c@279: /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*-  vi:set ts=8 sts=4 sw=4: */
c@279: 
c@279: /*
c@279:     QM DSP Library
c@279: 
c@279:     Centre for Digital Music, Queen Mary, University of London.
c@279:     This file copyright 2008-2009 Matthew Davies and QMUL.
c@279:     All rights reserved.
c@279: */
c@279: 
c@279: #include "DownBeat.h"
c@279: 
c@279: #include "maths/MathAliases.h"
c@279: #include "maths/MathUtilities.h"
c@280: #include "maths/KLDivergence.h"
c@279: #include "dsp/transforms/FFT.h"
c@279: 
c@279: #include <iostream>
c@279: #include <cstdlib>
c@279: 
c@279: DownBeat::DownBeat(float originalSampleRate,
c@279:                    size_t decimationFactor,
c@279:                    size_t dfIncrement) :
c@280:     m_bpb(0),
c@279:     m_rate(originalSampleRate),
c@279:     m_factor(decimationFactor),
c@279:     m_increment(dfIncrement),
c@279:     m_decimator1(0),
c@279:     m_decimator2(0),
c@279:     m_buffer(0),
c@283:     m_decbuf(0),
c@279:     m_bufsiz(0),
c@279:     m_buffill(0),
c@279:     m_beatframesize(0),
c@279:     m_beatframe(0)
c@279: {
c@279:     // beat frame size is next power of two up from 1.3 seconds at the
c@279:     // downsampled rate (happens to produce 4096 for 44100 or 48000 at
c@279:     // 16x decimation, which is our expected normal situation)
c@280:     m_beatframesize = MathUtilities::nextPowerOfTwo
c@280:         (int((m_rate / decimationFactor) * 1.3));
c@282: //    std::cerr << "rate = " << m_rate << ", bfs = " << m_beatframesize << std::endl;
c@279:     m_beatframe = new double[m_beatframesize];
c@279:     m_fftRealOut = new double[m_beatframesize];
c@279:     m_fftImagOut = new double[m_beatframesize];
c@289:     m_fft = new FFTReal(m_beatframesize);
c@279: }
c@279: 
c@279: DownBeat::~DownBeat()
c@279: {
c@279:     delete m_decimator1;
c@279:     delete m_decimator2;
c@279:     if (m_buffer) free(m_buffer);
c@279:     delete[] m_decbuf;
c@279:     delete[] m_beatframe;
c@279:     delete[] m_fftRealOut;
c@279:     delete[] m_fftImagOut;
c@289:     delete m_fft;
c@279: }
c@279: 
c@279: void
c@280: DownBeat::setBeatsPerBar(int bpb)
c@280: {
c@280:     m_bpb = bpb;
c@280: }
c@280: 
c@280: void
c@279: DownBeat::makeDecimators()
c@279: {
c@283: //    std::cerr << "m_factor = " << m_factor << std::endl;
c@279:     if (m_factor < 2) return;
c@279:     int highest = Decimator::getHighestSupportedFactor();
c@279:     if (m_factor <= highest) {
c@279:         m_decimator1 = new Decimator(m_increment, m_factor);
c@282: //        std::cerr << "DownBeat: decimator 1 factor " << m_factor << ", size " << m_increment << std::endl;
c@279:         return;
c@279:     }
c@279:     m_decimator1 = new Decimator(m_increment, highest);
c@282: //    std::cerr << "DownBeat: decimator 1 factor " << highest << ", size " << m_increment << std::endl;
c@279:     m_decimator2 = new Decimator(m_increment / highest, m_factor / highest);
c@282: //    std::cerr << "DownBeat: decimator 2 factor " << m_factor / highest << ", size " << m_increment / highest << std::endl;
c@280:     m_decbuf = new float[m_increment / highest];
c@279: }
c@279: 
c@279: void
c@280: DownBeat::pushAudioBlock(const float *audio)
c@279: {
c@279:     if (m_buffill + (m_increment / m_factor) > m_bufsiz) {
c@279:         if (m_bufsiz == 0) m_bufsiz = m_increment * 16;
c@279:         else m_bufsiz = m_bufsiz * 2;
c@279:         if (!m_buffer) {
c@280:             m_buffer = (float *)malloc(m_bufsiz * sizeof(float));
