Mercurial > hg > nnls-chroma
changeset 82:e8ceb7d3bd4f matthiasm-plugin
merge
| author | Matthias Mauch <mail@matthiasmauch.net> |
|---|---|
| date | Mon, 15 Nov 2010 11:01:56 +0900 |
| parents | 4270f3039ab0 (diff) f50f701df33e (current diff) |
| children | 3277f6cef3fe fc28c50caace |
| files | NNLSChroma.cpp |
| diffstat | 7 files changed, 624 insertions(+), 601 deletions(-) [+] |
line wrap: on
line diff
--- a/Chordino.cpp Mon Nov 15 09:23:49 2010 +0900 +++ b/Chordino.cpp Mon Nov 15 11:01:56 2010 +0900 @@ -246,12 +246,16 @@ calculate tuning from (using the angle of the complex number defined by the cumulative mean real and imag values) **/ - float meanTuningImag = sinvalue * m_meanTuning1 - sinvalue * m_meanTuning2; - float meanTuningReal = m_meanTuning0 + cosvalue * m_meanTuning1 + cosvalue * m_meanTuning2; + float meanTuningImag = 0; + float meanTuningReal = 0; + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + meanTuningReal += m_meanTunings[iBPS] * cosvalues[iBPS]; + meanTuningImag += m_meanTunings[iBPS] * sinvalues[iBPS]; + } float cumulativetuning = 440 * pow(2,atan2(meanTuningImag, meanTuningReal)/(24*M_PI)); float normalisedtuning = atan2(meanTuningImag, meanTuningReal)/(2*M_PI); int intShift = floor(normalisedtuning * 3); - float intFactor = normalisedtuning * 3 - intShift; // intFactor is a really bad name for this + float floatShift = normalisedtuning * 3 - intShift; // floatShift is a really bad name for this char buffer0 [50]; @@ -285,13 +289,13 @@ if (m_tuneLocal) { intShift = floor(m_localTuning[count] * 3); - intFactor = m_localTuning[count] * 3 - intShift; // intFactor is a really bad name for this + floatShift = m_localTuning[count] * 3 - intShift; // floatShift is a really bad name for this } - // cerr << intShift << " " << intFactor << endl; + // cerr << intShift << " " << floatShift << endl; for (unsigned k = 2; k < currentLogSpectum.values.size() - 3; ++k) { // interpolate all inner bins - tempValue = currentLogSpectum.values[k + intShift] * (1-intFactor) + currentLogSpectum.values[k+intShift+1] * intFactor; + tempValue = currentLogSpectum.values[k + intShift] * (1-floatShift) + currentLogSpectum.values[k+intShift+1] * floatShift; currentTunedSpec.values.push_back(tempValue); } @@ -370,11 +374,11 @@ if (some_b_greater_zero) { if (m_useNNLS == 0) { - for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { + for (unsigned iNote = nBPS/2 + 2; iNote < nNote - nBPS/2; iNote += nBPS) { currval = 0; - currval += b[iNote + 1 + -1] * 0.5; - currval += b[iNote + 1 + 0] * 1.0; - currval += b[iNote + 1 + 1] * 0.5; + for (int iBPS = -nBPS/2; iBPS < nBPS/2+1; ++iBPS) { + currval += b[iNote + iBPS] * (1-abs(iBPS*1.0/(nBPS/2+1))); + } chroma[iSemitone % 12] += currval * treblewindow[iSemitone]; basschroma[iSemitone % 12] += currval * basswindow[iSemitone]; iSemitone++; @@ -386,11 +390,11 @@ vector<int> signifIndex; int index=0; sumb /= 84.0; - for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { + for (unsigned iNote = nBPS/2 + 2; iNote < nNote - nBPS/2; iNote += nBPS) { float currval = 0; - currval += b[iNote + 1 + -1]; - currval += b[iNote + 1 + 0]; - currval += b[iNote + 1 + 1]; + for (int iBPS = -nBPS/2; iBPS < nBPS/2+1; ++iBPS) { + currval += b[iNote + iBPS]; + } if (currval > 0) signifIndex.push_back(index); index++; } @@ -400,6 +404,7 @@ int indx[84+1000]; int mode; int dictsize = nNote*signifIndex.size(); + // cerr << "dictsize is " << dictsize << "and values size" << f3.values.size()<< endl; float *curr_dict = new float[dictsize]; for (unsigned iNote = 0; iNote < signifIndex.size(); ++iNote) { for (unsigned iBin = 0; iBin < nNote; iBin++) {
--- a/NNLSBase.cpp Mon Nov 15 09:23:49 2010 +0900 +++ b/NNLSBase.cpp Mon Nov 15 11:01:56 2010 +0900 @@ -34,12 +34,8 @@ m_blockSize(0), m_stepSize(0), m_lengthOfNoteIndex(0), - m_meanTuning0(0), - m_meanTuning1(0), - m_meanTuning2(0), - m_localTuning0(0), - m_localTuning1(0), - m_localTuning2(0), + m_meanTunings(0), + m_localTunings(0), m_whitening(1.0), m_preset(0.0), m_localTuning(0), @@ -54,7 +50,9 @@ m_rollon(0), m_s(0.7), m_useNNLS(1), - m_useHMM(1) + m_useHMM(1), + sinvalues(0), + cosvalues(0) { if (debug_on) cerr << "--> NNLSBase" << endl; @@ -352,6 +350,14 @@ cerr << "--> initialise"; } + // make things for tuning estimation + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + sinvalues.push_back(sin(2*M_PI*(iBPS*1.0/nBPS))); + cosvalues.push_back(cos(2*M_PI*(iBPS*1.0/nBPS))); + } + + + // make hamming window of length 1/2 octave int hamwinlength = nBPS * 6 + 1; float hamwinsum = 0; for (int i = 0; i < hamwinlength; ++i) { @@ -359,6 +365,13 @@ hamwinsum += 0.54 - 0.46 * cos((2*M_PI*i)/(hamwinlength-1)); } for (int i = 0; i < hamwinlength; ++i) hw[i] = hw[i] / hamwinsum; + + + // initialise the tuning + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + m_meanTunings.push_back(0); + m_localTunings.push_back(0); + } if (channels < getMinChannelCount() || channels > getMaxChannelCount()) return false; @@ -411,12 +424,10 @@ m_frameCount = 0; // m_dictID = 0; m_logSpectrum.clear(); - m_meanTuning0 = 0; - m_meanTuning1 = 0; - m_meanTuning2 = 0; - m_localTuning0 = 0; - m_localTuning1 = 0; - m_localTuning2 = 0; + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + m_meanTunings[iBPS] = 0; + m_localTunings[iBPS] = 0; + } m_localTuning.clear(); } @@ -473,24 +484,28 @@ float one_over_N = 1.0/m_frameCount; // update means of complex tuning variables - m_meanTuning0 *= float(m_frameCount-1)*one_over_N; - m_meanTuning1 *= float(m_frameCount-1)*one_over_N; - m_meanTuning2 *= float(m_frameCount-1)*one_over_N; - - for (int iTone = 0; iTone < 160; iTone = iTone + 3) { - m_meanTuning0 += nm[iTone + 0]*one_over_N; - m_meanTuning1 += nm[iTone + 1]*one_over_N; - m_meanTuning2 += nm[iTone + 2]*one_over_N; + for (int iBPS = 0; iBPS < nBPS; ++iBPS) m_meanTunings[iBPS] *= float(m_frameCount-1)*one_over_N; + + for (int iTone = 0; iTone < round(nNote*0.62/nBPS)*nBPS+1; iTone = iTone + nBPS) { + for (int iBPS = 0; iBPS < nBPS; ++iBPS) m_meanTunings[iBPS] += nm[iTone + iBPS]*one_over_N; float ratioOld = 0.997; - m_localTuning0 *= ratioOld; m_localTuning0 += nm[iTone + 0] * (1 - ratioOld); - m_localTuning1 *= ratioOld; m_localTuning1 += nm[iTone + 1] * (1 - ratioOld); - m_localTuning2 *= ratioOld; m_localTuning2 += nm[iTone + 2] * (1 - ratioOld); + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + m_localTunings[iBPS] *= ratioOld; + m_localTunings[iBPS] += nm[iTone + iBPS] * (1 - ratioOld); + } } - // if (m_tuneLocal) { // local tuning - float localTuningImag = sinvalue * m_localTuning1 - sinvalue * m_localTuning2; - float localTuningReal = m_localTuning0 + cosvalue * m_localTuning1 + cosvalue * m_localTuning2; + // float localTuningImag = sinvalue * m_localTunings[1] - sinvalue * m_localTunings[2]; + // float localTuningReal = m_localTunings[0] + cosvalue * m_localTunings[1] + cosvalue * m_localTunings[2]; + + float localTuningImag = 0; + float localTuningReal = 0; + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + localTuningReal += m_localTunings[iBPS] * cosvalues[iBPS]; + localTuningImag += m_localTunings[iBPS] * sinvalues[iBPS]; + } + float normalisedtuning = atan2(localTuningImag, localTuningReal)/(2*M_PI); m_localTuning.push_back(normalisedtuning); @@ -514,526 +529,526 @@ NNLSBase::FeatureSet NNLSBase::getRemainingFeatures() { - if (debug_on) cerr << "--> getRemainingFeatures" << endl; - FeatureSet fsOut; - if (m_logSpectrum.size() == 0) return fsOut; - int nChord = m_chordnames.size(); + // if (debug_on) cerr << "--> getRemainingFeatures" << endl; + FeatureSet fsOut; + // if (m_logSpectrum.size() == 0) return fsOut; + // int nChord = m_chordnames.size(); + // // + // /** Calculate Tuning + // calculate tuning from (using the angle of the complex number defined by the + // cumulative mean real and imag values) + // **/ + // float meanTuningImag = sinvalue * m_meanTunings[1] - sinvalue * m_meanTunings[2]; + // float meanTuningReal = m_meanTunings[0] + cosvalue * m_meanTunings[1] + cosvalue * m_meanTunings[2]; + // float cumulativetuning = 440 * pow(2,atan2(meanTuningImag, meanTuningReal)/(24*M_PI)); + // float normalisedtuning = atan2(meanTuningImag, meanTuningReal)/(2*M_PI); + // int intShift = floor(normalisedtuning * 3); + // float floatShift = normalisedtuning * 3 - intShift; // floatShift is a really bad name for this + // + // char buffer0 [50]; + // + // sprintf(buffer0, "estimated tuning: %0.1f Hz", cumulativetuning); + // + // // cerr << "normalisedtuning: " << normalisedtuning << '\n'; + // + // // push tuning to FeatureSet fsOut + // Feature f0; // tuning + // f0.hasTimestamp = true; + // f0.timestamp = Vamp::RealTime::frame2RealTime(0, lrintf(m_inputSampleRate));; + // f0.label = buffer0; + // fsOut[0].push_back(f0); + // + // /** Tune Log-Frequency Spectrogram + // calculate a tuned log-frequency spectrogram (f2): use the tuning estimated above (kinda f0) to + // perform linear interpolation on the existing log-frequency spectrogram (kinda f1). + // **/ + // cerr << endl << "[NNLS Chroma Plugin] Tuning Log-Frequency Spectrogram ... "; + // + // float tempValue = 0; + // float dbThreshold = 0; // relative to the background spectrum + // float thresh = pow(10,dbThreshold/20); + // // cerr << "tune local ? " << m_tuneLocal << endl; + // int count = 0; + // + // for (FeatureList::iterator i = m_logSpectrum.begin(); i != m_logSpectrum.end(); ++i) { + // Feature f1 = *i; + // Feature f2; // tuned log-frequency spectrum + // f2.hasTimestamp = true; + // f2.timestamp = f1.timestamp; + // f2.values.push_back(0.0); f2.values.push_back(0.0); // set lower edge to zero + // + // if (m_tuneLocal == 1.0) { + // intShift = floor(m_localTuning[count] * 3); + // floatShift = m_localTuning[count] * 3 - intShift; // floatShift is a really bad name for this + // } + // + // // cerr << intShift << " " << floatShift << endl; + // + // for (unsigned k = 2; k < f1.values.size() - 3; ++k) { // interpolate all inner bins + // tempValue = f1.values[k + intShift] * (1-floatShift) + f1.values[k+intShift+1] * floatShift; + // f2.values.push_back(tempValue); + // } + // + // f2.values.push_back(0.0); f2.values.push_back(0.0); f2.values.push_back(0.0); // upper edge + // vector<float> runningmean = SpecialConvolution(f2.values,hw); + // vector<float> runningstd; + // for (int i = 0; i < nNote; i++) { // first step: squared values into vector (variance) + // runningstd.push_back((f2.values[i] - runningmean[i]) * (f2.values[i] - runningmean[i])); + // } + // runningstd = SpecialConvolution(runningstd,hw); // second step convolve + // for (int i = 0; i < nNote; i++) { + // runningstd[i] = sqrt(runningstd[i]); // square root to finally have running std + // if (runningstd[i] > 0) { + // // f2.values[i] = (f2.values[i] / runningmean[i]) > thresh ? + // // (f2.values[i] - runningmean[i]) / pow(runningstd[i],m_whitening) : 0; + // f2.values[i] = (f2.values[i] - runningmean[i]) > 0 ? + // (f2.values[i] - runningmean[i]) / pow(runningstd[i],m_whitening) : 0; + // } + // if (f2.values[i] < 0) { + // cerr << "ERROR: negative value in logfreq spectrum" << endl; + // } + // } + // fsOut[2].push_back(f2); + // count++; + // } + // cerr << "done." << endl; + // + // /** Semitone spectrum and chromagrams + // Semitone-spaced log-frequency spectrum derived from the tuned log-freq spectrum above. the spectrum + // is inferred using a non-negative least squares algorithm. + // Three different kinds of chromagram are calculated, "treble", "bass", and "both" (which means + // bass and treble stacked onto each other). + // **/ + // if (m_useNNLS == 0) { + // cerr << "[NNLS Chroma Plugin] Mapping to semitone spectrum and chroma ... "; + // } else { + // cerr << "[NNLS Chroma Plugin] Performing NNLS and mapping to chroma ... "; + // } // - /** Calculate Tuning - calculate tuning from (using the angle of the complex number defined by the - cumulative mean real and imag values) - **/ - float meanTuningImag = sinvalue * m_meanTuning1 - sinvalue * m_meanTuning2; - float meanTuningReal = m_meanTuning0 + cosvalue * m_meanTuning1 + cosvalue * m_meanTuning2; - float cumulativetuning = 440 * pow(2,atan2(meanTuningImag, meanTuningReal)/(24*M_PI)); - float normalisedtuning = atan2(meanTuningImag, meanTuningReal)/(2*M_PI); - int intShift = floor(normalisedtuning * 3); - float intFactor = normalisedtuning * 3 - intShift; // intFactor is a really bad name for this - - char buffer0 [50]; - - sprintf(buffer0, "estimated tuning: %0.1f Hz", cumulativetuning); - - // cerr << "normalisedtuning: " << normalisedtuning << '\n'; - - // push tuning to FeatureSet fsOut - Feature f0; // tuning - f0.hasTimestamp = true; - f0.timestamp = Vamp::RealTime::frame2RealTime(0, lrintf(m_inputSampleRate));; - f0.label = buffer0; - fsOut[0].push_back(f0); - - /** Tune Log-Frequency Spectrogram - calculate a tuned log-frequency spectrogram (f2): use the tuning estimated above (kinda f0) to - perform linear interpolation on the existing log-frequency spectrogram (kinda f1). - **/ - cerr << endl << "[NNLS Chroma Plugin] Tuning Log-Frequency Spectrogram ... "; - - float tempValue = 0; - float dbThreshold = 0; // relative to the background spectrum - float thresh = pow(10,dbThreshold/20); - // cerr << "tune local ? " << m_tuneLocal << endl; - int count = 0; - - for (FeatureList::iterator i = m_logSpectrum.begin(); i != m_logSpectrum.end(); ++i) { - Feature f1 = *i; - Feature f2; // tuned log-frequency spectrum - f2.hasTimestamp = true; - f2.timestamp = f1.timestamp; - f2.values.push_back(0.0); f2.values.push_back(0.0); // set lower edge to zero - - if (m_tuneLocal == 1.0) { - intShift = floor(m_localTuning[count] * 3); - intFactor = m_localTuning[count] * 3 - intShift; // intFactor is a really bad name for this - } - - // cerr << intShift << " " << intFactor << endl; - - for (unsigned k = 2; k < f1.values.size() - 3; ++k) { // interpolate all inner bins - tempValue = f1.values[k + intShift] * (1-intFactor) + f1.values[k+intShift+1] * intFactor; - f2.values.push_back(tempValue); - } - - f2.values.push_back(0.0); f2.values.push_back(0.0); f2.values.push_back(0.0); // upper edge - vector<float> runningmean = SpecialConvolution(f2.values,hw); - vector<float> runningstd; - for (int i = 0; i < nNote; i++) { // first step: squared values into vector (variance) - runningstd.push_back((f2.values[i] - runningmean[i]) * (f2.values[i] - runningmean[i])); - } - runningstd = SpecialConvolution(runningstd,hw); // second step convolve - for (int i = 0; i < nNote; i++) { - runningstd[i] = sqrt(runningstd[i]); // square root to finally have running std - if (runningstd[i] > 0) { - // f2.values[i] = (f2.values[i] / runningmean[i]) > thresh ? - // (f2.values[i] - runningmean[i]) / pow(runningstd[i],m_whitening) : 0; - f2.values[i] = (f2.values[i] - runningmean[i]) > 0 ? - (f2.values[i] - runningmean[i]) / pow(runningstd[i],m_whitening) : 0; - } - if (f2.values[i] < 0) { - cerr << "ERROR: negative value in logfreq spectrum" << endl; - } - } - fsOut[2].push_back(f2); - count++; - } - cerr << "done." << endl; - - /** Semitone spectrum and chromagrams - Semitone-spaced log-frequency spectrum derived from the tuned log-freq spectrum above. the spectrum - is inferred using a non-negative least squares algorithm. - Three different kinds of chromagram are calculated, "treble", "bass", and "both" (which means - bass and treble stacked onto each other). - **/ - if (m_useNNLS == 0) { - cerr << "[NNLS Chroma Plugin] Mapping to semitone spectrum and chroma ... "; - } else { - cerr << "[NNLS Chroma Plugin] Performing NNLS and mapping to chroma ... "; - } - - - vector<vector<float> > chordogram; - vector<vector<int> > scoreChordogram; - vector<float> chordchange = vector<float>(fsOut[2].size(),0); - vector<float> oldchroma = vector<float>(12,0); - vector<float> oldbasschroma = vector<float>(12,0); - count = 0; - - for (FeatureList::iterator it = fsOut[2].begin(); it != fsOut[2].end(); ++it) { - Feature f2 = *it; // logfreq spectrum - Feature f3; // semitone spectrum - Feature f4; // treble chromagram - Feature f5; // bass chromagram - Feature f6; // treble and bass chromagram - - f3.hasTimestamp = true; - f3.timestamp = f2.timestamp; - - f4.hasTimestamp = true; - f4.timestamp = f2.timestamp; - - f5.hasTimestamp = true; - f5.timestamp = f2.timestamp; - - f6.hasTimestamp = true; - f6.timestamp = f2.timestamp; - - float b[nNote]; - - bool some_b_greater_zero = false; - float sumb = 0; - for (int i = 0; i < nNote; i++) { - // b[i] = m_dict[(nNote * count + i) % (nNote * 84)]; - b[i] = f2.values[i]; - sumb += b[i]; - if (b[i] > 0) { - some_b_greater_zero = true; - } - } - - // here's where the non-negative least squares algorithm calculates the note activation x - - vector<float> chroma = vector<float>(12, 0); - vector<float> basschroma = vector<float>(12, 0); - float currval; - unsigned iSemitone = 0; - - if (some_b_greater_zero) { - if (m_useNNLS == 0) { - for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { - currval = 0; - currval += b[iNote + 1 + -1] * 0.5; - currval += b[iNote + 1 + 0] * 1.0; - currval += b[iNote + 1 + 1] * 0.5; - f3.values.push_back(currval); - chroma[iSemitone % 12] += currval * treblewindow[iSemitone]; - basschroma[iSemitone % 12] += currval * basswindow[iSemitone]; - iSemitone++; - } - - } else { - float x[84+1000]; - for (int i = 1; i < 1084; ++i) x[i] = 1.0; - vector<int> signifIndex; - int index=0; - sumb /= 84.0; - for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { - float currval = 0; - currval += b[iNote + 1 + -1]; - currval += b[iNote + 1 + 0]; - currval += b[iNote + 1 + 1]; - if (currval > 0) signifIndex.push_back(index); - f3.values.push_back(0); // fill the values, change later - index++; - } - float rnorm; - float w[84+1000]; - float zz[84+1000]; - int indx[84+1000]; - int mode; - int dictsize = nNote*signifIndex.size(); - // cerr << "dictsize is " << dictsize << "and values size" << f3.values.size()<< endl; - float *curr_dict = new float[dictsize]; - for (unsigned iNote = 0; iNote < signifIndex.size(); ++iNote) { - for (unsigned iBin = 0; iBin < nNote; iBin++) { - curr_dict[iNote * nNote + iBin] = 1.0 * m_dict[signifIndex[iNote] * nNote + iBin]; - } - } - nnls(curr_dict, nNote, nNote, signifIndex.size(), b, x, &rnorm, w, zz, indx, &mode); - delete [] curr_dict; - for (unsigned iNote = 0; iNote < signifIndex.size(); ++iNote) { - f3.values[signifIndex[iNote]] = x[iNote]; - // cerr << mode << endl; - chroma[signifIndex[iNote] % 12] += x[iNote] * treblewindow[signifIndex[iNote]]; - basschroma[signifIndex[iNote] % 12] += x[iNote] * basswindow[signifIndex[iNote]]; - } - } - } - - - - - f4.values = chroma; - f5.values = basschroma; - chroma.insert(chroma.begin(), basschroma.begin(), basschroma.end()); // just stack the both chromas - f6.values = chroma; - - if (m_doNormalizeChroma > 0) { - vector<float> chromanorm = vector<float>(3,0); - switch (int(m_doNormalizeChroma)) { - case 0: // should never end up here - break; - case 1: - chromanorm[0] = *max_element(f4.values.begin(), f4.values.end()); - chromanorm[1] = *max_element(f5.values.begin(), f5.values.end()); - chromanorm[2] = max(chromanorm[0], chromanorm[1]); - break; - case 2: - for (vector<float>::iterator it = f4.values.begin(); it != f4.values.end(); ++it) { - chromanorm[0] += *it; - } - for (vector<float>::iterator it = f5.values.begin(); it != f5.values.end(); ++it) { - chromanorm[1] += *it; - } - for (vector<float>::iterator it = f6.values.begin(); it != f6.values.end(); ++it) { - chromanorm[2] += *it; - } - break; - case 3: - for (vector<float>::iterator it = f4.values.begin(); it != f4.values.end(); ++it) { - chromanorm[0] += pow(*it,2); - } - chromanorm[0] = sqrt(chromanorm[0]); - for (vector<float>::iterator it = f5.values.begin(); it != f5.values.end(); ++it) { - chromanorm[1] += pow(*it,2); - } - chromanorm[1] = sqrt(chromanorm[1]); - for (vector<float>::iterator it = f6.values.begin(); it != f6.values.end(); ++it) { - chromanorm[2] += pow(*it,2); - } - chromanorm[2] = sqrt(chromanorm[2]); - break; - } - if (chromanorm[0] > 0) { - for (int i = 0; i < f4.values.size(); i++) { - f4.values[i] /= chromanorm[0]; - } - } - if (chromanorm[1] > 0) { - for (int i = 0; i < f5.values.size(); i++) { - f5.values[i] /= chromanorm[1]; - } - } - if (chromanorm[2] > 0) { - for (int i = 0; i < f6.values.size(); i++) { - f6.values[i] /= chromanorm[2]; - } - } - - } - - // local chord estimation - vector<float> currentChordSalience; - float tempchordvalue = 0; - float sumchordvalue = 0; - - for (int iChord = 0; iChord < nChord; iChord++) { - tempchordvalue = 0; - for (int iBin = 0; iBin < 12; iBin++) { - tempchordvalue += m_chorddict[24 * iChord + iBin] * chroma[iBin]; - } - for (int iBin = 12; iBin < 24; iBin++) { - tempchordvalue += m_chorddict[24 * iChord + iBin] * chroma[iBin]; - } - sumchordvalue+=tempchordvalue; - currentChordSalience.push_back(tempchordvalue); - } - if (sumchordvalue > 0) { - for (int iChord = 0; iChord < nChord; iChord++) { - currentChordSalience[iChord] /= sumchordvalue; - } - } else { - currentChordSalience[nChord-1] = 1.0; - } - chordogram.push_back(currentChordSalience); - - fsOut[3].push_back(f3); - fsOut[4].push_back(f4); - fsOut[5].push_back(f5); - fsOut[6].push_back(f6); - count++; - } - cerr << "done." << endl; - - - /* Simple chord estimation - I just take the local chord estimates ("currentChordSalience") and average them over time, then - take the maximum. Very simple, don't do this at home... - */ - cerr << "[NNLS Chroma Plugin] Chord Estimation ... "; - count = 0; - int halfwindowlength = m_inputSampleRate / m_stepSize; - vector<int> chordSequence; - for (FeatureList::iterator it = fsOut[6].begin(); it != fsOut[6].end(); ++it) { // initialise the score chordogram - vector<int> temp = vector<int>(nChord,0); - scoreChordogram.push_back(temp); - } - for (FeatureList::iterator it = fsOut[6].begin(); it < fsOut[6].end()-2*halfwindowlength-1; ++it) { - int startIndex = count + 1; - int endIndex = count + 2 * halfwindowlength; - - float chordThreshold = 2.5/nChord;//*(2*halfwindowlength+1); - - vector<int> chordCandidates; - for (unsigned iChord = 0; iChord < nChord-1; iChord++) { - // float currsum = 0; - // for (unsigned iFrame = startIndex; iFrame < endIndex; ++iFrame) { - // currsum += chordogram[iFrame][iChord]; - // } - // if (currsum > chordThreshold) chordCandidates.push_back(iChord); - for (unsigned iFrame = startIndex; iFrame < endIndex; ++iFrame) { - if (chordogram[iFrame][iChord] > chordThreshold) { - chordCandidates.push_back(iChord); - break; - } - } - } - chordCandidates.push_back(nChord-1); - // cerr << chordCandidates.size() << endl; - - float maxval = 0; // will be the value of the most salient *chord change* in this frame - float maxindex = 0; //... and the index thereof - unsigned bestchordL = nChord-1; // index of the best "left" chord - unsigned bestchordR = nChord-1; // index of the best "right" chord - - for (int iWF = 1; iWF < 2*halfwindowlength; ++iWF) { - // now find the max values on both sides of iWF - // left side: - float maxL = 0; - unsigned maxindL = nChord-1; - for (unsigned kChord = 0; kChord < chordCandidates.size(); kChord++) { - unsigned iChord = chordCandidates[kChord]; - float currsum = 0; - for (unsigned iFrame = 0; iFrame < iWF-1; ++iFrame) { - currsum += chordogram[count+iFrame][iChord]; - } - if (iChord == nChord-1) currsum *= 0.8; - if (currsum > maxL) { - maxL = currsum; - maxindL = iChord; - } - } - // right side: - float maxR = 0; - unsigned maxindR = nChord-1; - for (unsigned kChord = 0; kChord < chordCandidates.size(); kChord++) { - unsigned iChord = chordCandidates[kChord]; - float currsum = 0; - for (unsigned iFrame = iWF-1; iFrame < 2*halfwindowlength; ++iFrame) { - currsum += chordogram[count+iFrame][iChord]; - } - if (iChord == nChord-1) currsum *= 0.8; - if (currsum > maxR) { - maxR = currsum; - maxindR = iChord; - } - } - if (maxL+maxR > maxval) { - maxval = maxL+maxR; - maxindex = iWF; - bestchordL = maxindL; - bestchordR = maxindR; - } - - } - // cerr << "maxindex: " << maxindex << ", bestchordR is " << bestchordR << ", of frame " << count << endl; - // add a score to every chord-frame-point that was part of a maximum - for (unsigned iFrame = 0; iFrame < maxindex-1; ++iFrame) { - scoreChordogram[iFrame+count][bestchordL]++; - } - for (unsigned iFrame = maxindex-1; iFrame < 2*halfwindowlength; ++iFrame) { - scoreChordogram[iFrame+count][bestchordR]++; - } - if (bestchordL != bestchordR) chordchange[maxindex+count] += (halfwindowlength - abs(maxindex-halfwindowlength)) * 2.0 / halfwindowlength; - count++; - } - // cerr << "******* agent finished *******" << endl; - count = 0; - for (FeatureList::iterator it = fsOut[6].begin(); it != fsOut[6].end(); ++it) { - float maxval = 0; // will be the value of the most salient chord in this frame - float maxindex = 0; //... and the index thereof - for (unsigned iChord = 0; iChord < nChord; iChord++) { - if (scoreChordogram[count][iChord] > maxval) { - maxval = scoreChordogram[count][iChord]; - maxindex = iChord; - // cerr << iChord << endl; - } - } - chordSequence.push_back(maxindex); - // cerr << "before modefilter, maxindex: " << maxindex << endl; - count++; - } - // cerr << "******* mode filter done *******" << endl; - - - // mode filter on chordSequence - count = 0; - string oldChord = ""; - for (FeatureList::iterator it = fsOut[6].begin(); it != fsOut[6].end(); ++it) { - Feature f6 = *it; - Feature f7; // chord estimate - f7.hasTimestamp = true; - f7.timestamp = f6.timestamp; - Feature f8; // chord estimate - f8.hasTimestamp = true; - f8.timestamp = f6.timestamp; - - vector<int> chordCount = vector<int>(nChord,0); - int maxChordCount = 0; - int maxChordIndex = nChord-1; - string maxChord; - int startIndex = max(count - halfwindowlength/2,0); - int endIndex = min(int(chordogram.size()), count + halfwindowlength/2); - for (int i = startIndex; i < endIndex; i++) { - chordCount[chordSequence[i]]++; - if (chordCount[chordSequence[i]] > maxChordCount) { - // cerr << "start index " << startIndex << endl; - maxChordCount++; - maxChordIndex = chordSequence[i]; - maxChord = m_chordnames[maxChordIndex]; - } - } - // chordSequence[count] = maxChordIndex; - // cerr << maxChordIndex << endl; - f8.values.push_back(chordchange[count]/(halfwindowlength*2)); - // cerr << chordchange[count] << endl; - fsOut[9].push_back(f8); - if (oldChord != maxChord) { - oldChord = maxChord; - - // char buffer1 [50]; - // if (maxChordIndex < nChord - 1) { - // sprintf(buffer1, "%s%s", notenames[maxChordIndex % 12 + 12], chordtypes[maxChordIndex]); - // } else { - // sprintf(buffer1, "N"); - // } - // f7.label = buffer1; - f7.label = m_chordnames[maxChordIndex]; - fsOut[7].push_back(f7); - } - count++; - } - Feature f7; // last chord estimate - f7.hasTimestamp = true; - f7.timestamp = fsOut[6][fsOut[6].size()-1].timestamp; - f7.label = "N"; - fsOut[7].push_back(f7); - cerr << "done." << endl; - // // musicity - // count = 0; - // int oldlabeltype = 0; // start value is 0, music is 1, speech is 2 - // vector<float> musicityValue; - // for (FeatureList::iterator it = fsOut[4].begin(); it != fsOut[4].end(); ++it) { - // Feature f4 = *it; - // - // int startIndex = max(count - musicitykernelwidth/2,0); - // int endIndex = min(int(chordogram.size()), startIndex + musicitykernelwidth - 1); - // float chromasum = 0; - // float diffsum = 0; - // for (int k = 0; k < 12; k++) { - // for (int i = startIndex + 1; i < endIndex; i++) { - // chromasum += pow(fsOut[4][i].values[k],2); - // diffsum += abs(fsOut[4][i-1].values[k] - fsOut[4][i].values[k]); - // } - // } - // diffsum /= chromasum; - // musicityValue.push_back(diffsum); - // count++; - // } - // - // float musicityThreshold = 0.44; - // if (m_stepSize == 4096) { - // musicityThreshold = 0.74; - // } - // if (m_stepSize == 4410) { - // musicityThreshold = 0.77; - // } - // - // count = 0; - // for (FeatureList::iterator it = fsOut[4].begin(); it != fsOut[4].end(); ++it) { - // Feature f4 = *it; - // Feature f8; // musicity - // Feature f9; // musicity segmenter - // - // f8.hasTimestamp = true; - // f8.timestamp = f4.timestamp; - // f9.hasTimestamp = true; - // f9.timestamp = f4.timestamp; - // - // int startIndex = max(count - musicitykernelwidth/2,0); - // int endIndex = min(int(chordogram.size()), startIndex + musicitykernelwidth - 1); - // int musicityCount = 0; - // for (int i = startIndex; i <= endIndex; i++) { - // if (musicityValue[i] > musicityThreshold) musicityCount++; - // } - // bool isSpeech = (2 * musicityCount > endIndex - startIndex + 1); - // - // if (isSpeech) { - // if (oldlabeltype != 2) { - // f9.label = "Speech"; - // fsOut[9].push_back(f9); - // oldlabeltype = 2; - // } - // } else { - // if (oldlabeltype != 1) { - // f9.label = "Music"; - // fsOut[9].push_back(f9); - // oldlabeltype = 1; - // } - // } - // f8.values.push_back(musicityValue[count]); - // fsOut[8].push_back(f8); - // count++; - // } + // + // vector<vector<float> > chordogram; + // vector<vector<int> > scoreChordogram; + // vector<float> chordchange = vector<float>(fsOut[2].size(),0); + // vector<float> oldchroma = vector<float>(12,0); + // vector<float> oldbasschroma = vector<float>(12,0); + // count = 0; + // + // for (FeatureList::iterator it = fsOut[2].begin(); it != fsOut[2].end(); ++it) { + // Feature f2 = *it; // logfreq spectrum + // Feature f3; // semitone spectrum + // Feature f4; // treble chromagram + // Feature f5; // bass chromagram + // Feature f6; // treble and bass chromagram + // + // f3.hasTimestamp = true; + // f3.timestamp = f2.timestamp; + // + // f4.hasTimestamp = true; + // f4.timestamp = f2.timestamp; + // + // f5.hasTimestamp = true; + // f5.timestamp = f2.timestamp; + // + // f6.hasTimestamp = true; + // f6.timestamp = f2.timestamp; + // + // float b[nNote]; + // + // bool some_b_greater_zero = false; + // float sumb = 0; + // for (int i = 0; i < nNote; i++) { + // // b[i] = m_dict[(nNote * count + i) % (nNote * 84)]; + // b[i] = f2.values[i]; + // sumb += b[i]; + // if (b[i] > 0) { + // some_b_greater_zero = true; + // } + // } + // + // // here's where the non-negative least squares algorithm calculates the note activation x + // + // vector<float> chroma = vector<float>(12, 0); + // vector<float> basschroma = vector<float>(12, 0); + // float currval; + // unsigned iSemitone = 0; + // + // if (some_b_greater_zero) { + // if (m_useNNLS == 0) { + // for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { + // currval = 0; + // currval += b[iNote + 1 + -1] * 0.5; + // currval += b[iNote + 1 + 0] * 1.0; + // currval += b[iNote + 1 + 1] * 0.5; + // f3.values.push_back(currval); + // chroma[iSemitone % 12] += currval * treblewindow[iSemitone]; + // basschroma[iSemitone % 12] += currval * basswindow[iSemitone]; + // iSemitone++; + // } + // + // } else { + // float x[84+1000]; + // for (int i = 1; i < 1084; ++i) x[i] = 1.0; + // vector<int> signifIndex; + // int index=0; + // sumb /= 84.0; + // for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { + // float currval = 0; + // currval += b[iNote + 1 + -1]; + // currval += b[iNote + 1 + 0]; + // currval += b[iNote + 1 + 1]; + // if (currval > 0) signifIndex.push_back(index); + // f3.values.push_back(0); // fill the values, change later + // index++; + // } + // float rnorm; + // float w[84+1000]; + // float zz[84+1000]; + // int indx[84+1000]; + // int mode; + // int dictsize = nNote*signifIndex.size(); + // // cerr << "dictsize is " << dictsize << "and values size" << f3.values.size()<< endl; + // float *curr_dict = new float[dictsize]; + // for (unsigned iNote = 0; iNote < signifIndex.size(); ++iNote) { + // for (unsigned iBin = 0; iBin < nNote; iBin++) { + // curr_dict[iNote * nNote + iBin] = 1.0 * m_dict[signifIndex[iNote] * nNote + iBin]; + // } + // } + // nnls(curr_dict, nNote, nNote, signifIndex.size(), b, x, &rnorm, w, zz, indx, &mode); + // delete [] curr_dict; + // for (unsigned iNote = 0; iNote < signifIndex.size(); ++iNote) { + // f3.values[signifIndex[iNote]] = x[iNote]; + // // cerr << mode << endl; + // chroma[signifIndex[iNote] % 12] += x[iNote] * treblewindow[signifIndex[iNote]]; + // basschroma[signifIndex[iNote] % 12] += x[iNote] * basswindow[signifIndex[iNote]]; + // } + // } + // } + // + // + // + // + // f4.values = chroma; + // f5.values = basschroma; + // chroma.insert(chroma.begin(), basschroma.begin(), basschroma.end()); // just stack the both chromas + // f6.values = chroma; + // + // if (m_doNormalizeChroma > 0) { + // vector<float> chromanorm = vector<float>(3,0); + // switch (int(m_doNormalizeChroma)) { + // case 0: // should never end up here + // break; + // case 1: + // chromanorm[0] = *max_element(f4.values.begin(), f4.values.end()); + // chromanorm[1] = *max_element(f5.values.begin(), f5.values.end()); + // chromanorm[2] = max(chromanorm[0], chromanorm[1]); + // break; + // case 2: + // for (vector<float>::iterator it = f4.values.begin(); it != f4.values.end(); ++it) { + // chromanorm[0] += *it; + // } + // for (vector<float>::iterator it = f5.values.begin(); it != f5.values.end(); ++it) { + // chromanorm[1] += *it; + // } + // for (vector<float>::iterator it = f6.values.begin(); it != f6.values.end(); ++it) { + // chromanorm[2] += *it; + // } + // break; + // case 3: + // for (vector<float>::iterator it = f4.values.begin(); it != f4.values.end(); ++it) { + // chromanorm[0] += pow(*it,2); + // } + // chromanorm[0] = sqrt(chromanorm[0]); + // for (vector<float>::iterator it = f5.values.begin(); it != f5.values.end(); ++it) { + // chromanorm[1] += pow(*it,2); + // } + // chromanorm[1] = sqrt(chromanorm[1]); + // for (vector<float>::iterator it = f6.values.begin(); it != f6.values.end(); ++it) { + // chromanorm[2] += pow(*it,2); + // } + // chromanorm[2] = sqrt(chromanorm[2]); + // break; + // } + // if (chromanorm[0] > 0) { + // for (int i = 0; i < f4.values.size(); i++) { + // f4.values[i] /= chromanorm[0]; + // } + // } + // if (chromanorm[1] > 0) { + // for (int i = 0; i < f5.values.size(); i++) { + // f5.values[i] /= chromanorm[1]; + // } + // } + // if (chromanorm[2] > 0) { + // for (int i = 0; i < f6.values.size(); i++) { + // f6.values[i] /= chromanorm[2]; + // } + // } + // + // } + // + // // local chord estimation + // vector<float> currentChordSalience; + // float tempchordvalue = 0; + // float sumchordvalue = 0; + // + // for (int iChord = 0; iChord < nChord; iChord++) { + // tempchordvalue = 0; + // for (int iBin = 0; iBin < 12; iBin++) { + // tempchordvalue += m_chorddict[24 * iChord + iBin] * chroma[iBin]; + // } + // for (int iBin = 12; iBin < 24; iBin++) { + // tempchordvalue += m_chorddict[24 * iChord + iBin] * chroma[iBin]; + // } + // sumchordvalue+=tempchordvalue; + // currentChordSalience.push_back(tempchordvalue); + // } + // if (sumchordvalue > 0) { + // for (int iChord = 0; iChord < nChord; iChord++) { + // currentChordSalience[iChord] /= sumchordvalue; + // } + // } else { + // currentChordSalience[nChord-1] = 1.0; + // } + // chordogram.push_back(currentChordSalience); + // + // fsOut[3].push_back(f3); + // fsOut[4].push_back(f4); + // fsOut[5].push_back(f5); + // fsOut[6].push_back(f6); + // count++; + // } + // cerr << "done." << endl; + // + // + // /* Simple chord estimation + // I just take the local chord estimates ("currentChordSalience") and average them over time, then + // take the maximum. Very simple, don't do this at home... + // */ + // cerr << "[NNLS Chroma Plugin] Chord Estimation ... "; + // count = 0; + // int halfwindowlength = m_inputSampleRate / m_stepSize; + // vector<int> chordSequence; + // for (FeatureList::iterator it = fsOut[6].begin(); it != fsOut[6].end(); ++it) { // initialise the score chordogram + // vector<int> temp = vector<int>(nChord,0); + // scoreChordogram.push_back(temp); + // } + // for (FeatureList::iterator it = fsOut[6].begin(); it < fsOut[6].end()-2*halfwindowlength-1; ++it) { + // int startIndex = count + 1; + // int endIndex = count + 2 * halfwindowlength; + // + // float chordThreshold = 2.5/nChord;//*(2*halfwindowlength+1); + // + // vector<int> chordCandidates; + // for (unsigned iChord = 0; iChord < nChord-1; iChord++) { + // // float currsum = 0; + // // for (unsigned iFrame = startIndex; iFrame < endIndex; ++iFrame) { + // // currsum += chordogram[iFrame][iChord]; + // // } + // // if (currsum > chordThreshold) chordCandidates.push_back(iChord); + // for (unsigned iFrame = startIndex; iFrame < endIndex; ++iFrame) { + // if (chordogram[iFrame][iChord] > chordThreshold) { + // chordCandidates.push_back(iChord); + // break; + // } + // } + // } + // chordCandidates.push_back(nChord-1); + // // cerr << chordCandidates.size() << endl; + // + // float maxval = 0; // will be the value of the most salient *chord change* in this frame + // float maxindex = 0; //... and the index thereof + // unsigned bestchordL = nChord-1; // index of the best "left" chord + // unsigned bestchordR = nChord-1; // index of the best "right" chord + // + // for (int iWF = 1; iWF < 2*halfwindowlength; ++iWF) { + // // now find the max values on both sides of iWF + // // left side: + // float maxL = 0; + // unsigned maxindL = nChord-1; + // for (unsigned kChord = 0; kChord < chordCandidates.size(); kChord++) { + // unsigned iChord = chordCandidates[kChord]; + // float currsum = 0; + // for (unsigned iFrame = 0; iFrame < iWF-1; ++iFrame) { + // currsum += chordogram[count+iFrame][iChord]; + // } + // if (iChord == nChord-1) currsum *= 0.8; + // if (currsum > maxL) { + // maxL = currsum; + // maxindL = iChord; + // } + // } + // // right side: + // float maxR = 0; + // unsigned maxindR = nChord-1; + // for (unsigned kChord = 0; kChord < chordCandidates.size(); kChord++) { + // unsigned iChord = chordCandidates[kChord]; + // float currsum = 0; + // for (unsigned iFrame = iWF-1; iFrame < 2*halfwindowlength; ++iFrame) { + // currsum += chordogram[count+iFrame][iChord]; + // } + // if (iChord == nChord-1) currsum *= 0.8; + // if (currsum > maxR) { + // maxR = currsum; + // maxindR = iChord; + // } + // } + // if (maxL+maxR > maxval) { + // maxval = maxL+maxR; + // maxindex = iWF; + // bestchordL = maxindL; + // bestchordR = maxindR; + // } + // + // } + // // cerr << "maxindex: " << maxindex << ", bestchordR is " << bestchordR << ", of frame " << count << endl; + // // add a score to every chord-frame-point that was part of a maximum + // for (unsigned iFrame = 0; iFrame < maxindex-1; ++iFrame) { + // scoreChordogram[iFrame+count][bestchordL]++; + // } + // for (unsigned iFrame = maxindex-1; iFrame < 2*halfwindowlength; ++iFrame) { + // scoreChordogram[iFrame+count][bestchordR]++; + // } + // if (bestchordL != bestchordR) chordchange[maxindex+count] += (halfwindowlength - abs(maxindex-halfwindowlength)) * 2.0 / halfwindowlength; + // count++; + // } + // // cerr << "******* agent finished *******" << endl; + // count = 0; + // for (FeatureList::iterator it = fsOut[6].begin(); it != fsOut[6].end(); ++it) { + // float maxval = 0; // will be the value of the most salient chord in this frame + // float maxindex = 0; //... and the index thereof + // for (unsigned iChord = 0; iChord < nChord; iChord++) { + // if (scoreChordogram[count][iChord] > maxval) { + // maxval = scoreChordogram[count][iChord]; + // maxindex = iChord; + // // cerr << iChord << endl; + // } + // } + // chordSequence.push_back(maxindex); + // // cerr << "before modefilter, maxindex: " << maxindex << endl; + // count++; + // } + // // cerr << "******* mode filter done *******" << endl; + // + // + // // mode filter on chordSequence + // count = 0; + // string oldChord = ""; + // for (FeatureList::iterator it = fsOut[6].begin(); it != fsOut[6].end(); ++it) { + // Feature f6 = *it; + // Feature f7; // chord estimate + // f7.hasTimestamp = true; + // f7.timestamp = f6.timestamp; + // Feature f8; // chord estimate + // f8.hasTimestamp = true; + // f8.timestamp = f6.timestamp; + // + // vector<int> chordCount = vector<int>(nChord,0); + // int maxChordCount = 0; + // int maxChordIndex = nChord-1; + // string maxChord; + // int startIndex = max(count - halfwindowlength/2,0); + // int endIndex = min(int(chordogram.size()), count + halfwindowlength/2); + // for (int i = startIndex; i < endIndex; i++) { + // chordCount[chordSequence[i]]++; + // if (chordCount[chordSequence[i]] > maxChordCount) { + // // cerr << "start index " << startIndex << endl; + // maxChordCount++; + // maxChordIndex = chordSequence[i]; + // maxChord = m_chordnames[maxChordIndex]; + // } + // } + // // chordSequence[count] = maxChordIndex; + // // cerr << maxChordIndex << endl; + // f8.values.push_back(chordchange[count]/(halfwindowlength*2)); + // // cerr << chordchange[count] << endl; + // fsOut[9].push_back(f8); + // if (oldChord != maxChord) { + // oldChord = maxChord; + // + // // char buffer1 [50]; + // // if (maxChordIndex < nChord - 1) { + // // sprintf(buffer1, "%s%s", notenames[maxChordIndex % 12 + 12], chordtypes[maxChordIndex]); + // // } else { + // // sprintf(buffer1, "N"); + // // } + // // f7.label = buffer1; + // f7.label = m_chordnames[maxChordIndex]; + // fsOut[7].push_back(f7); + // } + // count++; + // } + // Feature f7; // last chord estimate + // f7.hasTimestamp = true; + // f7.timestamp = fsOut[6][fsOut[6].size()-1].timestamp; + // f7.label = "N"; + // fsOut[7].push_back(f7); + // cerr << "done." << endl; + // // // musicity + // // count = 0; + // // int oldlabeltype = 0; // start value is 0, music is 1, speech is 2 + // // vector<float> musicityValue; + // // for (FeatureList::iterator it = fsOut[4].begin(); it != fsOut[4].end(); ++it) { + // // Feature f4 = *it; + // // + // // int startIndex = max(count - musicitykernelwidth/2,0); + // // int endIndex = min(int(chordogram.size()), startIndex + musicitykernelwidth - 1); + // // float chromasum = 0; + // // float diffsum = 0; + // // for (int k = 0; k < 12; k++) { + // // for (int i = startIndex + 1; i < endIndex; i++) { + // // chromasum += pow(fsOut[4][i].values[k],2); + // // diffsum += abs(fsOut[4][i-1].values[k] - fsOut[4][i].values[k]); + // // } + // // } + // // diffsum /= chromasum; + // // musicityValue.push_back(diffsum); + // // count++; + // // } + // // + // // float musicityThreshold = 0.44; + // // if (m_stepSize == 4096) { + // // musicityThreshold = 0.74; + // // } + // // if (m_stepSize == 4410) { + // // musicityThreshold = 0.77; + // // } + // // + // // count = 0; + // // for (FeatureList::iterator it = fsOut[4].begin(); it != fsOut[4].end(); ++it) { + // // Feature f4 = *it; + // // Feature f8; // musicity + // // Feature f9; // musicity segmenter + // // + // // f8.hasTimestamp = true; + // // f8.timestamp = f4.timestamp; + // // f9.hasTimestamp = true; + // // f9.timestamp = f4.timestamp; + // // + // // int startIndex = max(count - musicitykernelwidth/2,0); + // // int endIndex = min(int(chordogram.size()), startIndex + musicitykernelwidth - 1); + // // int musicityCount = 0; + // // for (int i = startIndex; i <= endIndex; i++) { + // // if (musicityValue[i] > musicityThreshold) musicityCount++; + // // } + // // bool isSpeech = (2 * musicityCount > endIndex - startIndex + 1); + // // + // // if (isSpeech) { + // // if (oldlabeltype != 2) { + // // f9.label = "Speech"; + // // fsOut[9].push_back(f9); + // // oldlabeltype = 2; + // // } + // // } else { + // // if (oldlabeltype != 1) { + // // f9.label = "Music"; + // // fsOut[9].push_back(f9); + // // oldlabeltype = 1; + // // } + // // } + // // f8.values.push_back(musicityValue[count]); + // // fsOut[8].push_back(f8); + // // count++; + // // } return fsOut; }
--- a/NNLSBase.h Mon Nov 15 09:23:49 2010 +0900 +++ b/NNLSBase.h Mon Nov 15 11:01:56 2010 +0900 @@ -60,12 +60,8 @@ size_t m_blockSize; size_t m_stepSize; int m_lengthOfNoteIndex; - float m_meanTuning0; - float m_meanTuning1; - float m_meanTuning2; - float m_localTuning0; - float m_localTuning1; - float m_localTuning2; + vector<float> m_meanTunings; + vector<float> m_localTunings; float m_whitening; float m_preset; float m_s; @@ -82,6 +78,8 @@ float m_doNormalizeChroma; float m_rollon; vector<float> hw; + vector<float> sinvalues; + vector<float> cosvalues; };
--- a/NNLSChroma.cpp Mon Nov 15 09:23:49 2010 +0900 +++ b/NNLSChroma.cpp Mon Nov 15 11:01:56 2010 +0900 @@ -219,12 +219,16 @@ calculate tuning from (using the angle of the complex number defined by the cumulative mean real and imag values) **/ - float meanTuningImag = sinvalue * m_meanTuning1 - sinvalue * m_meanTuning2; - float meanTuningReal = m_meanTuning0 + cosvalue * m_meanTuning1 + cosvalue * m_meanTuning2; + float meanTuningImag = 0; + float meanTuningReal = 0; + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + meanTuningReal += m_meanTunings[iBPS] * cosvalues[iBPS]; + meanTuningImag += m_meanTunings[iBPS] * sinvalues[iBPS]; + } float cumulativetuning = 440 * pow(2,atan2(meanTuningImag, meanTuningReal)/(24*M_PI)); float normalisedtuning = atan2(meanTuningImag, meanTuningReal)/(2*M_PI); int intShift = floor(normalisedtuning * 3); - float intFactor = normalisedtuning * 3 - intShift; // intFactor is a really bad name for this + float floatShift = normalisedtuning * 3 - intShift; // floatShift is a really bad name for this char buffer0 [50]; @@ -244,8 +248,6 @@ // cerr << "tune local ? " << m_tuneLocal << endl; int count = 0; - cerr << nNote; - cerr << endl << "-------------------------------------"<< endl; for (FeatureList::iterator i = m_logSpectrum.begin(); i != m_logSpectrum.end(); ++i) { Feature f1 = *i; @@ -256,13 +258,13 @@ if (m_tuneLocal) { intShift = floor(m_localTuning[count] * 3); - intFactor = m_localTuning[count] * 3 - intShift; // intFactor is a really bad name for this + floatShift = m_localTuning[count] * 3 - intShift; // floatShift is a really bad name for this } - // cerr << intShift << " " << intFactor << endl; + // cerr << intShift << " " << floatShift << endl; for (unsigned k = 2; k < f1.values.size() - 3; ++k) { // interpolate all inner bins - tempValue = f1.values[k + intShift] * (1-intFactor) + f1.values[k+intShift+1] * intFactor; + tempValue = f1.values[k + intShift] * (1-floatShift) + f1.values[k+intShift+1] * floatShift; f2.values.push_back(tempValue); } @@ -349,11 +351,11 @@ if (some_b_greater_zero) { if (m_useNNLS == 0) { - for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { + for (unsigned iNote = nBPS/2 + 2; iNote < nNote - nBPS/2; iNote += nBPS) { currval = 0; - currval += b[iNote + 1 + -1] * 0.5; - currval += b[iNote + 1 + 0] * 1.0; - currval += b[iNote + 1 + 1] * 0.5; + for (int iBPS = -nBPS/2; iBPS < nBPS/2+1; ++iBPS) { + currval += b[iNote + iBPS] * (1-abs(iBPS*1.0/(nBPS/2+1))); + } f3.values.push_back(currval); chroma[iSemitone % 12] += currval * treblewindow[iSemitone]; basschroma[iSemitone % 12] += currval * basswindow[iSemitone]; @@ -366,11 +368,11 @@ vector<int> signifIndex; int index=0; sumb /= 84.0; - for (unsigned iNote = 2; iNote < nNote - 2; iNote += 3) { + for (unsigned iNote = nBPS/2 + 2; iNote < nNote - nBPS/2; iNote += nBPS) { float currval = 0; - currval += b[iNote + 1 + -1]; - currval += b[iNote + 1 + 0]; - currval += b[iNote + 1 + 1]; + for (int iBPS = -nBPS/2; iBPS < nBPS/2+1; ++iBPS) { + currval += b[iNote + iBPS]; + } if (currval > 0) signifIndex.push_back(index); f3.values.push_back(0); // fill the values, change later index++;
--- a/Tuning.cpp Mon Nov 15 09:23:49 2010 +0900 +++ b/Tuning.cpp Mon Nov 15 11:01:56 2010 +0900 @@ -187,8 +187,15 @@ calculate tuning from (using the angle of the complex number defined by the cumulative mean real and imag values) **/ - float meanTuningImag = sinvalue * m_meanTuning1 - sinvalue * m_meanTuning2; - float meanTuningReal = m_meanTuning0 + cosvalue * m_meanTuning1 + cosvalue * m_meanTuning2; + + float meanTuningImag = 0; + float meanTuningReal = 0; + for (int iBPS = 0; iBPS < nBPS; ++iBPS) { + meanTuningReal += m_meanTunings[iBPS] * cosvalues[iBPS]; + meanTuningImag += m_meanTunings[iBPS] * sinvalues[iBPS]; + } + + float cumulativetuning = 440 * pow(2,atan2(meanTuningImag, meanTuningReal)/(24*M_PI)); char buffer0 [50];
--- a/chromamethods.cpp Mon Nov 15 09:23:49 2010 +0900 +++ b/chromamethods.cpp Mon Nov 15 11:01:56 2010 +0900 @@ -105,7 +105,7 @@ bool logFreqMatrix(int fs, int blocksize, float *outmatrix) { - int binspersemitone = 3; // this must be 3 + int binspersemitone = nBPS; int minoctave = 0; // this must be 0 int maxoctave = 7; // this must be 7 int oversampling = 80; @@ -128,14 +128,14 @@ int maxMIDI = 21 + maxoctave * 12; // this includes one additional semitone! vector<float> cq_f; float oob = 1.0/binspersemitone; // one over binspersemitone - cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-69))); // 0.083333 is approx 1/12 - cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI+oob-69))); - for (int i = minMIDI + 1; i < maxMIDI; ++i) { - for (int k = -1; k < 2; ++k) { + // cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-69))); // 0.083333 is approx 1/12 + // cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI+oob-69))); + for (int i = minMIDI; i < maxMIDI; ++i) { + for (int k = 0; k < binspersemitone; ++k) { cq_f.push_back(440 * pow(2.0,0.083333333333 * (i+oob*k-69))); } } - cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-oob-69))); + // cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-oob-69))); cq_f.push_back(440 * pow(2.0,0.083333 * (maxMIDI-69))); int nFFT = fft_f.size(); @@ -171,7 +171,7 @@ } void dictionaryMatrix(float* dm, float s_param) { - int binspersemitone = 3; // this must be 3 + int binspersemitone = nBPS; int minoctave = 0; // this must be 0 int maxoctave = 7; // this must be 7 // float s_param = 0.7; @@ -181,14 +181,14 @@ int maxMIDI = 21 + maxoctave * 12; // this includes one additional semitone! vector<float> cq_f; float oob = 1.0/binspersemitone; // one over binspersemitone - cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-69))); // 0.083333 is approx 1/12 - cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI+oob-69))); - for (int i = minMIDI + 1; i < maxMIDI; ++i) { - for (int k = -1; k < 2; ++k) { + // cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-69))); // 0.083333 is approx 1/12 + // cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI+oob-69))); + for (int i = minMIDI; i < maxMIDI; ++i) { + for (int k = 0; k < binspersemitone; ++k) { cq_f.push_back(440 * pow(2.0,0.083333333333 * (i+oob*k-69))); } } - cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-oob-69))); + // cq_f.push_back(440 * pow(2.0,0.083333 * (minMIDI-oob-69))); cq_f.push_back(440 * pow(2.0,0.083333 * (maxMIDI-69))); float curr_f;
--- a/chromamethods.h Mon Nov 15 09:23:49 2010 +0900 +++ b/chromamethods.h Mon Nov 15 11:01:56 2010 +0900 @@ -22,7 +22,7 @@ #include <vector> #include <string> -const int nBPS = 3; // bins per semitone +const int nBPS = 5; // bins per semitone const int nOctave = 7; const int nNote = nOctave * 12 * nBPS + 2 * (nBPS/2+1); // a core over all octaves, plus some overlap at top and bottom @@ -53,10 +53,6 @@ static const float basswindow[] = {0.001769, 0.015848, 0.043608, 0.084265, 0.136670, 0.199341, 0.270509, 0.348162, 0.430105, 0.514023, 0.597545, 0.678311, 0.754038, 0.822586, 0.882019, 0.930656, 0.967124, 0.990393, 0.999803, 0.995091, 0.976388, 0.944223, 0.899505, 0.843498, 0.777785, 0.704222, 0.624888, 0.542025, 0.457975, 0.375112, 0.295778, 0.222215, 0.156502, 0.100495, 0.055777, 0.023612, 0.004909, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000}; static const float treblewindow[] = {0.000350, 0.003144, 0.008717, 0.017037, 0.028058, 0.041719, 0.057942, 0.076638, 0.097701, 0.121014, 0.146447, 0.173856, 0.203090, 0.233984, 0.266366, 0.300054, 0.334860, 0.370590, 0.407044, 0.444018, 0.481304, 0.518696, 0.555982, 0.592956, 0.629410, 0.665140, 0.699946, 0.733634, 0.766016, 0.796910, 0.826144, 0.853553, 0.878986, 0.902299, 0.923362, 0.942058, 0.958281, 0.971942, 0.982963, 0.991283, 0.996856, 0.999650, 0.999650, 0.996856, 0.991283, 0.982963, 0.971942, 0.958281, 0.942058, 0.923362, 0.902299, 0.878986, 0.853553, 0.826144, 0.796910, 0.766016, 0.733634, 0.699946, 0.665140, 0.629410, 0.592956, 0.555982, 0.518696, 0.481304, 0.444018, 0.407044, 0.370590, 0.334860, 0.300054, 0.266366, 0.233984, 0.203090, 0.173856, 0.146447, 0.121014, 0.097701, 0.076638, 0.057942, 0.041719, 0.028058, 0.017037, 0.008717, 0.003144, 0.000350}; -static const float sinvalue = 0.866025404; -static const float cosvalue = -0.5; - - #include "nnls.h"