comparison PYIN.cpp @ 91:854d9403c5be

note separation based on RMS

     m_oSmoothedPitchTrack(0),
     m_oNotes(0),
     m_threshDistr(2.0f),
     m_outputUnvoiced(0.0f),
     m_pitchProb(0),
-    m_timestamp(0)
+    m_timestamp(0),
+    m_level(0)
 {
 }
 
 PYIN::~PYIN()
 {

     m_yin.setThresholdDistr(m_threshDistr);
     m_yin.setFrameSize(m_blockSize);
     
     m_pitchProb.clear();
     m_timestamp.clear();
+    m_level.clear();
 /*    
     std::cerr << "PYIN::reset"
           << ", blockSize = " << m_blockSize
           << std::endl;
 */

     double *dInputBuffers = new double[m_blockSize];
     for (size_t i = 0; i < m_blockSize; ++i) dInputBuffers[i] = inputBuffers[0][i];
     
     Yin::YinOutput yo = m_yin.processProbabilisticYin(dInputBuffers);
     delete [] dInputBuffers;
+
+    m_level.push_back(yo.rms);
 
     // First, get the things out of the way that we don't want to output 
     // immediately, but instead save for later.
     vector<pair<double, double> > tempPitchProb;
     for (size_t iCandidate = 0; iCandidate < yo.freqProb.size(); ++iCandidate)

     size_t nFrame = m_pitchProb.size();
     
     std::vector<float> notePitchTrack; // collects pitches for one note at a time
     for (size_t iFrame = 0; iFrame < nFrame; ++iFrame)
     {
-        isVoiced = mnOut[iFrame].noteState < 3 && smoothedPitch[iFrame].size() > 0;
+        isVoiced = mnOut[iFrame].noteState < 3
+                   && smoothedPitch[iFrame].size() > 0
+                   && (iFrame == nFrame-1 
+                       || ((m_level[iFrame+1]/m_level[iFrame]) < 1.25));
+        // std::cerr << m_level[iFrame]/m_level[iFrame-1] << std::endl;
         if (isVoiced && iFrame != nFrame-1)
         {
             if (oldIsVoiced == 0) // beginning of a note
             {
                 onsetFrame = iFrame;
                 notePitchTrack.clear();
             }
             float pitch = smoothedPitch[iFrame][0].first;
             notePitchTrack.push_back(pitch); // add to the note's pitch track
         } else { // not currently voiced
-            if (oldIsVoiced == 1 && notePitchTrack.size() > 17) // end of note
+            if (oldIsVoiced == 1 && notePitchTrack.size() > 14) // end of note
             {
                 std::sort(notePitchTrack.begin(), notePitchTrack.end());
                 float medianPitch = notePitchTrack[notePitchTrack.size()/2];
                 float medianFreq = std::pow(2,(medianPitch - 69) / 12) * 440;
                 f.values.clear();