c@279:         } else {
c@282: //            std::cerr << "DownBeat::pushAudioBlock: realloc m_buffer to " << m_bufsiz << std::endl;
c@280:             m_buffer = (float *)realloc(m_buffer, m_bufsiz * sizeof(float));
c@279:         }
c@279:     }
c@283:     if (!m_decimator1 && m_factor > 1) makeDecimators();
c@283: //    float rmsin = 0, rmsout = 0;
c@283: //    for (int i = 0; i < m_increment; ++i) {
c@283: //        rmsin += audio[i] * audio[i];
c@283: //    }
c@279:     if (m_decimator2) {
c@279:         m_decimator1->process(audio, m_decbuf);
c@279:         m_decimator2->process(m_decbuf, m_buffer + m_buffill);
c@283:     } else if (m_decimator1) {
c@283:         m_decimator1->process(audio, m_buffer + m_buffill);
c@279:     } else {
c@283:         // just copy across (m_factor is presumably 1)
c@283:         for (int i = 0; i < m_increment; ++i) {
c@283:             (m_buffer + m_buffill)[i] = audio[i];
c@283:         }
c@279:     }
c@283: //    for (int i = 0; i < m_increment / m_factor; ++i) {
c@283: //        rmsout += m_buffer[m_buffill + i] * m_buffer[m_buffill + i];
c@283: //    }
c@282: //    std::cerr << "pushAudioBlock: rms in " << sqrt(rmsin) << ", out " << sqrt(rmsout) << std::endl;
c@279:     m_buffill += m_increment / m_factor;
c@279: }
c@279:     
c@280: const float *
c@279: DownBeat::getBufferedAudio(size_t &length) const
c@279: {
c@279:     length = m_buffill;
c@279:     return m_buffer;
c@279: }
c@279: 
c@279: void
c@280: DownBeat::resetAudioBuffer()
c@280: {
c@280:     if (m_buffer) free(m_buffer);
c@283:     m_buffer = 0;
c@280:     m_buffill = 0;
c@280:     m_bufsiz = 0;
c@280: }
c@280: 
c@280: void
c@280: DownBeat::findDownBeats(const float *audio,
c@279:                         size_t audioLength,
c@279:                         const d_vec_t &beats,
c@279:                         i_vec_t &downbeats)
c@279: {
c@279:     // FIND DOWNBEATS BY PARTITIONING THE INPUT AUDIO FILE INTO BEAT SEGMENTS
c@279:     // WHERE THE AUDIO FRAMES ARE DOWNSAMPLED  BY A FACTOR OF 16 (fs ~= 2700Hz)
c@279:     // THEN TAKING THE JENSEN-SHANNON DIVERGENCE BETWEEN BEAT SYNCHRONOUS SPECTRAL FRAMES
c@279: 
c@279:     // IMPLEMENTATION (MOSTLY) FOLLOWS:
c@279:     //  DAVIES AND PLUMBLEY "A SPECTRAL DIFFERENCE APPROACH TO EXTRACTING DOWNBEATS IN MUSICAL AUDIO"
c@279:     //  EUSIPCO 2006, FLORENCE, ITALY
c@279: 
c@279:     d_vec_t newspec(m_beatframesize / 2); // magnitude spectrum of current beat
c@279:     d_vec_t oldspec(m_beatframesize / 2); // magnitude spectrum of previous beat
c@281: 
c@281:     m_beatsd.clear();
c@279: 
c@279:     if (audioLength == 0) return;
c@279: 
c@279:     for (size_t i = 0; i + 1 < beats.size(); ++i) {
c@279: 
c@279:         // Copy the extents of the current beat from downsampled array
c@279:         // into beat frame buffer
c@279: 
c@279:         size_t beatstart = (beats[i] * m_increment) / m_factor;
c@280:         size_t beatend = (beats[i+1] * m_increment) / m_factor;
c@279:         if (beatend >= audioLength) beatend = audioLength - 1;
c@279:         if (beatend < beatstart) beatend = beatstart;
c@279:         size_t beatlen = beatend - beatstart;
c@279: 
c@279:         // Also apply a Hanning window to the beat frame buffer, sized
c@279:         // to the beat extents rather than the frame size.  (Because
c@279:         // the size varies, it's easier to do this by hand than use
c@279:         // our Window abstraction.)
c@279: 
c@283: //        std::cerr << "beatlen = " << beatlen << std::endl;
c@283: 
c@283: //        float rms = 0;
c@283:         for (size_t j = 0; j < beatlen && j < m_beatframesize; ++j) {
c@279:             double mul = 0.5 * (1.0 - cos(TWO_PI * (double(j) / double(beatlen))));
c@279:             m_beatframe[j] = audio[beatstart + j] * mul;
c@283: //            rms += m_beatframe[j] * m_beatframe[j];
c@279:         }
c@283: //        rms = sqrt(rms);
c@282: //        std::cerr << "beat " << i << ": audio rms " << rms << std::endl;
c@279: 
c@279:         for (size_t j = beatlen; j < m_beatframesize; ++j) {
c@279:             m_beatframe[j] = 0.0;
c@279:         }
c@279: 
c@279:         // Now FFT beat frame
c@279:         
c@289:         m_fft->process(false, m_beatframe, m_fftRealOut, m_fftImagOut);
c@279:         
c@279:         // Calculate magnitudes
c@279: 
c@279:         for (size_t j = 0; j < m_beatframesize/2; ++j) {
c@279:             newspec[j] = sqrt(m_fftRealOut[j] * m_fftRealOut[j] +
c@279:                               m_fftImagOut[j] * m_fftImagOut[j]);
c@279:         }
c@279: 
c@279:         // Preserve peaks by applying adaptive threshold
c@279: 
c@279:         MathUtilities::adaptiveThreshold(newspec);
c@279: 
c@279:         // Calculate JS divergence between new and old spectral frames
c@279: 
c@281:         if (i > 0) { // otherwise we have no previous frame
c@281:             m_beatsd.push_back(measureSpecDiff(oldspec, newspec));
c@282: //            std::cerr << "specdiff: " << m_beatsd[m_beatsd.size()-1] << std::endl;
c@281:         }
c@279: 
c@279:         // Copy newspec across to old
c@279: 
c@279:         for (size_t j = 0; j < m_beatframesize/2; ++j) {
c@279:             oldspec[j] = newspec[j];
c@279:         }
c@279:     }
c@279: 
c@279:     // We now have all spectral difference measures in specdiff
c@279: 
c@295:     unsigned int timesig = m_bpb;
c@280:     if (timesig == 0) timesig = 4;
c@280: 
c@279:     d_vec_t dbcand(timesig); // downbeat candidates
c@279: 
c@280:     for (int beat = 0; beat < timesig; ++beat) {
c@280:         dbcand[beat] = 0;
c@280:     }
c@280: 
c@279:     // look for beat transition which leads to greatest spectral change
c@279:     for (int beat = 0; beat < timesig; ++beat) {
c@281:         int count = 0;
c@281:         for (int example = beat - 1; example < m_beatsd.size(); example += timesig) {
c@281:             if (example < 0) continue;
c@281:             dbcand[beat] += (m_beatsd[example]) / timesig;
c@281:             ++count;
c@279:         }
c@281:         if (count > 0) m_beatsd[beat] /= count;
c@282: //        std::cerr << "dbcand[" << beat << "] = " << dbcand[beat] << std::endl;
c@279:     }
c@279: 
c@280: 
c@279:     // first downbeat is beat at index of maximum value of dbcand
c@279:     int dbind = MathUtilities::getMax(dbcand);
c@279: 
c@279:     // remaining downbeats are at timesig intervals from the first
c@279:     for (int i = dbind; i < beats.size(); i += timesig) {
c@279:         downbeats.push_back(i);
c@279:     }
c@279: }
c@279: 
c@279: double
c@279: DownBeat::measureSpecDiff(d_vec_t oldspec, d_vec_t newspec)
c@279: {
c@279:     // JENSEN-SHANNON DIVERGENCE BETWEEN SPECTRAL FRAMES
c@279: 
c@295:     unsigned int SPECSIZE = 512;   // ONLY LOOK AT FIRST 512 SAMPLES OF SPECTRUM. 
c@279:     if (SPECSIZE > oldspec.size()/4) {
c@279:         SPECSIZE = oldspec.size()/4;
c@279:     }
c@279:     double SD = 0.;
c@279:     double sd1 = 0.;
c@279: 
c@279:     double sumnew = 0.;
c@279:     double sumold = 0.;
c@279:   
c@295:     for (unsigned int i = 0;i < SPECSIZE;i++)
c@279:     {
c@279:         newspec[i] +=EPS;
c@279:         oldspec[i] +=EPS;
c@279:         
c@279:         sumnew+=newspec[i];
c@279:         sumold+=oldspec[i];
c@279:     } 
c@279:     
c@295:     for (unsigned int i = 0;i < SPECSIZE;i++)
c@279:     {
c@279:         newspec[i] /= (sumnew);
c@279:         oldspec[i] /= (sumold);
c@279:         
c@279:         // IF ANY SPECTRAL VALUES ARE 0 (SHOULDN'T BE ANY!) SET THEM TO 1
c@279:         if (newspec[i] == 0)
c@279:         {
c@279:             newspec[i] = 1.;
c@279:         }
c@279:         
c@279:         if (oldspec[i] == 0)
c@279:         {
c@279:             oldspec[i] = 1.;
c@279:         }
c@279:         
c@279:         // JENSEN-SHANNON CALCULATION
c@279:         sd1 = 0.5*oldspec[i] + 0.5*newspec[i];	
c@279:         SD = SD + (-sd1*log(sd1)) + (0.5*(oldspec[i]*log(oldspec[i]))) + (0.5*(newspec[i]*log(newspec[i])));
c@279:     }
c@279:     
c@279:     return SD;
c@279: }
c@279: 
c@281: void
c@281: DownBeat::getBeatSD(vector<double> &beatsd) const
c@281: {
c@281:     for (int i = 0; i < m_beatsd.size(); ++i) beatsd.push_back(m_beatsd[i]);
c@281: }
c@281: