:     
max@1: 	
max@1:     	// make hamming window of length 1/2 octave
max@1:     	int hamwinlength = nBPS * 6 + 1;
max@1:         float hamwinsum = 0;
max@1:         for (int i = 0; i < hamwinlength; ++i) { 
max@1:             hw.push_back(0.54 - 0.46 * cos((2*M_PI*i)/(hamwinlength-1)));    
max@1:             hamwinsum += 0.54 - 0.46 * cos((2*M_PI*i)/(hamwinlength-1));
max@1:         }
max@1:         for (int i = 0; i < hamwinlength; ++i) hw[i] = hw[i] / hamwinsum;
max@1:     
max@1:     
max@1:         // initialise the tuning
max@1:         for (int iBPS = 0; iBPS < nBPS; ++iBPS) {
max@1:             meanTunings.push_back(0);
max@1:             localTunings.push_back(0);
max@1:         }
max@1: 	
max@1:         blockSize = blockSize;
max@1:         frameCount = 0;
max@1:         int tempn = nNote * blockSize/2;
max@1:         // cerr << "length of tempkernel : " <<  tempn << endl;
max@1:         float *tempkernel;
max@1: 
max@1:         tempkernel = new float[tempn];
max@1: 
max@1:         logFreqMatrix(inputSampleRate, blockSize, tempkernel);
max@1:         kernelValue.clear();
max@1:         kernelFftIndex.clear();
max@1:         kernelNoteIndex.clear();
max@1:         int countNonzero = 0;
max@1:         for (int iNote = 0; iNote < nNote; ++iNote) { 
max@1:             // I don't know if this is wise: manually making a sparse matrix
max@1:             for (int iFFT = 0; iFFT < static_cast<int>(blockSize/2); ++iFFT) {
max@1:                 if (tempkernel[iFFT + blockSize/2 * iNote] > 0) {
max@1:                     kernelValue.push_back(tempkernel[iFFT + blockSize/2 * iNote]);
max@1:                     if (tempkernel[iFFT + blockSize/2 * iNote] > 0) {
max@1:                         countNonzero++;
max@1:                     }
max@1:                     kernelFftIndex.push_back(iFFT);
max@1:                     kernelNoteIndex.push_back(iNote);				
max@1:                 }
max@1:             }
max@1:         }
max@1:         delete [] tempkernel;
max@1:     }    
max@1: };
max@1: 
max@1: 
max@1: /* --------------------------------- */
max@1: /* ----- SONG PARTITIONER ---------- */
max@1: /* --------------------------------- */
max@1: 
max@1: 
max@1: /* --- ATTRIBUTES --- */
max@1: 
max@1: float SongPartitioner::m_stepSecs = 0.01161;            // 512 samples at 44100
max@1: size_t SongPartitioner::m_chromaFramesizeFactor = 16;   // 16 times as long as beat tracker's
max@1: size_t SongPartitioner::m_chromaStepsizeFactor = 4;     // 4 times as long as beat tracker's
max@1: 
max@1: 
max@1: /* --- METHODS --- */
max@1: 
max@1: /* --- Constructor --- */
max@1: SongPartitioner::SongPartitioner(float inputSampleRate) :
max@1:     Vamp::Plugin(inputSampleRate),
max@1:     m_d(0),
max@1:     m_bpb(4),
max@1:     m_pluginFrameCount(0)
max@1: {
max@1: }
max@1: 
max@1: 
max@1: /* --- Desctructor --- */
max@1: SongPartitioner::~SongPartitioner()
max@1: {
max@1:     delete m_d;
max@1: }
max@1: 
max@1: 
max@1: /* --- Methods --- */
max@1: string SongPartitioner::getIdentifier() const
max@1: {
max@1:     return "qm-songpartitioner";
max@1: }
max@1: 
max@1: string SongPartitioner::getName() const
max@1: {
max@1:     return "Song Partitioner";
max@1: }
max@1: 
max@1: string SongPartitioner::getDescription() const
max@1: {
max@1:     return "Estimate contiguous segments pertaining to song parts such as verse and chorus.";
max@1: }
max@1: 
max@1: string SongPartitioner::getMaker() const
max@1: {
max@1:     return "Queen Mary, University of London";
max@1: }
max@1: 
max@1: int SongPartitioner::getPluginVersion() const
max@1: {
max@1:     return 2;
max@1: }
max@1: 
max@1: string SongPartitioner::getCopyright() const
max@1: {
max@1:     return "Plugin by Matthew Davies, Christian Landone, Chris Cannam, Matthias Mauch and Massimiliano Zanoni  Copyright (c) 2006-2012 QMUL - All Rights Reserved";
max@1: }
max@1: 
max@1: SongPartitioner::ParameterList SongPartitioner::getParameterDescriptors() const
max@1: {
max@1:     ParameterList list;
max@1: 
max@1:     ParameterDescriptor desc;
max@1: 
max@1:     desc.identifier = "bpb";
max@1:     desc.name = "Beats per Bar";
max@1:     desc.description = "The number of beats in each bar";
max@1:     desc.minValue = 2;
max@1:     desc.maxValue = 16;
max@1:     desc.defaultValue = 4;
max@1:     desc.isQuantized = true;
max@1:     desc.quantizeStep = 1;
max@1:     list.push_back(desc);
max@1: 
max@1:     return list;
max@1: }
max@1: 
max@1: float SongPartitioner::getParameter(std::string name) const
max@1: {
max@1:     if (name == "bpb") return m_bpb;
max@1:     return 0.0;
max@1: }
max@1: 
max@1: void SongPartitioner::setParameter(std::string name, float value)
max@1: {
max@1:     if (name == "bpb") m_bpb = lrintf(value);
max@1: }
max@1: 
max@1: 
max@1: // Return the StepSize for Chroma Extractor 
max@1: size_t SongPartitioner::getPreferredStepSize() const
max@1: {
max@1:     size_t step = size_t(m_inputSampleRate * m_stepSecs + 0.0001);
max@1:     if (step < 1) step = 1;
max@1: 
max@1:     return step;
max@1: }
max@1: 
max@1: // Return the BlockSize for Chroma Extractor 
max@1: size_t SongPartitioner::getPreferredBlockSize() const
max@1: {
max@1:     size_t theoretical = getPreferredStepSize() * 2;
max@1:     theoretical *= m_chromaFramesizeFactor; 
max@1: 
max@1:     return theoretical;
max@1: }
max@1: 
max@1: 
max@1: // Initialize the plugin and define Beat Tracker and Chroma Extractor Objects
max@1: bool SongPartitioner::initialise(size_t channels, size_t stepSize, size_t blockSize)
max@1: {
max@1:     if (m_d) {
max@1: 	delete m_d;
max@1: 	m_d = 0;
max@1:     }
max@1: 
max@1:     if (channels < getMinChannelCount() ||
max@1: 	channels > getMaxChannelCount()) {
max@1:         std::cerr << "SongPartitioner::initialise: Unsupported channel count: "
max@1:                   << channels << std::endl;
max@1:         return false;
max@1:     }
max@1: 
max@1:     if (stepSize != getPreferredStepSize()) {
max@1:         std::cerr << "ERROR: SongPartitioner::initialise: Unsupported step size for this sample rate: "
max@1:                   << stepSize << " (wanted " << (getPreferredStepSize()) << ")" << std::endl;
max@1:         return false;
max@1:     }
max@1: 
max@1:     if (blockSize != getPreferredBlockSize()) {
max@1:         std::cerr << "WARNING: SongPartitioner::initialise: Sub-optimal block size for this sample rate: "
max@1:                   << blockSize << " (wanted " << getPreferredBlockSize() << ")" << std::endl;
max@1:     }
max@1: 
max@1:     // Beat tracker and Chroma extractor has two different configuration parameters 
max@1:     
max@1:     // Configuration Parameters for Beat Tracker
max@1:     DFConfig dfConfig;
max@1:     dfConfig.DFType = DF_COMPLEXSD;
max@1:     dfConfig.stepSize = stepSize;
max@1:     dfConfig.frameLength = blockSize / m_chromaFramesizeFactor;
max@1:     dfConfig.dbRise = 3;
max@1:     dfConfig.adaptiveWhitening = false;
max@1:     dfConfig.whiteningRelaxCoeff = -1;
max@1:     dfConfig.whiteningFloor = -1;
max@1:     
max@1:     // Initialise Beat Tracker
max@1:     m_d = new BeatTrackerData(m_inputSampleRate, dfConfig);
max@1:     m_d->downBeat->setBeatsPerBar(m_bpb);
max@1:     
max@1:     // Initialise Chroma Extractor
max@1:     m_chromadata = new ChromaData(m_inputSampleRate, blockSize);
max@1:     m_chromadata->initialise();
max@1:     
max@1:     return true;
max@1: }
max@1: 
max@1: void SongPartitioner::reset()
max@1: {
max@1:     if (m_d) m_d->reset();
max@1:     m_pluginFrameCount = 0;
max@1: }
max@1: 
max@1: SongPartitioner::OutputList SongPartitioner::getOutputDescriptors() const
max@1: {
max@1:     OutputList list;
max@1:     size_t outputCounter = 0;
max@1: 
max@1:     OutputDescriptor beat;
max@1:     beat.identifier = "beats";
max@1:     beat.name = "Beats";
max@1:     beat.description = "Beat locations labelled with metrical position";
max@1:     beat.unit = "";
max@1:     beat.hasFixedBinCount = true;
max@1:     beat.binCount = 0;
max@1:     beat.sampleType = OutputDescriptor::VariableSampleRate;
max@1:     beat.sampleRate = 1.0 / m_stepSecs;
max@1:     m_beatOutputNumber = outputCounter++;
max@1: 
max@1:     OutputDescriptor bars;
max@1:     bars.identifier = "bars";
max@1:     bars.name = "Bars";
max@1:     bars.description = "Bar locations";
max@1:     bars.unit = "";
max@1:     bars.hasFixedBinCount = true;
max@1:     bars.binCount = 0;
max@1:     bars.sampleType = OutputDescriptor::VariableSampleRate;
max@1:     bars.sampleRate = 1.0 / m_stepSecs;
max@1:     m_barsOutputNumber = outputCounter++;
max@1: 
max@1:     OutputDescriptor beatcounts;
max@1:     beatcounts.identifier = "beatcounts";
max@1:     beatcounts.name = "Beat Count";
max@1:     beatcounts.description = "Beat counter function";
max@1:     beatcounts.unit = "";
max@1:     beatcounts.hasFixedBinCount = true;
max@1:     beatcounts.binCount = 1;
max@1:     beatcounts.sampleType = OutputDescriptor::VariableSampleRate;
max@1:     beatcounts.sampleRate = 1.0 / m_stepSecs;
max@1:     m_beatcountsOutputNumber = outputCounter++;
max@1: 
max@1:     OutputDescriptor beatsd;
max@1:     beatsd.identifier = "beatsd";
max@1:     beatsd.name = "Beat Spectral Difference";
max@1:     beatsd.description = "Beat spectral difference function used for bar-line detection";
max@1:     beatsd.unit = "";
max@1:     beatsd.hasFixedBinCount = true;
max@1:     beatsd.binCount = 1;
max@1:     beatsd.sampleType = OutputDescriptor::VariableSampleRate;
max@1:     beatsd.sampleRate = 1.0 / m_stepSecs;
max@1:     m_beatsdOutputNumber = outputCounter++;
max@1:     
max@1:     OutputDescriptor logscalespec;
max@1:     logscalespec.identifier = "logscalespec";
max@1:     logscalespec.name = "Log-Frequency Spectrum";
max@1:     logscalespec.description = "Spectrum with linear frequency on a log scale.";
max@1:     logscalespec.unit = "";
max@1:     logscalespec.hasFixedBinCount = true;
max@1:     logscalespec.binCount = nNote;
max@1:     logscalespec.hasKnownExtents = false;
max@1:     logscalespec.isQuantized = false;
max@1:     logscalespec.sampleType = OutputDescriptor::FixedSampleRate;
max@1:     logscalespec.hasDuration = false;
max@1:     logscalespec.sampleRate = m_inputSampleRate/2048;
max@1:     m_logscalespecOutputNumber = outputCounter++;
max@1:     
max@1:     OutputDescriptor bothchroma;
max@1:     bothchroma.identifier = "bothchroma";
max@1:     bothchroma.name = "Chromagram and Bass Chromagram";
max@1:     bothchroma.description = "Tuning-adjusted chromagram and bass chromagram (stacked on top of each other) from NNLS approximate transcription.";
max@1:     bothchroma.unit = "";
max@1:     bothchroma.hasFixedBinCount = true;
max@1:     bothchroma.binCount = 24;
max@1:     bothchroma.hasKnownExtents = false;
max@1:     bothchroma.isQuantized = false;
max@1:     bothchroma.sampleType = OutputDescriptor::FixedSampleRate;
max@1:     bothchroma.hasDuration = false;
max@1:     bothchroma.sampleRate = m_inputSampleRate/2048;
max@1:     m_bothchromaOutputNumber = outputCounter++;
max@1:     
max@1:     OutputDescriptor qchromafw;
max@1:     qchromafw.identifier = "qchromafw";
max@1:     qchromafw.name = "Pseudo-Quantised Chromagram and Bass Chromagram";
max@1:     qchromafw.description = "Pseudo-Quantised Chromagram and Bass Chromagram (frames between two beats are identical).";
max@1:     qchromafw.unit = "";
max@1:     qchromafw.hasFixedBinCount = true;
max@1:     qchromafw.binCount = 24;
max@1:     qchromafw.hasKnownExtents = false;
max@1:     qchromafw.isQuantized = false;
max@1:     qchromafw.sampleType = OutputDescriptor::FixedSampleRate;
max@1:     qchromafw.hasDuration = false;
max@1:     qchromafw.sampleRate = m_inputSampleRate/2048;
max@1:     m_qchromafwOutputNumber = outputCounter++;    
max@1:     
max@1:     OutputDescriptor qchroma;
max@1:     qchroma.identifier = "qchroma";
max@1:     qchroma.name = "Quantised Chromagram and Bass Chromagram";
max@1:     qchroma.description = "Quantised Chromagram and Bass Chromagram.";
max@1:     qchroma.unit = "";
max@1:     qchroma.hasFixedBinCount = true;
max@1:     qchroma.binCount = 24;
max@1:     qchroma.hasKnownExtents = false;
max@1:     qchroma.isQuantized = false;
max@1:     qchroma.sampleType = OutputDescriptor::FixedSampleRate;
max@1:     qchroma.hasDuration = true;
max@1:     m_qchromaOutputNumber = outputCounter++;
max@1: 
max@1:     OutputDescriptor segm;
max@1:     segm.identifier = "segm";
max@1:     segm.name = "Segmentation";
max@1:     segm.description = "Segmentation";
max@1:     segm.unit = "segment-type";
max@1:     segm.hasFixedBinCount = true;
max@1:     //segm.binCount = 24;
max@1:     segm.binCount = 1;
max@1:     segm.hasKnownExtents = true;
max@1:     segm.minValue = 1;
max@1:     segm.maxValue = 5;
max@1:     segm.isQuantized = true;
max@1:     segm.quantizeStep = 1;
max@1:     segm.sampleType = OutputDescriptor::VariableSampleRate;
max@1:     segm.hasDuration = true;
max@1:     m_segmOutputNumber = outputCounter++;
max@1:     
max@1:     
max@1:    /* 
max@1:     OutputList list;
max@1:     OutputDescriptor segmentation;
max@1:     segmentation.identifier = "segmentation";
max@1:     segmentation.name = "Segmentation";
max@1:     segmentation.description = "Segmentation";
max@1:     segmentation.unit = "segment-type";
max@1:     segmentation.hasFixedBinCount = true;
max@1:     segmentation.binCount = 1;
max@1:     segmentation.hasKnownExtents = true;
max@1:     segmentation.minValue = 1;
max@1:     segmentation.maxValue = nSegmentTypes;
max@1:     segmentation.isQuantized = true;
max@1:     segmentation.quantizeStep = 1;
max@1:     segmentation.sampleType = OutputDescriptor::VariableSampleRate;
max@1:     segmentation.sampleRate = m_inputSampleRate / getPreferredStepSize();
max@1:     list.push_back(segmentation);
max@1:     return list;
max@1:     */
max@1:     
max@1:     
max@1:     list.push_back(beat);
max@1:     list.push_back(bars);
max@1:     list.push_back(beatcounts);
max@1:     list.push_back(beatsd);
max@1:     list.push_back(logscalespec);
max@1:     list.push_back(bothchroma);
max@1:     list.push_back(qchromafw);
max@1:     list.push_back(qchroma);
max@1:     list.push_back(segm);
max@1: 
max@1:     return list;
max@1: }
max@1: 
max@1: // Executed for each frame - called from the host  
max@1: 
max@1: // We use time domain input, because DownBeat requires it -- so we
max@1: // use the time-domain version of DetectionFunction::process which
max@1: // does its own FFT.  It requires doubles as input, so we need to
max@1: // make a temporary copy
max@1: 
max@1: // We only support a single input channel
max@1: SongPartitioner::FeatureSet SongPartitioner::process(const float *const *inputBuffers,Vamp::RealTime timestamp)
max@1: {
max@1:     if (!m_d) {
max@1: 	cerr << "ERROR: SongPartitioner::process: "
max@1: 	     << "SongPartitioner has not been initialised"
max@1: 	     << endl;
max@1: 	return FeatureSet();
max@1:     }
max@1: 
max@1:     const int fl = m_d->dfConfig.frameLength;
max@1: #ifndef __GNUC__
max@1:     double *dfinput = (double *)alloca(fl * sizeof(double));
max@1: #else
max@1:     double dfinput[fl];
max@1: #endif
max@1:     int sampleOffset = ((m_chromaFramesizeFactor-1) * fl) / 2;
max@1:     
max@1:     // Since chroma needs a much longer frame size, we only ever use the very
max@1:     // beginning of the frame for beat tracking.
max@1:     for (int i = 0; i < fl; ++i) dfinput[i] = inputBuffers[0][i];
max@1:     double output = m_d->df->process(dfinput);
max@1: 
max@1:     if (m_d->dfOutput.empty()) m_d->origin = timestamp;
max@1: 
max@1: //    std::cerr << "df[" << m_d->dfOutput.size() << "] is " << output << std::endl;
max@1:     m_d->dfOutput.push_back(output);
max@1: 
max@1:     // Downsample and store the incoming audio block.
max@1:     // We have an overlap on the incoming audio stream (step size is
max@1:     // half block size) -- this function is configured to take only a
max@1:     // step size's worth, so effectively ignoring the overlap.  Note
max@1:     // however that this means we omit the last blocksize - stepsize
max@1:     // samples completely for the purposes of barline detection
max@1:     // (hopefully not a problem)
max@1:     m_d->downBeat->pushAudioBlock(inputBuffers[0]);
max@1: 
max@1:     // The following is not done every time, but only every m_chromaFramesizeFactor times,
max@1:     // because the chroma does not need dense time frames.
max@1:     
max@1:     if (m_pluginFrameCount % m_chromaStepsizeFactor == 0)
max@1:     {    
max@1:         
max@1:         // Window the full time domain, data, FFT it and process chroma stuff.
max@1:     
max@1:         #ifndef __GNUC__
max@1:             float *windowedBuffers = (float *)alloca(m_chromadata->blockSize * sizeof(float));
max@1:         #else
max@1:             float windowedBuffers[m_chromadata->blockSize];
max@1:         #endif
max@1:         m_chromadata->window.cut(&inputBuffers[0][0], &windowedBuffers[0]);
max@1:     
max@1:         // adjust timestamp (we want the middle of the frame)
max@1:         timestamp = timestamp + Vamp::RealTime::frame2RealTime(sampleOffset, lrintf(m_inputSampleRate));
max@1: 
max@1:         m_chromadata->baseProcess(&windowedBuffers[0], timestamp);
max@1:         
max@1:     }
max@1:     m_pluginFrameCount++;
max@1:     
max@1:     FeatureSet fs;
max@1:     fs[m_logscalespecOutputNumber].push_back(
max@1:         m_chromadata->logSpectrum.back());
max@1:     return fs;
max@1: }
max@1: 
max@1: SongPartitioner::FeatureSet SongPartitioner::getRemainingFeatures()
max@1: {
max@1:     if (!m_d) {
max@1: 	cerr << "ERROR: SongPartitioner::getRemainingFeatures: "
max@1: 	     << "SongPartitioner has not been initialised"
max@1: 	     << endl;
max@1: 	return FeatureSet();
max@1:     }
max@1: 
max@1:     FeatureSet masterFeatureset = BeatTrack();
max@1:     FeatureList chromaList = ChromaFeatures();
max@1:     
max@1:     for (size_t i = 0; i < chromaList.size(); ++i)
max@1:     {
max@1:         masterFeatureset[m_bothchromaOutputNumber].push_back(chromaList[i]);
max@1:     }
max@1:     
max@1:     // quantised and pseudo-quantised (beat-wise) chroma
max@1:     std::vector<FeatureList> quantisedChroma = BeatQuantiser(chromaList, masterFeatureset[m_beatOutputNumber]);
max@1:     
max@1:     masterFeatureset[m_qchromafwOutputNumber] = quantisedChroma[0];
max@1:     masterFeatureset[m_qchromaOutputNumber] = quantisedChroma[1];
max@1:     
max@1:     // Segmentation
max@1:     masterFeatureset[m_segmOutputNumber] = Segmenter(quantisedChroma[1]);
max@1:     
max@1:     return(masterFeatureset);
max@1: }
max@1: 
max@1: /* ------ Beat Tracker ------ */
max@1: 
max@1: SongPartitioner::FeatureSet SongPartitioner::BeatTrack()
max@1: {
max@1:     vector<double> df;
max@1:     vector<double> beatPeriod;
max@1:     vector<double> tempi;
max@1:     
max@1:     for (size_t i = 2; i < m_d->dfOutput.size(); ++i) { // discard first two elts
max@1:         df.push_back(m_d->dfOutput[i]);
max@1:         beatPeriod.push_back(0.0);
max@1:     }
max@1:     if (df.empty()) return FeatureSet();
max@1: 
max@1:     TempoTrackV2 tt(m_inputSampleRate, m_d->dfConfig.stepSize);
max@1:     tt.calculateBeatPeriod(df, beatPeriod, tempi);
max@1: 
max@1:     vector<double> beats;
max@1:     tt.calculateBeats(df, beatPeriod, beats);
max@1: 
max@1:     vector<int> downbeats;
max@1:     size_t downLength = 0;
max@1:     const float *downsampled = m_d->downBeat->getBufferedAudio(downLength);
max@1:     m_d->downBeat->findDownBeats(downsampled, downLength, beats, downbeats);
max@1: 
max@1:     vector<double> beatsd;
max@1:     m_d->downBeat->getBeatSD(beatsd);
max@1:     
max@1:     /*std::cout << "BeatTracker: found downbeats at: ";
max@1:     for (int i = 0; i < downbeats.size(); ++i) std::cout << downbeats[i] << " " << std::endl;*/
max@1:     
max@1:     FeatureSet returnFeatures;
max@1: 
max@1:     char label[20];
max@1: 
max@1:     int dbi = 0;
max@1:     int beat = 0;
max@1:     int bar = 0;
max@1: 
max@1:     if (!downbeats.empty()) {
max@1:         // get the right number for the first beat; this will be
max@1:         // incremented before use (at top of the following loop)
max@1:         int firstDown = downbeats[0];
max@1:         beat = m_bpb - firstDown - 1;
max@1:         if (beat == m_bpb) beat = 0;
max@1:     }
max@1: 
max@1:     for (size_t i = 0; i < beats.size(); ++i) {
max@1:         
max@1:         size_t frame = beats[i] * m_d->dfConfig.stepSize;
max@1:         
max@1:         if (dbi < downbeats.size() && i == downbeats[dbi]) {
max@1:             beat = 0;
max@1:             ++bar;
max@1:             ++dbi;
max@1:         } else {
max@1:             ++beat;
max@1:         }
max@1:         
max@1:         /* Ooutput Section */
max@1:         
max@1:         // outputs are:
max@1:         //
max@1:         // 0 -> beats
max@1:         // 1 -> bars
max@1:         // 2 -> beat counter function
max@1:         
max@1:         Feature feature;
max@1:         feature.hasTimestamp = true;
max@1:         feature.timestamp = m_d->origin + Vamp::RealTime::frame2RealTime (frame, lrintf(m_inputSampleRate));
max@1:         
max@1:         sprintf(label, "%d", beat + 1);
max@1:         feature.label = label;
max@1:         returnFeatures[m_beatOutputNumber].push_back(feature);          // labelled beats
max@1:         
max@1:         feature.values.push_back(beat + 1);
max@1:         returnFeatures[m_beatcountsOutputNumber].push_back(feature);    // beat function
max@1:         
max@1:         if (i > 0 && i <= beatsd.size()) {
max@1:             feature.values.clear();
max@1:             feature.values.push_back(beatsd[i-1]);
max@1:             feature.label = "";
max@1:             returnFeatures[m_beatsdOutputNumber].push_back(feature);    // beat spectral difference
max@1:         }
max@1:         
max@1:         if (beat == 0) {
max@1:             feature.values.clear();
max@1:             sprintf(label, "%d", bar);
max@1:             feature.label = label;
max@1:             returnFeatures[m_barsOutputNumber].push_back(feature);      // bars
max@1:         }
max@1:     }
max@1: 
max@1:     return returnFeatures;
max@1: }
max@1: 
max@1: 
max@1: /* ------ Chroma Extractor ------ */
max@1: 
max@1: SongPartitioner::FeatureList SongPartitioner::ChromaFeatures()
max@1: {
max@1:         
max@1:     FeatureList returnFeatureList;
max@1:     FeatureList tunedlogfreqspec;
max@1:     
max@1:     if (m_chromadata->logSpectrum.size() == 0) return returnFeatureList;
max@1: 
max@1:     /**  Calculate Tuning
max@1:          calculate tuning from (using the angle of the complex number defined by the 
max@1:          cumulative mean real and imag values)
max@1:     **/
max@1:     float meanTuningImag = 0;
max@1:     float meanTuningReal = 0;
max@1:     for (int iBPS = 0; iBPS < nBPS; ++iBPS) {
max@1:         meanTuningReal += m_chromadata->meanTunings[iBPS] * m_chromadata->cosvalues[iBPS];
max@1:         meanTuningImag += m_chromadata->meanTunings[iBPS] * m_chromadata->sinvalues[iBPS];
max@1:     }
max@1:     float cumulativetuning = 440 * pow(2,atan2(meanTuningImag, meanTuningReal)/(24*M_PI));
max@1:     float normalisedtuning = atan2(meanTuningImag, meanTuningReal)/(2*M_PI);
max@1:     int intShift = floor(normalisedtuning * 3);
max@1:     float floatShift = normalisedtuning * 3 - intShift; // floatShift is a really bad name for this
max@1:      
max@1:     char buffer0 [50];
max@1:  
max@1:     sprintf(buffer0, "estimated tuning: %0.1f Hz", cumulativetuning);
max@1:                  
max@1:     /** Tune Log-Frequency Spectrogram
max@1:         calculate a tuned log-frequency spectrogram (f2): use the tuning estimated above (kinda f0) to 
max@1:         perform linear interpolation on the existing log-frequency spectrogram (kinda f1).
max@1:     **/
max@1:     cerr << endl << "[NNLS Chroma Plugin] Tuning Log-Frequency Spectrogram ... ";
max@1:              
max@1:     float tempValue = 0;
max@1: 
max@1:     int count = 0;
max@1:     
max@1:     for (FeatureList::iterator i = m_chromadata->logSpectrum.begin(); i != m_chromadata->logSpectrum.end(); ++i) 
max@1:     {
max@1:         
max@1:         Feature f1 = *i;
max@1:         Feature f2; // tuned log-frequency spectrum
max@1:         
max@1:         f2.hasTimestamp = true;
max@1:         f2.timestamp = f1.timestamp;
max@1:         
max@1:         f2.values.push_back(0.0); 
max@1:         f2.values.push_back(0.0); // set lower edge to zero
max@1: 
max@1:         if (m_chromadata->tuneLocal) {
max@1:             intShift = floor(m_chromadata->localTuning[count] * 3);
max@1:             floatShift = m_chromadata->localTuning[count] * 3 - intShift; 
max@1:             // floatShift is a really bad name for this
max@1:         }
max@1: 
max@1:         for (int k = 2; k < (int)f1.values.size() - 3; ++k) 
max@1:         { // interpolate all inner bins
max@1:             tempValue = f1.values[k + intShift] * (1-floatShift) + f1.values[k+intShift+1] * floatShift;
max@1:             f2.values.push_back(tempValue);
max@1:         }
max@1:          
max@1:         f2.values.push_back(0.0); 
max@1:         f2.values.push_back(0.0); 
max@1:         f2.values.push_back(0.0); // upper edge
max@1: 
max@1:         vector<float> runningmean = SpecialConvolution(f2.values,m_chromadata->hw);
max@1:         vector<float> runningstd;
max@1:         for (int i = 0; i < nNote; i++) { // first step: squared values into vector (variance)
max@1:             runningstd.push_back((f2.values[i] - runningmean[i]) * (f2.values[i] - runningmean[i]));
max@1:         }
max@1:         runningstd = SpecialConvolution(runningstd,m_chromadata->hw); // second step convolve
max@1:         for (int i = 0; i < nNote; i++) 
max@1:         { 
max@1:             
max@1:             runningstd[i] = sqrt(runningstd[i]); 
max@1:             // square root to finally have running std
max@1:             
max@1:             if (runningstd[i] > 0) 
max@1:             {
max@1:                 f2.values[i] = (f2.values[i] - runningmean[i]) > 0 ?
max@1:                     (f2.values[i] - runningmean[i]) / pow(runningstd[i],m_chromadata->whitening) : 0;
max@1:             }
max@1:             
max@1:             if (f2.values[i] < 0) {
max@1:                 
max@1:                 cerr << "ERROR: negative value in logfreq spectrum" << endl;
max@1:                 
max@1:             }
max@1:         }
max@1:         tunedlogfreqspec.push_back(f2);
max@1:         count++;
max@1:     }
max@1:     cerr << "done." << endl;    
max@1:     /** Semitone spectrum and chromagrams
max@1:         Semitone-spaced log-frequency spectrum derived 
max@1:         from the tuned log-freq spectrum above. the spectrum
max@1:         is inferred using a non-negative least squares algorithm.
max@1:         Three different kinds of chromagram are calculated, "treble", "bass", and "both" (which means 
max@1:         bass and treble stacked onto each other).
max@1:     **/
max@1:     if (m_chromadata->useNNLS == 0) {
max@1:         cerr << "[NNLS Chroma Plugin] Mapping to semitone spectrum and chroma ... ";
max@1:     } else {
max@1:         cerr << "[NNLS Chroma Plugin] Performing NNLS and mapping to chroma ... ";
max@1:     }
max@1:     
max@1:     vector<float> oldchroma = vector<float>(12,0);
max@1:     vector<float> oldbasschroma = vector<float>(12,0);
max@1:     count = 0;
max@1: 
max@1:     for (FeatureList::iterator it = tunedlogfreqspec.begin(); it != tunedlogfreqspec.end(); ++it) {
max@1:         Feature logfreqsp = *it; // logfreq spectrum
max@1:         Feature bothchroma; // treble and bass chromagram
max@1:                             
max@1:         bothchroma.hasTimestamp = true;
max@1:         bothchroma.timestamp = logfreqsp.timestamp;
max@1:         
max@1:         float b[nNote];
max@1: 
max@1:         bool some_b_greater_zero = false;
max@1:         float sumb = 0;
max@1:         for (int i = 0; i < nNote; i++) {
max@1:             b[i] = logfreqsp.values[i];
max@1:             sumb += b[i];
max@1:             if (b[i] > 0) {
max@1:                 some_b_greater_zero = true;
max@1:             }            
max@1:         }
max@1:     
max@1:         // here's where the non-negative least squares algorithm calculates the note activation x
max@1: 
max@1:         vector<float> chroma = vector<float>(12, 0);
max@1:         vector<float> basschroma = vector<float>(12, 0);
max@1:         float currval;
max@1:         int iSemitone = 0;
max@1:      
max@1:         if (some_b_greater_zero) {
max@1:             if (m_chromadata->useNNLS == 0) {
max@1:                 for (int iNote = nBPS/2 + 2; iNote < nNote - nBPS/2; iNote += nBPS) {
max@1:                     currval = 0;
max@1:                     for (int iBPS = -nBPS/2; iBPS < nBPS/2+1; ++iBPS) {
max@1:                         currval += b[iNote + iBPS] * (1-abs(iBPS*1.0/(nBPS/2+1)));                       
max@1:                     }
max@1:                     chroma[iSemitone % 12] += currval * treblewindow[iSemitone];
max@1:                     basschroma[iSemitone % 12] += currval * basswindow[iSemitone];
max@1:                     iSemitone++;
max@1:                 }
max@1:          
max@1:             } else {
max@1:                 float x[84+1000];
max@1:                 for (int i = 1; i < 1084; ++i) x[i] = 1.0;
max@1:                 vector<int> signifIndex;
max@1:                 int index=0;
max@1:                 sumb /= 84.0;
max@1:                 for (int iNote = nBPS/2 + 2; iNote < nNote - nBPS/2; iNote += nBPS) {
max@1:                     float currval = 0;
max@1:                     for (int iBPS = -nBPS/2; iBPS < nBPS/2+1; ++iBPS) {
max@1:                         currval += b[iNote + iBPS]; 
max@1:                     }
max@1:                     if (currval > 0) signifIndex.push_back(index);
max@1:                     index++;
max@1:                 }
max@1:                 float rnorm;
max@1:                 float w[84+1000];
max@1:                 float zz[84+1000];
max@1:                 int indx[84+1000];
max@1:                 int mode;
max@1:                 int dictsize = nNote*signifIndex.size();
max@1: 
max@1:                 float *curr_dict = new float[dictsize];
max@1:                 for (int iNote = 0; iNote < (int)signifIndex.size(); ++iNote) {
max@1:                     for (int iBin = 0; iBin < nNote; iBin++) {
max@1:                         curr_dict[iNote * nNote + iBin] = 
max@1:                             1.0 * m_chromadata->dict[signifIndex[iNote] * nNote + iBin];
max@1:                     }
max@1:                 }
max@1:                 nnls(curr_dict, nNote, nNote, signifIndex.size(), b, x, &rnorm, w, zz, indx, &mode);
max@1:                 delete [] curr_dict;
max@1:                 for (int iNote = 0; iNote < (int)signifIndex.size(); ++iNote) {
max@1:                     // cerr << mode << endl;
max@1:                     chroma[signifIndex[iNote] % 12] += x[iNote] * treblewindow[signifIndex[iNote]];
max@1:                     basschroma[signifIndex[iNote] % 12] += x[iNote] * basswindow[signifIndex[iNote]];
max@1:                 }
max@1:             }    
max@1:         }
max@1:  
max@1:         chroma.insert(chroma.begin(), basschroma.begin(), basschroma.end()); 
max@1:         // just stack the both chromas 
max@1:         
max@1:         bothchroma.values = chroma; 
max@1:         returnFeatureList.push_back(bothchroma);
max@1:         count++;
max@1:     }
max@1:     cerr << "done." << endl;
max@1: 
max@1:     return returnFeatureList;     
max@1: }
max@1: 
max@1: /* ------ Beat Quantizer ------ */
max@1: 
max@4: std::vector<Vamp::Plugin::FeatureList>
max@4: SongPartitioner::BeatQuantiser(Vamp::Plugin::FeatureList chromagram, Vamp::Plugin::FeatureList beats)
max@1: {
max@1:     std::vector<FeatureList> returnVector;
max@1:     
max@1:     FeatureList fwQchromagram; // frame-wise beat-quantised chroma
max@1:     FeatureList bwQchromagram; // beat-wise beat-quantised chroma
max@1:     
max@4:     int nChromaFrame = (int) chromagram.size();
max@4:     int nBeat = (int) beats.size();
max@1:     
max@1:     if (nBeat == 0 && nChromaFrame == 0) return returnVector;
max@1:     
max@1:     size_t nBin = chromagram[0].values.size();
max@1:     
max@1:     vector<float> tempChroma = vector<float>(nBin);
max@1:     
max@1:     Vamp::RealTime beatTimestamp = Vamp::RealTime::zeroTime;
max@1:     int currBeatCount = -1; // start before first beat
max@1:     int framesInBeat = 0;
max@1:     
max@4:     for (int iChroma = 0; iChroma < nChromaFrame; ++iChroma)
max@1:     {
max@4:         Vamp::RealTime frameTimestamp = chromagram[iChroma].timestamp;
max@4: 		Vamp::RealTime tempBeatTimestamp;
max@4: 		
max@4: 		if (currBeatCount != beats.size()-1) tempBeatTimestamp = beats[currBeatCount+1].timestamp;
max@4: 		else tempBeatTimestamp = chromagram[nChromaFrame-1].timestamp;
max@4: 		
max@4:         if (frameTimestamp > tempBeatTimestamp ||
max@1:             iChroma == nChromaFrame-1)
max@1:         {
max@1:             // new beat (or last chroma frame)
max@1:             // 1. finish all the old beat processing
max@4: 			if (framesInBeat > 0)
max@4: 			{
max@4:             	for (int i = 0; i < nBin; ++i) tempChroma[i] /= framesInBeat; // average
max@4: 			}
max@1:             
max@1:             Feature bwQchromaFrame;
max@1:             bwQchromaFrame.hasTimestamp = true;
max@1:             bwQchromaFrame.timestamp = beatTimestamp;
max@1:             bwQchromaFrame.values = tempChroma;
max@1:             bwQchromaFrame.duration = beats[currBeatCount+1].timestamp - beats[currBeatCount].timestamp;
max@1:             bwQchromagram.push_back(bwQchromaFrame);
max@1:             
max@1:             for (int iFrame = -framesInBeat; iFrame < 0; ++iFrame)
max@1:             {
max@1:                 Feature fwQchromaFrame;
max@1:                 fwQchromaFrame.hasTimestamp = true;
max@1:                 fwQchromaFrame.timestamp = chromagram[iChroma+iFrame].timestamp;
max@1:                 fwQchromaFrame.values = tempChroma; // all between two beats get the same
max@1:                 fwQchromagram.push_back(fwQchromaFrame);
max@1:             }
max@1:             
max@1:             // 2. increments / resets for current (new) beat
max@1:             currBeatCount++;
max@1:             beatTimestamp = beats[currBeatCount].timestamp;
max@1:             for (size_t i = 0; i < nBin; ++i) tempChroma[i] = 0; // average
max@1:             framesInBeat = 0;
max@1:         }
max@1:         framesInBeat++;
max@1:         for (size_t i = 0; i < nBin; ++i) tempChroma[i] += chromagram[iChroma].values[i];
max@1:     }
max@1:     returnVector.push_back(fwQchromagram);
max@1:     returnVector.push_back(bwQchromagram);
max@1: }
max@1: 
max@1: /* -------------------------------- */
max@1: /* ------ Support Functions  ------ */
max@1: /* -------------------------------- */
max@1: 
max@1: // one-dimesion median filter
max@1: arma::vec medfilt1(arma::vec v, int medfilt_length)
max@1: {    
max@1:     int halfWin = medfilt_length/2;
max@1:     
max@1:     // result vector
max@1:     arma::vec res = arma::zeros<arma::vec>(v.size());
max@1:     
max@1:     // padding 
max@1:     arma::vec padV = arma::zeros<arma::vec>(v.size()+medfilt_length-1);
max@1:     
max@1:     for (unsigned i=medfilt_length/2; i < medfilt_length/2+v.size(); ++ i)
max@1:     {
max@1:         padV(i) = v(i-medfilt_length/2);
max@1:     }    
max@1:     
max@1:     // Median filter
max@1:     arma::vec win = arma::zeros<arma::vec>(medfilt_length);
max@1:     
max@1:     for (unsigned i=0; i < v.size(); ++i)
max@1:     {
max@1:         win = padV.subvec(i,i+halfWin*2);
max@1:         win = sort(win);
max@1:         res(i) = win(halfWin);
max@1:     }
max@1:     
max@1:     return res;
max@1: }
max@1: 
max@1: 
max@1: // Quantile
max@1: double quantile(arma::vec v, double p)
max@1: {
max@1:     arma::vec sortV = arma::sort(v);
max@1:     int n = sortV.size();
max@1:     arma::vec x = arma::zeros<vec>(n+2);
max@1:     arma::vec y = arma::zeros<vec>(n+2);
max@1:     
max@1:     x(0) = 0;
max@1:     x(n+1) = 100; 
max@1:     
max@1:     for (unsigned i=1; i<n+1; ++i)
max@1:         x(i) = 100*(0.5+(i-1))/n;
max@1:         
max@1:     y(0) = sortV(0);
max@1:     y.subvec(1,n) = sortV;
max@1:     y(n+1) = sortV(n-1);
max@1:     
max@1:     arma::uvec x2index = find(x>=p*100);
max@1:     
max@1:     // Interpolation
max@1:     double x1 = x(x2index(0)-1);
max@1:     double x2 = x(x2index(0));
max@1:     double y1 = y(x2index(0)-1);
max@1:     double y2 = y(x2index(0));
max@1:     
max@1:     double res = (y2-y1)/(x2-x1)*(p*100-x1)+y1;
max@1:     
max@1:     return res;
max@1: }
max@1: 
max@1: // Max Filtering
max@1: arma::mat maxfilt1(arma::mat inmat, int len)
max@1: {
max@1:     arma::mat outmat = inmat;
max@1:     
max@1:     for (int i=0; i<inmat.n_rows; ++i)
max@1:     {
max@1:         if (arma::sum(inmat.row(i)) > 0)
max@1:         {
max@1:             // Take a window of rows
max@1:             int startWin;
max@1:             int endWin;
max@1:             
max@1:             if (0 > i-len)
max@1:                 startWin = 0;
max@1:             else
max@1:                 startWin = i-len;
max@1:             
max@1:             if (inmat.n_rows-1 < i+len-1)
max@1:                 endWin = inmat.n_rows-1;
max@1:             else
max@1:                 endWin = i+len-1;
max@1:     
max@1:             outmat(i,span::all) = arma::max(inmat(span(startWin,endWin),span::all));
max@1:         }
max@1:     }
max@1:     
max@1:     return outmat;
max@1:     
max@1: }
max@1: 
max@1: // Null Parts
max@1: Part nullpart(vector<Part> parts, arma::vec barline)
max@1: {
max@1:     arma::uvec nullindices = arma::ones<arma::uvec>(barline.size());
max@1:     for (unsigned iPart=0; iPart<parts.size(); ++iPart)
max@1:     {
max@7:         //for (unsigned iIndex=0; iIndex < parts[0].indices.size(); ++iIndex) 
max@7:         for (unsigned iIndex=0; iIndex < parts[iPart].indices.size(); ++iIndex) 
max@1:             for (unsigned i=0; i<parts[iPart].n; ++i) 
max@1:             {
max@1:                 unsigned ind = parts[iPart].indices[iIndex]+i;
max@1:                 nullindices(ind) = 0;
max@1:             }
max@1:     }
max@7: 
max@1:     Part newPart;
max@1:     newPart.n = 1;
max@1:     uvec q = find(nullindices > 0);
max@1:     
max@1:     for (unsigned i=0; i<q.size();++i) 
max@1:         newPart.indices.push_back(q(i));
max@7: 
max@1:     newPart.letter = '-';
max@1:     newPart.value = 0;
max@1:     newPart.level = 0;
max@1:     
max@1:     return newPart;    
max@1: }
max@1: 
max@1: 
max@1: // Merge Nulls
max@1: void mergenulls(vector<Part> &parts)
max@1: {
max@1:     for (unsigned iPart=0; iPart<parts.size(); ++iPart)
max@1:     {
max@1:         
max@1:         vector<Part> newVectorPart;
max@1:         
max@1:         if (parts[iPart].letter.compare("-")==0)
max@1:         {
max@1:             sort (parts[iPart].indices.begin(), parts[iPart].indices.end());
max@1:             unsigned newpartind = -1;
max@1:             
max@1:             vector<int> indices;
max@1:             indices.push_back(-2);
max@1:             
max@1:             for (unsigned iIndex=0; iIndex<parts[iPart].indices.size(); ++iIndex) 
max@1:                 indices.push_back(parts[iPart].indices[iIndex]);
max@1:             
max@1:             for (unsigned iInd=1; iInd < indices.size(); ++iInd)
max@1:             { 
max@1:                 if (indices[iInd] - indices[iInd-1] > 1)
max@1:                 {
max@1:                     newpartind++;
max@1: 
max@1:                     Part newPart;
max@1:                     newPart.letter = 'n';
max@1:                     std::stringstream out;
max@1:                     out << newpartind+1;
max@1:                     newPart.letter.append(out.str());
max@1:                     newPart.value = 20+newpartind+1;
max@1:                     newPart.n = 1;
max@1:                     newPart.indices.push_back(indices[iInd]);
max@1:                     newPart.level = 0;   
max@1:                     
max@1:                     newVectorPart.push_back(newPart);
max@1:                 }
max@1:                 else
max@1:                 {
max@1:                     newVectorPart[newpartind].n = newVectorPart[newpartind].n+1;
max@1:                 }
max@1:             }
max@1:             parts.erase (parts.end());
max@1:             
max@1:             for (unsigned i=0; i<newVectorPart.size(); ++i)
max@1:                 parts.push_back(newVectorPart[i]);
max@1:         }
max@1:     }
max@1: }
max@1: 
max@1: /* ------ Segmentation ------ */
max@1: 
max@1: vector<Part> songSegment(Vamp::Plugin::FeatureList quatisedChromagram)
max@1: {
max@1:     
max@1:     
max@1:     /* ------ Parameters ------ */
max@1:     double thresh_beat = 0.85;
max@1:     double thresh_seg = 0.80;
max@1:     int medfilt_length = 5;
max@1:     int minlength = 28;
max@1:     int maxlength = 128;
max@1:     double quantilePerc = 0.1;
max@1:     /* ------------------------ */
max@1:     
max@1:     
max@1:     // Collect Info
max@1:     int nBeat = quatisedChromagram.size();                      // Number of feature vector
max@1:     int nFeatValues = quatisedChromagram[0].values.size();      // Number of values for each feature vector
max@1:     
max@1:     arma::irowvec timeStamp = arma::zeros<arma::imat>(1,nBeat);       // Vector of Time Stamps
max@1:     
max@1: 	// Save time stamp as a Vector
max@1:     if (quatisedChromagram[0].hasTimestamp)
max@1:     {
max@1:         for (unsigned i = 0; i < nBeat; ++ i)
max@1:             timeStamp[i] = quatisedChromagram[i].timestamp.nsec;
max@1:     }
max@1:     
max@1:     
max@1:     // Build a ObservationTOFeatures Matrix
max@1:     arma::mat featVal = arma::zeros<mat>(nBeat,nFeatValues/2);
max@1:     
max@1:     for (unsigned i = 0; i < nBeat; ++ i)
max@1:         for (unsigned j = 0; j < nFeatValues/2; ++ j)
max@1:         {
max@1:             featVal(i,j) = (quatisedChromagram[i].values[j]+quatisedChromagram[i].values[j+12]) * 0.8;
max@1:         }
max@1:     
max@1:     // Set to arbitrary value to feature vectors with low std
max@1:     arma::mat a = stddev(featVal,1,1);
max@1:     
max@1:     // Feature Colleration Matrix
max@1:     arma::mat simmat0 = 1-arma::cor(arma::trans(featVal));
max@1:     
max@1: 
max@1:     for (unsigned i = 0; i < nBeat; ++ i)
max@1:     {
max@1:         if (a(i)<0.000001)
max@1:         {
max@1:             featVal(i,1) = 1000;  // arbitrary  
max@1:             
max@1:             for (unsigned j = 0; j < nFeatValues/2; ++j)
max@1:             {
max@1:                 simmat0(i,j) = 1;
max@1:                 simmat0(j,i) = 1;
max@1:             }
max@1:         }
max@1:     }
max@1:     
max@1:     arma::mat simmat = 1-simmat0/2;
max@1:     
max@1:     // -------- To delate when the proble with the add of beat will be solved -------
max@1:     for (unsigned i = 0; i < nBeat; ++ i)
max@1:      for (unsigned j = 0; j < nBeat; ++ j)
max@1:          if (!std::isfinite(simmat(i,j)))
max@1:              simmat(i,j)=0;
max@1:     // ------------------------------------------------------------------------------
max@1:     
max@1:     // Median Filtering applied to the Correlation Matrix
max@1:     // The median filter is for each diagonal of the Matrix
max@1:     arma::mat median_simmat = arma::zeros<arma::mat>(nBeat,nBeat);
max@1:     
max@1:     for (unsigned i = 0; i < nBeat; ++ i)
max@1:     {
max@1:         arma::vec temp = medfilt1(simmat.diag(i),medfilt_length);
max@1:         median_simmat.diag(i) = temp;
max@1:         median_simmat.diag(-i) = temp;
max@1:     }
max@1: 
max@1:     for (unsigned i = 0; i < nBeat; ++ i)
max@1:         for (unsigned j = 0; j < nBeat; ++ j)
max@1:             if (!std::isfinite(median_simmat(i,j)))
max@1:                 median_simmat(i,j) = 0;
max@1:     
max@1:     // -------------- NOT CONVERTED -------------------------------------    
max@1:     //    if param.seg.standardise
max@1:     //        med_median_simmat = repmat(median(median_simmat),nBeat,1);
max@1:     //    std_median_simmat = repmat(std(median_simmat),nBeat,1);
max@1:     //    median_simmat = (median_simmat - med_median_simmat) ./ std_median_simmat;
max@1:     //    end
max@1:     // --------------------------------------------------------
max@1:     
max@1:     // Retrieve Bar Bounderies
max@1:     arma::uvec dup = find(median_simmat > thresh_beat);
max@1:     arma::mat potential_duplicates = arma::zeros<arma::mat>(nBeat,nBeat);
max@1:     potential_duplicates.elem(dup) = arma::ones<arma::vec>(dup.size());
max@1:     potential_duplicates = trimatu(potential_duplicates);
max@1:     
max@1:     unsigned nPartlengths = round((maxlength-minlength)/4)+1;
max@1:     arma::vec partlengths = zeros<arma::vec>(nPartlengths);
max@1:     
max@1:     for (unsigned i = 0; i < nPartlengths; ++ i)
max@1:         partlengths(i) = (i*4)+ minlength;
max@1:     
max@1:     // initialise arrays
max@1:     arma::cube simArray = zeros<arma::cube>(nBeat,nBeat,nPartlengths);
max@1:     arma::cube decisionArray2 = zeros<arma::cube>(nBeat,nBeat,nPartlengths);
max@1: 
max@1:     int conta = 0;
max@1:        
max@1:     //for (unsigned iLength = 0; iLength < nPartlengths; ++ iLength)
max@1:     for (unsigned iLength = 0; iLength < 20; ++ iLength)
max@1:     {
max@1:         unsigned len = partlengths(iLength);
max@1:         unsigned nUsedBeat = nBeat - len + 1;                   // number of potential rep beginnings: they can't overlap at the end of the song
max@1:         
max@1:         for (unsigned iBeat = 0; iBeat < nUsedBeat; ++ iBeat)   // looping over all columns (arbitrarily chosen columns)
max@1:         {
max@1:             arma::uvec help2 = find(potential_duplicates(span(0,nUsedBeat-1),iBeat)==1);
max@1:             
max@1:             for (unsigned i=0; i<help2.size(); ++i)
max@1:             {
max@1: 
max@1:                 // measure how well two length len segments go together
max@1:                 int kBeat = help2(i);
max@1:                 arma::vec distrib = median_simmat(span(iBeat,iBeat+len-1),span(kBeat,kBeat+len-1)).diag(0);
max@1:                 simArray(iBeat,kBeat,iLength) = quantile(distrib,quantilePerc);
max@1:             }
max@1:         }
max@1:         
max@1:         arma::mat tempM = simArray(span(0,nUsedBeat-1),span(0,nUsedBeat-1),span(iLength,iLength));
max@1:         simArray.slice(iLength)(span(0,nUsedBeat-1),span(0,nUsedBeat-1)) = tempM + arma::trans(tempM) - (eye<mat>(nUsedBeat,nUsedBeat)%tempM); 
max@1:         
max@1:         // convolution
max@1:         arma::vec K = arma::zeros<vec>(3);
max@1:         K << 0.01 << 0.98 << 0.01;
max@1:         
max@1:         
max@1:         for (unsigned i=0; i<simArray.n_rows; ++i)
max@1:         {
max@1:             arma::rowvec t = arma::conv((arma::rowvec)simArray.slice(iLength).row(i),K);
max@1:             simArray.slice(iLength)(i,span::all) = t.subvec(1,t.size()-2);
max@1:         }
max@1:  
max@1:         // take only over-average bars that do not overlap
max@1:         
max@1:         arma::mat temp = arma::zeros<mat>(simArray.n_rows, simArray.n_cols);
max@1:         temp(span::all, span(0,nUsedBeat-1)) = simArray.slice(iLength)(span::all,span(0,nUsedBeat-1));
max@1:         
max@1:         for (unsigned i=0; i<temp.n_rows; ++i)
max@1:             for (unsigned j=0; j<nUsedBeat; ++j)
max@1:                 if (temp(i,j) < thresh_seg)
max@1:                     temp(i,j) = 0;
max@1:         
max@1:         decisionArray2.slice(iLength) = temp;
max@1: 
max@1:         arma::mat maxMat = maxfilt1(decisionArray2.slice(iLength),len-1);
max@1:         
max@1:         for (unsigned i=0; i<decisionArray2.n_rows; ++i)
max@1:             for (unsigned j=0; j<decisionArray2.n_cols; ++j)
max@1:                 if (decisionArray2.slice(iLength)(i,j) < maxMat(i,j))
max@1:                     decisionArray2.slice(iLength)(i,j) = 0;
max@1:         
max@1:         decisionArray2.slice(iLength) = decisionArray2.slice(iLength) % arma::trans(decisionArray2.slice(iLength));
max@1:         
max@1:         for (unsigned i=0; i<simArray.n_rows; ++i)
max@1:             for (unsigned j=0; j<simArray.n_cols; ++j)
max@1:                 if (simArray.slice(iLength)(i,j) < thresh_seg)
max@1:                     potential_duplicates(i,j) = 0; 
max@1:     }
max@1:     
max@1:     // Milk the data
max@1:     
max@1:     arma::mat bestval;
max@1:     
max@1:     for (unsigned iLength=0; iLength<nPartlengths; ++iLength)
max@1:     {
max@1:         arma::mat temp = arma::zeros<arma::mat>(decisionArray2.n_rows,decisionArray2.n_cols);
max@1: 
max@1:        for (unsigned rows=0; rows<decisionArray2.n_rows; ++rows)
max@1:             for (unsigned cols=0; cols<decisionArray2.n_cols; ++cols)
max@1:                 if (decisionArray2.slice(iLength)(rows,cols) > 0)
max@1:                     temp(rows,cols) = 1;
max@1:         
max@1:         arma::vec currLogicSum = arma::sum(temp,1);
max@1:         
max@1:         for (unsigned iBeat=0; iBeat<nBeat; ++iBeat)
max@1:             if (currLogicSum(iBeat) > 1)
max@1:             {
max@1:                 arma::vec t = decisionArray2.slice(iLength)(span::all,iBeat);
max@1:                 double currSum = sum(t);
max@1:                 
max@1:                 unsigned count = 0;
max@1:                 for (unsigned i=0; i<t.size(); ++i)
max@1:                     if (t(i)>0)
max@1:                         count++;
max@1:                 
max@1:                 currSum = (currSum/count)/2;
max@1:                 
max@1:                 arma::rowvec t1;
max@1:                 t1 << (currLogicSum(iBeat)-1) * partlengths(iLength) << currSum << iLength << iBeat << currLogicSum(iBeat);
max@1:                 
max@1:                 bestval = join_cols(bestval,t1);
max@1:             }
max@1:     }
max@1:     
max@1:     // Definition of the resulting vector
max@1:     vector<Part> parts;
max@1:     
max@1:     // make a table of all valid sets of parts
max@1:     
max@1:     char partletters[] = {'A','B','C','D','E','F','G', 'H','I','J','K','L','M','N','O','P','Q','R','S'};
max@1:     unsigned partvalues[] = {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19};
max@1:     arma::vec valid_sets = arma::ones<arma::vec>(bestval.n_rows);
max@1:     
max@1:     if (!bestval.is_empty())
max@1:     {
max@1:         
max@1:         // In questo punto viene introdotto un errore alla 3 cifra decimale
max@1:         
max@1:         arma::colvec t = arma::zeros<arma::colvec>(bestval.n_rows);
max@1:         for (unsigned i=0; i<bestval.n_rows; ++i)
max@1:         {
max@1:             t(i) = bestval(i,1)*2;
max@1:         }
max@1:         
max@1:         double m = t.max();
max@1:         
max@1:         bestval(span::all,1) = bestval(span::all,1) / m; 
max@1:         bestval(span::all,0) = bestval(span::all,0) + bestval(span::all,1);
max@1:         
max@1:         arma::mat bestval2;
max@1:         for (unsigned i=0; i<bestval.n_cols; ++i)
max@1:             if (i!=1)
max@1:                 bestval2 = join_rows(bestval2,bestval.col(i));
max@1:         
max@6:         for (unsigned kSeg=0; kSeg<6; ++kSeg)
max@1:         {
max@1:             arma::mat currbestvals = arma::zeros<arma::mat>(bestval2.n_rows, bestval2.n_cols);
max@1:             for (unsigned i=0; i<bestval2.n_rows; ++i)
max@1:                 for (unsigned j=0; j<bestval2.n_cols; ++j)
max@1:                     if (valid_sets(i))
max@1:                         currbestvals(i,j) = bestval2(i,j);
max@1:             
max@1:             arma::vec t1 = currbestvals.col(0);
max@1:             double ma;
max@1:             uword maIdx;
max@1:             ma = t1.max(maIdx);
max@6:             
max@6:             if ((maIdx == 0)&&(ma == 0))
max@6:                 break;
max@1: 
max@1:             double bestLength = partlengths(currbestvals(maIdx,1));
max@1:             arma::rowvec bestIndices = decisionArray2.slice(currbestvals(maIdx,1))(currbestvals(maIdx,2),span::all);
max@1:                 
max@1:             arma::rowvec bestIndicesMap = arma::zeros<arma::rowvec>(bestIndices.size());
max@1:             for (unsigned i=0; i<bestIndices.size(); ++i)
max@1:                 if (bestIndices(i)>0)
max@1:                     bestIndicesMap(i) = 1;
max@1:                    
max@1:             arma::rowvec mask = arma::zeros<arma::rowvec>(bestLength*2-1);
max@1:             for (unsigned i=0; i<bestLength; ++i)
max@1:                 mask(i+bestLength-1) = 1;
max@1:             
max@1:             arma::rowvec t2 = arma::conv(bestIndicesMap,mask); 
max@1:             arma::rowvec island = t2.subvec(mask.size()/2,t2.size()-1-mask.size()/2);
max@1:             
max@1:             // Save results in the structure
max@1:             Part newPart;
max@1:             newPart.n = bestLength;
max@1:             uvec q1 = find(bestIndices > 0);
max@1:             
max@1:             for (unsigned i=0; i<q1.size();++i)
max@1:                 newPart.indices.push_back(q1(i));
max@1:             
max@1:             newPart.letter = partletters[kSeg];
max@1:             newPart.value = partvalues[kSeg];
max@1:             newPart.level = kSeg+1;
max@1:             parts.push_back(newPart);
max@1:             
max@1:             uvec q2 = find(valid_sets==1);
max@1:             
max@1:             for (unsigned i=0; i<q2.size(); ++i)
max@1:             {
max@1:                 unsigned iSet = q2(i);
max@1:                 unsigned s = partlengths(bestval2(iSet,1));
max@1:                 
max@1:                 arma::rowvec mask1 = arma::zeros<arma::rowvec>(s*2-1);
max@1:                 for (unsigned i=0; i<s; ++i)
max@1:                     mask1(i+s-1) = 1;
max@1:                 
max@1:                 arma::rowvec Ind = decisionArray2.slice(bestval2(iSet,1))(bestval2(iSet,2),span::all);
max@1:                 arma::rowvec IndMap = arma::zeros<arma::rowvec>(Ind.size());
max@1:                 for (unsigned i=0; i<Ind.size(); ++i)
max@1:                     if (Ind(i)>0)
max@1:                         IndMap(i) = 2;
max@1:                 
max@1:                 arma::rowvec t3 = arma::conv(IndMap,mask1); 
max@6:                 arma::rowvec currislands = t3.subvec(mask1.size()/2,t3.size()-1-mask1.size()/2);       
max@1:                 arma::rowvec islandsdMult = currislands%island;
max@6:                 
max@1:                 arma::uvec islandsIndex = find(islandsdMult > 0);
max@1:                 
max@6:                 if (islandsIndex.size() > 0)
max@1:                     valid_sets(iSet) = 0;
max@1:             }
max@1:         }
max@1:     }
max@1:     else
max@1:     {
max@1:         Part newPart;
max@1:         newPart.n = nBeat;
max@1:         newPart.indices.push_back(1);
max@1:         newPart.letter = 'A';
max@1:         newPart.value = 1;
max@1:         newPart.level = 1;
max@1:         parts.push_back(newPart);
max@1:     }
max@6:    
max@1:     arma::vec bar = linspace(1,nBeat,nBeat);    
max@1:     Part np = nullpart(parts,bar);
max@7:     
max@1:     parts.push_back(np);
max@1:     
max@1:     // -------------- NOT CONVERTED -------------------------------------  
max@1:     // if param.seg.editor
max@1:     //    [pa, ta] = partarray(parts);
max@1:     //    parts = editorssearch(pa, ta, parts);
max@1:     //    parts = [parts, nullpart(parts,1:nBeat)];
max@1:     // end
max@1:     // ------------------------------------------------------------------
max@1: 
max@1:     
max@1:     mergenulls(parts);
max@1:     
max@1:     
max@1:     // -------------- NOT CONVERTED -------------------------------------  
max@1:     // if param.seg.editor
max@1:     //    [pa, ta] = partarray(parts);
max@1:     //    parts = editorssearch(pa, ta, parts);
max@1:     //    parts = [parts, nullpart(parts,1:nBeat)];
max@1:     // end
max@1:     // ------------------------------------------------------------------
max@1:     
max@1:     return parts;
max@1: }
max@1: 
max@1: 
max@1: 
max@1: void songSegmentChroma(Vamp::Plugin::FeatureList quatisedChromagram, vector<Part> &parts)
max@1: {
max@1:     // Collect Info
max@1:     int nBeat = quatisedChromagram.size();                      // Number of feature vector
max@1:     int nFeatValues = quatisedChromagram[0].values.size();      // Number of values for each feature vector
max@1: 
max@1:     arma::mat synchTreble = arma::zeros<mat>(nBeat,nFeatValues/2);
max@1:     
max@1:     for (unsigned i = 0; i < nBeat; ++ i)
max@1:         for (unsigned j = 0; j < nFeatValues/2; ++ j)
max@1:         {
max@1:             synchTreble(i,j) = quatisedChromagram[i].values[j];
max@1:         }
max@1:     
max@1:     arma::mat synchBass = arma::zeros<mat>(nBeat,nFeatValues/2);
max@1:     
max@1:     for (unsigned i = 0; i < nBeat; ++ i)
max@1:         for (unsigned j = 0; j < nFeatValues/2; ++ j)
max@1:         {
max@1:             synchBass(i,j) = quatisedChromagram[i].values[j+12];
max@1:         }
max@1: 
max@1:     // Process
max@1:     
max@1:     arma::mat segTreble = arma::zeros<arma::mat>(quatisedChromagram.size(),quatisedChromagram[0].values.size()/2);
max@1:     arma::mat segBass = arma::zeros<arma::mat>(quatisedChromagram.size(),quatisedChromagram[0].values.size()/2);
max@1:     
max@1:     for (unsigned iPart=0; iPart<parts.size(); ++iPart)
max@1:     {
max@1:         parts[iPart].nInd = parts[iPart].indices.size();
max@1:         
max@1:         for (unsigned kOccur=0; kOccur<parts[iPart].nInd; ++kOccur)
max@1:         {
max@1:             int kStartIndex = parts[iPart].indices[kOccur];
max@1:             int kEndIndex = kStartIndex + parts[iPart].n-1;
max@1:             
max@1:             segTreble.rows(kStartIndex,kEndIndex) = segTreble.rows(kStartIndex,kEndIndex) + synchTreble.rows(kStartIndex,kEndIndex);
max@1:             segBass.rows(kStartIndex,kEndIndex) = segBass.rows(kStartIndex,kEndIndex) + synchBass.rows(kStartIndex,kEndIndex);
max@1:         }
max@1:     }
max@1: }
max@1: 
max@1: 
max@1: // Segment Integration
max@1: vector<Part> songSegmentIntegration(vector<Part> &parts)
max@1: {
max@1:     // Break up parts (every part will have one instance)
max@1:     vector<Part> newPartVector;
max@1:     vector<int> partindices;
max@1:     
max@1:     for (unsigned iPart=0; iPart<parts.size(); ++iPart)
max@1:     {
max@1:         parts[iPart].nInd = parts[iPart].indices.size();
max@1:         for (unsigned iInstance=0; iInstance<parts[iPart].nInd; ++iInstance)
max@1:         {
max@1:             Part newPart;
max@1:             newPart.n = parts[iPart].n;
max@1:             newPart.letter = parts[iPart].letter;
max@1:             newPart.value = parts[iPart].value;
max@1:             newPart.level = parts[iPart].level;
max@1:             newPart.indices.push_back(parts[iPart].indices[iInstance]);
max@1:             newPart.nInd = 1;
max@1:             partindices.push_back(parts[iPart].indices[iInstance]);
max@1:             
max@1:             newPartVector.push_back(newPart);
max@1:         }
max@1:     }
max@1:     
max@1:     
max@1:     // Sort the parts in order of occurrence
max@1:     sort (partindices.begin(), partindices.end());
max@1:     
max@1:     for (unsigned i=0; i<partindices.size(); ++i)
max@1:     {
max@1:         bool found = false;
max@1:         int in=0;    
max@1:         while (!found)
max@1:         {
max@1:             if (newPartVector[in].indices[0] == partindices[i])
max@1:             {
max@1:                 newPartVector.push_back(newPartVector[in]);
max@1:                 newPartVector.erase(newPartVector.begin()+in);
max@1:                 found = true;
max@1:             }
max@1:             else
max@1:                 in++;
max@1:         }  
max@1:     }
max@1:     
max@1:     // Clear the vector
max@1:     for (unsigned iNewpart=1; iNewpart < newPartVector.size(); ++iNewpart)
max@1:     {
max@1:         if (newPartVector[iNewpart].n < 12)
max@1:         {
max@1:             newPartVector[iNewpart-1].n = newPartVector[iNewpart-1].n + newPartVector[iNewpart].n;
max@1:             newPartVector.erase(newPartVector.begin()+iNewpart);
max@1:         }
max@1:     }
max@1: 
max@1:     return newPartVector;
max@1: }
max@1: 
max@1: // Segmenter
max@1: Vamp::Plugin::FeatureList SongPartitioner::Segmenter(Vamp::Plugin::FeatureList quatisedChromagram)
max@1: {
max@1:     /* --- Display Information --- */
max@1:     int numBeat = quatisedChromagram.size();
max@1:     int numFeats = quatisedChromagram[0].values.size();
max@1: 
max@1:     vector<Part> parts;
max@1:     vector<Part> finalParts;
max@1:     
max@1:     parts = songSegment(quatisedChromagram);
max@1:     songSegmentChroma(quatisedChromagram,parts);
max@7:     
max@1:     finalParts = songSegmentIntegration(parts);
max@1:     
max@1:     
max@1:     // TEMP ----
max@6:     /*for (unsigned i=0;i<finalParts.size(); ++i)
max@1:      {
max@6:      std::cout << "Parts n° " << i << std::endl;
max@6:      std::cout << "n°: " << finalParts[i].n << std::endl;
max@6:      std::cout << "letter: " <<  finalParts[i].letter << std::endl;
max@1:      
max@6:      std::cout << "indices: ";
max@6:      for (unsigned j=0;j<finalParts[i].indices.size(); ++j)
max@6:          std::cout << finalParts[i].indices[j] << " ";
max@6:        
max@6:      std::cout << std::endl;
max@6:      std::cout <<  "level: " << finalParts[i].level << std::endl;
max@1:      }*/
max@1:     
max@1:     // ---------
max@1:     
max@1:     
max@1:     // Output
max@1: 
max@1:     Vamp::Plugin::FeatureList results;
max@1:     
max@1:     
max@1:     Feature seg;
max@1:     
max@1:     arma::vec indices;
max@1:     unsigned idx=0;
max@1:     vector<int> values;
max@1:     vector<string> letters;
max@1:     
max@1:     for (unsigned iPart=0; iPart<finalParts.size()-1; ++iPart)
max@1:     {
max@1:         unsigned iInstance=0;
max@1:         seg.hasTimestamp = true;
max@1:          
max@1:         int ind = finalParts[iPart].indices[iInstance];
max@1:         int ind1 = finalParts[iPart+1].indices[iInstance];
max@1:          
max@1:         seg.timestamp = quatisedChromagram[ind].timestamp;
max@1:         seg.hasDuration = true;
max@1:         seg.duration = quatisedChromagram[ind1].timestamp-quatisedChromagram[ind].timestamp;
max@1:         seg.values.clear();
max@1:         seg.values.push_back(finalParts[iPart].value);
max@1:         seg.label = finalParts[iPart].letter;
max@1:          
max@1:         results.push_back(seg);
max@1:     }
max@1:     
max@1:     int ind = finalParts[finalParts.size()-1].indices[0];
max@1:     seg.timestamp = quatisedChromagram[ind].timestamp;
max@1:     seg.hasDuration = true;
max@1:     seg.duration = quatisedChromagram[quatisedChromagram.size()-1].timestamp-quatisedChromagram[ind].timestamp;
max@1:     seg.values.clear();
max@1:     seg.values.push_back(finalParts[finalParts.size()-1].value);
max@1:     seg.label = finalParts[finalParts.size()-1].letter;
max@1:     
max@1:     results.push_back(seg);
max@1: 
max@1:     return results;    
max@1: }
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: 
max@1: