Chris@3: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
Chris@3: /*
Chris@31:     This file is Copyright (c) 2012 Chris Cannam
Chris@31:   
Chris@3:     Permission is hereby granted, free of charge, to any person
Chris@3:     obtaining a copy of this software and associated documentation
Chris@3:     files (the "Software"), to deal in the Software without
Chris@3:     restriction, including without limitation the rights to use, copy,
Chris@3:     modify, merge, publish, distribute, sublicense, and/or sell copies
Chris@3:     of the Software, and to permit persons to whom the Software is
Chris@3:     furnished to do so, subject to the following conditions:
Chris@3: 
Chris@3:     The above copyright notice and this permission notice shall be
Chris@3:     included in all copies or substantial portions of the Software.
Chris@3: 
Chris@3:     THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
Chris@3:     EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
Chris@3:     MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
Chris@3:     NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR
Chris@3:     ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF
Chris@3:     CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
Chris@3:     WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
Chris@3: */
Chris@3: 
Chris@31: #include "CepstralPitchTracker.h"
Chris@47: #include "MeanFilter.h"
Chris@3: 
Chris@26: #include "vamp-sdk/FFT.h"
Chris@26: 
Chris@3: #include <vector>
Chris@3: #include <algorithm>
Chris@3: 
Chris@3: #include <cstdio>
Chris@3: #include <cmath>
Chris@3: #include <complex>
Chris@3: 
Chris@3: using std::string;
Chris@7: using std::vector;
Chris@16: using Vamp::RealTime;
Chris@7: 
Chris@16: 
Chris@31: CepstralPitchTracker::CepstralPitchTracker(float inputSampleRate) :
Chris@3:     Plugin(inputSampleRate),
Chris@3:     m_channels(0),
Chris@3:     m_stepSize(256),
Chris@3:     m_blockSize(1024),
Chris@3:     m_fmin(50),
Chris@25:     m_fmax(900),
Chris@18:     m_vflen(1),
Chris@3:     m_binFrom(0),
Chris@3:     m_binTo(0),
Chris@15:     m_bins(0)
Chris@3: {
Chris@3: }
Chris@3: 
Chris@31: CepstralPitchTracker::~CepstralPitchTracker()
Chris@3: {
Chris@3: }
Chris@3: 
Chris@3: string
Chris@31: CepstralPitchTracker::getIdentifier() const
Chris@3: {
Chris@39:     return "cepstral-pitchtracker";
Chris@3: }
Chris@3: 
Chris@3: string
Chris@31: CepstralPitchTracker::getName() const
Chris@3: {
Chris@39:     return "Cepstral Pitch Tracker";
Chris@3: }
Chris@3: 
Chris@3: string
Chris@31: CepstralPitchTracker::getDescription() const
Chris@3: {
Chris@3:     return "Estimate f0 of monophonic material using a cepstrum method.";
Chris@3: }
Chris@3: 
Chris@3: string
Chris@31: CepstralPitchTracker::getMaker() const
Chris@3: {
Chris@3:     return "Chris Cannam";
Chris@3: }
Chris@3: 
Chris@3: int
Chris@31: CepstralPitchTracker::getPluginVersion() const
Chris@3: {
Chris@3:     // Increment this each time you release a version that behaves
Chris@3:     // differently from the previous one
Chris@3:     return 1;
Chris@3: }
Chris@3: 
Chris@3: string
Chris@31: CepstralPitchTracker::getCopyright() const
Chris@3: {
Chris@3:     return "Freely redistributable (BSD license)";
Chris@3: }
Chris@3: 
Chris@31: CepstralPitchTracker::InputDomain
Chris@31: CepstralPitchTracker::getInputDomain() const
Chris@3: {
Chris@3:     return FrequencyDomain;
Chris@3: }
Chris@3: 
Chris@3: size_t
Chris@31: CepstralPitchTracker::getPreferredBlockSize() const
Chris@3: {
Chris@3:     return 1024;
Chris@3: }
Chris@3: 
Chris@3: size_t 
Chris@31: CepstralPitchTracker::getPreferredStepSize() const
Chris@3: {
Chris@3:     return 256;
Chris@3: }
Chris@3: 
Chris@3: size_t
Chris@31: CepstralPitchTracker::getMinChannelCount() const
Chris@3: {
Chris@3:     return 1;
Chris@3: }
Chris@3: 
Chris@3: size_t
Chris@31: CepstralPitchTracker::getMaxChannelCount() const
Chris@3: {
Chris@3:     return 1;
Chris@3: }
Chris@3: 
Chris@31: CepstralPitchTracker::ParameterList
Chris@31: CepstralPitchTracker::getParameterDescriptors() const
Chris@3: {
Chris@3:     ParameterList list;
Chris@3:     return list;
Chris@3: }
Chris@3: 
Chris@3: float
Chris@31: CepstralPitchTracker::getParameter(string identifier) const
Chris@3: {
Chris@3:     return 0.f;
Chris@3: }
Chris@3: 
Chris@3: void
Chris@31: CepstralPitchTracker::setParameter(string identifier, float value) 
Chris@3: {
Chris@3: }
Chris@3: 
Chris@31: CepstralPitchTracker::ProgramList
Chris@31: CepstralPitchTracker::getPrograms() const
Chris@3: {
Chris@3:     ProgramList list;
Chris@3:     return list;
Chris@3: }
Chris@3: 
Chris@3: string
Chris@31: CepstralPitchTracker::getCurrentProgram() const
Chris@3: {
Chris@3:     return ""; // no programs
Chris@3: }
Chris@3: 
Chris@3: void
Chris@31: CepstralPitchTracker::selectProgram(string name)
Chris@3: {
Chris@3: }
Chris@3: 
Chris@31: CepstralPitchTracker::OutputList
Chris@31: CepstralPitchTracker::getOutputDescriptors() const
Chris@3: {
Chris@3:     OutputList outputs;
Chris@3: 
Chris@3:     OutputDescriptor d;
Chris@3: 
Chris@3:     d.identifier = "f0";
Chris@3:     d.name = "Estimated f0";
Chris@3:     d.description = "Estimated fundamental frequency";
Chris@3:     d.unit = "Hz";
Chris@3:     d.hasFixedBinCount = true;
Chris@3:     d.binCount = 1;
Chris@3:     d.hasKnownExtents = true;
Chris@3:     d.minValue = m_fmin;
Chris@3:     d.maxValue = m_fmax;
Chris@3:     d.isQuantized = false;
Chris@3:     d.sampleType = OutputDescriptor::FixedSampleRate;
Chris@3:     d.sampleRate = (m_inputSampleRate / m_stepSize);
Chris@3:     d.hasDuration = false;
Chris@3:     outputs.push_back(d);
Chris@3: 
Chris@16:     d.identifier = "notes";
Chris@16:     d.name = "Notes";
Chris@16:     d.description = "Derived fixed-pitch note frequencies";
Chris@16:     d.unit = "Hz";
Chris@16:     d.hasFixedBinCount = true;
Chris@16:     d.binCount = 1;
Chris@16:     d.hasKnownExtents = true;
Chris@16:     d.minValue = m_fmin;
Chris@16:     d.maxValue = m_fmax;
Chris@16:     d.isQuantized = false;
Chris@16:     d.sampleType = OutputDescriptor::FixedSampleRate;
Chris@16:     d.sampleRate = (m_inputSampleRate / m_stepSize);
Chris@16:     d.hasDuration = true;
Chris@16:     outputs.push_back(d);
Chris@16: 
Chris@3:     return outputs;
Chris@3: }
Chris@3: 
Chris@3: bool
Chris@31: CepstralPitchTracker::initialise(size_t channels, size_t stepSize, size_t blockSize)
Chris@3: {
Chris@3:     if (channels < getMinChannelCount() ||
Chris@3: 	channels > getMaxChannelCount()) return false;
Chris@3: 
Chris@31: //    std::cerr << "CepstralPitchTracker::initialise: channels = " << channels
Chris@3: //	      << ", stepSize = " << stepSize << ", blockSize = " << blockSize
Chris@3: //	      << std::endl;
Chris@3: 
Chris@3:     m_channels = channels;
Chris@3:     m_stepSize = stepSize;
Chris@3:     m_blockSize = blockSize;
Chris@3: 
Chris@3:     m_binFrom = int(m_inputSampleRate / m_fmax);
Chris@3:     m_binTo = int(m_inputSampleRate / m_fmin); 
Chris@3: 
Chris@3:     if (m_binTo >= (int)m_blockSize / 2) {
Chris@3:         m_binTo = m_blockSize / 2 - 1;
Chris@3:     }
Chris@3: 
Chris@3:     m_bins = (m_binTo - m_binFrom) + 1;
Chris@3: 
Chris@3:     reset();
Chris@3: 
Chris@3:     return true;
Chris@3: }
Chris@3: 
Chris@3: void
Chris@31: CepstralPitchTracker::reset()
Chris@3: {
Chris@3: }
Chris@3: 
Chris@3: void
Chris@35: CepstralPitchTracker::addFeaturesFrom(NoteHypothesis h, FeatureSet &fs)
Chris@30: {
Chris@35:     NoteHypothesis::Estimates es = h.getAcceptedEstimates();
Chris@30: 
Chris@35:     for (int i = 0; i < (int)es.size(); ++i) {
Chris@30: 	Feature f;
Chris@30: 	f.hasTimestamp = true;
Chris@30: 	f.timestamp = es[i].time;
Chris@30: 	f.values.push_back(es[i].freq);
Chris@30: 	fs[0].push_back(f);
Chris@30:     }
Chris@30: 
Chris@30:     Feature nf;
Chris@30:     nf.hasTimestamp = true;
Chris@30:     nf.hasDuration = true;
Chris@35:     NoteHypothesis::Note n = h.getAveragedNote();
Chris@30:     nf.timestamp = n.time;
Chris@30:     nf.duration = n.duration;
Chris@30:     nf.values.push_back(n.freq);
Chris@30:     fs[1].push_back(nf);
Chris@30: }
Chris@30: 
Chris@18: double
Chris@31: CepstralPitchTracker::cubicInterpolate(const double y[4], double x)
Chris@18: {
Chris@18:     double a0 = y[3] - y[2] - y[0] + y[1];
Chris@18:     double a1 = y[0] - y[1] - a0;
Chris@18:     double a2 = y[2] - y[0];
Chris@18:     double a3 = y[1];
Chris@18:     return
Chris@18:         a0 * x * x * x +
Chris@18:         a1 * x * x +
Chris@18:         a2 * x +
Chris@18:         a3;
Chris@18: }
Chris@18: 
Chris@18: double
Chris@31: CepstralPitchTracker::findInterpolatedPeak(const double *in, int maxbin)
Chris@18: {
Chris@18:     if (maxbin < 2 || maxbin > m_bins - 3) {
Chris@18:         return maxbin;
Chris@18:     }
Chris@18: 
Chris@18:     double maxval = 0.0;
Chris@18:     double maxidx = maxbin;
Chris@18: 
Chris@18:     const int divisions = 10;
Chris@18:     double y[4];
Chris@18: 
Chris@18:     y[0] = in[maxbin-1];
Chris@18:     y[1] = in[maxbin];
Chris@18:     y[2] = in[maxbin+1];
Chris@18:     y[3] = in[maxbin+2];
Chris@18:     for (int i = 0; i < divisions; ++i) {
Chris@18:         double probe = double(i) / double(divisions);
Chris@18:         double value = cubicInterpolate(y, probe);
Chris@18:         if (value > maxval) {
Chris@18:             maxval = value; 
Chris@18:             maxidx = maxbin + probe;
Chris@18:         }
Chris@18:     }
Chris@18: 
Chris@18:     y[3] = y[2];
Chris@18:     y[2] = y[1];
Chris@18:     y[1] = y[0];
Chris@18:     y[0] = in[maxbin-2];
Chris@18:     for (int i = 0; i < divisions; ++i) {
Chris@18:         double probe = double(i) / double(divisions);
Chris@18:         double value = cubicInterpolate(y, probe);
Chris@18:         if (value > maxval) {
Chris@18:             maxval = value; 
Chris@18:             maxidx = maxbin - 1 + probe;
Chris@18:         }
Chris@18:     }
Chris@18: 
Chris@18: /*
Chris@18:     std::cerr << "centre = " << maxbin << ": ["
Chris@18:               << in[maxbin-2] << ","
Chris@18:               << in[maxbin-1] << ","
Chris@18:               << in[maxbin] << ","
Chris@18:               << in[maxbin+1] << ","
Chris@18:               << in[maxbin+2] << "] -> " << maxidx << std::endl;
Chris@18: */
Chris@18: 
Chris@18:     return maxidx;
Chris@18: }
Chris@18: 
Chris@31: CepstralPitchTracker::FeatureSet
Chris@31: CepstralPitchTracker::process(const float *const *inputBuffers, RealTime timestamp)
Chris@3: {
Chris@3:     FeatureSet fs;
Chris@3: 
Chris@3:     int bs = m_blockSize;
Chris@3:     int hs = m_blockSize/2 + 1;
Chris@3: 
Chris@3:     double *rawcep = new double[bs];
Chris@3:     double *io = new double[bs];
Chris@3:     double *logmag = new double[bs];
Chris@3: 
Chris@4:     // The "inverse symmetric" method. Seems to be the most reliable
Chris@3:         
Chris@25:     double magmean = 0.0;
Chris@25: 
Chris@3:     for (int i = 0; i < hs; ++i) {
Chris@3: 
Chris@3: 	double power =
Chris@3: 	    inputBuffers[0][i*2  ] * inputBuffers[0][i*2  ] +
Chris@3: 	    inputBuffers[0][i*2+1] * inputBuffers[0][i*2+1];
Chris@3: 	double mag = sqrt(power);
Chris@25: 
Chris@25:         magmean += mag;
Chris@25: 
Chris@3: 	double lm = log(mag + 0.00000001);
Chris@3: 	
Chris@4: 	logmag[i] = lm;
Chris@4: 	if (i > 0) logmag[bs - i] = lm;
Chris@3:     }
Chris@3: 
Chris@25:     magmean /= hs;
Chris@25:     double threshold = 0.1; // for magmean
Chris@25:     
Chris@26:     Vamp::FFT::inverse(bs, logmag, 0, rawcep, io);
Chris@3:     
Chris@3:     delete[] logmag;
Chris@3:     delete[] io;
Chris@3: 
Chris@3:     int n = m_bins;
Chris@3:     double *data = new double[n];
Chris@47:     MeanFilter(m_vflen).filterSubsequence(rawcep, data, m_blockSize, n, m_binFrom);
Chris@3:     delete[] rawcep;
Chris@3: 
Chris@3:     double maxval = 0.0;
Chris@6:     int maxbin = -1;
Chris@3: 
Chris@3:     for (int i = 0; i < n; ++i) {
Chris@3:         if (data[i] > maxval) {
Chris@3:             maxval = data[i];
Chris@3:             maxbin = i;
Chris@3:         }
Chris@3:     }
Chris@3: 
Chris@15:     if (maxbin < 0) {
Chris@15:         delete[] data;
Chris@15:         return fs;
Chris@15:     }
Chris@15: 
Chris@15:     double nextPeakVal = 0.0;
Chris@15:     for (int i = 1; i+1 < n; ++i) {
Chris@15:         if (data[i] > data[i-1] &&
Chris@15:             data[i] > data[i+1] &&
Chris@15:             i != maxbin &&
Chris@15:             data[i] > nextPeakVal) {
Chris@15:             nextPeakVal = data[i];
Chris@15:         }
Chris@15:     }
Chris@8: 
Chris@18:     double cimax = findInterpolatedPeak(data, maxbin);
Chris@18:     double peakfreq = m_inputSampleRate / (cimax + m_binFrom);
Chris@15: 
Chris@15:     double confidence = 0.0;
Chris@15:     if (nextPeakVal != 0.0) {
Chris@27:         confidence = (maxval - nextPeakVal) * 10.0;
Chris@25:         if (magmean < threshold) confidence = 0.0;
Chris@39: //        std::cerr << "magmean = " << magmean << ", confidence = " << confidence << std::endl;
Chris@15:     }
Chris@15: 
Chris@35:     NoteHypothesis::Estimate e;
Chris@8:     e.freq = peakfreq;
Chris@8:     e.time = timestamp;
Chris@15:     e.confidence = confidence;
Chris@8: 
Chris@28:     if (!m_good.accept(e)) {
Chris@13: 
Chris@11:         int candidate = -1;
Chris@13:         bool accepted = false;
Chris@13: 
Chris@35:         for (int i = 0; i < (int)m_possible.size(); ++i) {
Chris@28:             if (m_possible[i].accept(e)) {
Chris@35:                 if (m_possible[i].getState() == NoteHypothesis::Satisfied) {
Chris@28:                     accepted = true;
Chris@11:                     candidate = i;
Chris@11:                 }
Chris@11:                 break;
Chris@11:             }
Chris@11:         }
Chris@12: 
Chris@13:         if (!accepted) {
Chris@35:             NoteHypothesis h;
Chris@28:             h.accept(e); //!!! must succeed as h is new, so perhaps there should be a ctor for this
Chris@13:             m_possible.push_back(h);
Chris@13:         }
Chris@13: 
Chris@35:         if (m_good.getState() == NoteHypothesis::Expired) {
Chris@30:             addFeaturesFrom(m_good, fs);
Chris@12:         }
Chris@12:         
Chris@35:         if (m_good.getState() == NoteHypothesis::Expired ||
Chris@35:             m_good.getState() == NoteHypothesis::Rejected) {
Chris@11:             if (candidate >= 0) {
Chris@28:                 m_good = m_possible[candidate];
Chris@11:             } else {
Chris@35:                 m_good = NoteHypothesis();
Chris@11:             }
Chris@11:         }
Chris@8: 
Chris@14:         // reap rejected/expired hypotheses from possible list
Chris@14:         Hypotheses toReap = m_possible;
Chris@14:         m_possible.clear();
Chris@35:         for (int i = 0; i < (int)toReap.size(); ++i) {
Chris@35:             NoteHypothesis h = toReap[i];
Chris@35:             if (h.getState() != NoteHypothesis::Rejected && 
Chris@35:                 h.getState() != NoteHypothesis::Expired) {
Chris@14:                 m_possible.push_back(h);
Chris@14:             }
Chris@14:         }
Chris@14:     }  
Chris@14: 
Chris@3:     delete[] data;
Chris@3:     return fs;
Chris@3: }
Chris@3: 
Chris@31: CepstralPitchTracker::FeatureSet
Chris@31: CepstralPitchTracker::getRemainingFeatures()
Chris@3: {
Chris@3:     FeatureSet fs;
Chris@35:     if (m_good.getState() == NoteHypothesis::Satisfied) {
Chris@30:         addFeaturesFrom(m_good, fs);
Chris@11:     }
Chris@3:     return fs;
Chris@3: }