autoSpectrum Spectrum_compressFrequencyDomain (Spectrum me, double fmax, long interpolationDepth, int freqscale, int method) { try { double fdomain = my xmax - my xmin, factor = fdomain / fmax ; //long numberOfFrequencies = 1.0 + fmax / my dx; // keep dx the same, otherwise the "duration" changes double xmax = my xmax / factor; long numberOfFrequencies = (long) floor (my nx / factor); // keep dx the same, otherwise the "duration" changes autoSpectrum thee = Spectrum_create (xmax, numberOfFrequencies); thy z[1][1] = my z[1][1]; thy z[2][1] = my z[2][1]; double df = freqscale == 1 ? factor * my dx : log10 (fdomain) / (numberOfFrequencies - 1); for (long i = 2; i <= numberOfFrequencies; i++) { double f = my xmin + (freqscale == 1 ? (i - 1) * df : pow (10.0, (i - 1) * df)); double x, y, index = (f - my x1) / my dx + 1; if (index > my nx) { break; } if (method == 1) { x = NUM_interpolate_sinc (my z[1], my nx, index, interpolationDepth); y = NUM_interpolate_sinc (my z[2], my nx, index, interpolationDepth); } else { x = NUMundefined; // ppgb: better than data from random memory y = NUMundefined; } thy z[1][i] = x; thy z[2][i] = y; } return thee; } catch (MelderError) { Melder_throw (me, U": not compressed."); } }
autoSpectrum Spectrum_shiftFrequencies (Spectrum me, double shiftBy, double newMaximumFrequency, long interpolationDepth) { try { double xmax = my xmax; long numberOfFrequencies = my nx; if (newMaximumFrequency != 0) { numberOfFrequencies = (long) floor (newMaximumFrequency / my dx) + 1; xmax = newMaximumFrequency; } autoSpectrum thee = Spectrum_create (xmax, numberOfFrequencies); // shiftBy >= 0 for (long i = 1; i <= thy nx; i++) { double thyf = thy x1 + (i - 1) * thy dx; double myf = thyf - shiftBy; if (myf >= my xmin && myf <= my xmax) { double index = Sampled_xToIndex (me, myf); thy z[1][i] = NUM_interpolate_sinc (my z[1], my nx, index, interpolationDepth); thy z[2][i] = NUM_interpolate_sinc (my z[2], my nx, index, interpolationDepth); } } // Make imaginary part of first and last sample zero // so Spectrum_to_Sound uses FFT if numberOfSamples was power of 2! double amp = sqrt (thy z[1][1] * thy z[1][1] + thy z[2][1] * thy z[2][1]); thy z[1][1] = amp; thy z[2][1] = 0; amp = sqrt (thy z[1][thy nx] * thy z[1][thy nx] + thy z[2][thy nx] * thy z[2][thy nx]); thy z[1][thy nx] = amp; thy z[2][thy nx] = 0; return thee; } catch (MelderError) { Melder_throw (me, U": not shifted."); } }
Sound Sound_resample (Sound me, double samplingFrequency, long precision) { double upfactor = samplingFrequency * my dx; if (fabs (upfactor - 2) < 1e-6) return Sound_upsample (me); if (fabs (upfactor - 1) < 1e-6) return Data_copy (me); try { long numberOfSamples = lround ((my xmax - my xmin) * samplingFrequency); if (numberOfSamples < 1) Melder_throw (U"The resampled Sound would have no samples."); autoSound filtered = NULL; if (upfactor < 1.0) { // need anti-aliasing filter? long nfft = 1, antiTurnAround = 1000; while (nfft < my nx + antiTurnAround * 2) nfft *= 2; autoNUMvector <double> data (1, nfft); filtered.reset (Sound_create (my ny, my xmin, my xmax, my nx, my dx, my x1)); for (long channel = 1; channel <= my ny; channel ++) { for (long i = 1; i <= nfft; i ++) { data [i] = 0; } NUMvector_copyElements (my z [channel], & data [antiTurnAround], 1, my nx); NUMrealft (data.peek(), nfft, 1); // go to the frequency domain for (long i = (long) floor (upfactor * nfft); i <= nfft; i ++) { data [i] = 0; // filter away high frequencies } data [2] = 0.0; NUMrealft (data.peek(), nfft, -1); // return to the time domain double factor = 1.0 / nfft; double *to = filtered -> z [channel]; for (long i = 1; i <= my nx; i ++) { to [i] = data [i + antiTurnAround] * factor; } } me = filtered.peek(); // reference copy; remove at end } autoSound thee = Sound_create (my ny, my xmin, my xmax, numberOfSamples, 1.0 / samplingFrequency, 0.5 * (my xmin + my xmax - (numberOfSamples - 1) / samplingFrequency)); for (long channel = 1; channel <= my ny; channel ++) { double *from = my z [channel]; double *to = thy z [channel]; if (precision <= 1) { for (long i = 1; i <= numberOfSamples; i ++) { double x = Sampled_indexToX (thee.peek(), i); double index = Sampled_xToIndex (me, x); long leftSample = (long) floor (index); double fraction = index - leftSample; to [i] = leftSample < 1 || leftSample >= my nx ? 0.0 : (1 - fraction) * from [leftSample] + fraction * from [leftSample + 1]; } } else { for (long i = 1; i <= numberOfSamples; i ++) { double x = Sampled_indexToX (thee.peek(), i); double index = Sampled_xToIndex (me, x); to [i] = NUM_interpolate_sinc (my z [channel], my nx, index, precision); } } } return thee.transfer(); } catch (MelderError) { Melder_throw (me, U": not resampled."); } }
// // Vector_getValueAtX () returns the average of all the interpolated channels. // double Vector_getValueAtX (Vector me, double x, long ilevel, int interpolation) { double leftEdge = my x1 - 0.5 * my dx, rightEdge = leftEdge + my nx * my dx; if (x < leftEdge || x > rightEdge) return NUMundefined; if (ilevel > Vector_CHANNEL_AVERAGE) { Melder_assert (ilevel <= my ny); return NUM_interpolate_sinc (my z [ilevel], my nx, Sampled_xToIndex (me, x), interpolation == Vector_VALUE_INTERPOLATION_SINC70 ? NUM_VALUE_INTERPOLATE_SINC70 : interpolation == Vector_VALUE_INTERPOLATION_SINC700 ? NUM_VALUE_INTERPOLATE_SINC700 : interpolation); } double sum = 0.0; for (long channel = 1; channel <= my ny; channel ++) { sum += NUM_interpolate_sinc (my z [channel], my nx, Sampled_xToIndex (me, x), interpolation == Vector_VALUE_INTERPOLATION_SINC70 ? NUM_VALUE_INTERPOLATE_SINC70 : interpolation == Vector_VALUE_INTERPOLATION_SINC700 ? NUM_VALUE_INTERPOLATE_SINC700 : interpolation); } return sum / my ny; }
void ParamCurve_draw (ParamCurve me, Graphics g, double t1, double t2, double dt, double x1, double x2, double y1, double y2, int garnish) { if (t2 <= t1) { double tx1 = my x -> x1; double ty1 = my y -> x1; double tx2 = my x -> x1 + (my x -> nx - 1) * my x -> dx; double ty2 = my y -> x1 + (my y -> nx - 1) * my y -> dx; t1 = tx1 > ty1 ? tx1 : ty1; t2 = tx2 < ty2 ? tx2 : ty2; } if (x2 <= x1) Matrix_getWindowExtrema (my x, 0, 0, 1, 1, & x1, & x2); if (x1 == x2) { x1 -= 1.0; x2 += 1.0; } if (y2 <= y1) Matrix_getWindowExtrema (my y, 0, 0, 1, 1, & y1, & y2); if (y1 == y2) { y1 -= 1.0; y2 += 1.0; } if (dt <= 0.0) dt = my x -> dx < my y -> dx ? my x -> dx : my y -> dx; long numberOfPoints = (long) ceil ((t2 - t1) / dt) + 1; if (numberOfPoints > 0) { autoNUMvector <double> x (1, numberOfPoints); autoNUMvector <double> y (1, numberOfPoints); for (long i = 1; i <= numberOfPoints; i ++) { double t = i == numberOfPoints ? t2 : t1 + (i - 1) * dt; double index = my x -> f_xToIndex (t); x [i] = NUM_interpolate_sinc (my x -> z [1], my x -> nx, index, 50); index = my y -> f_xToIndex (t); y [i] = NUM_interpolate_sinc (my y -> z [1], my y -> nx, index, 50); } Graphics_setWindow (g, x1, x2, y1, y2); Graphics_setInner (g); Graphics_polyline (g, numberOfPoints, & x [1], & y [1]); Graphics_unsetInner (g); } if (garnish) { Graphics_drawInnerBox (g); Graphics_marksBottom (g, 2, 1, 1, 0); Graphics_marksLeft (g, 2, 1, 1, 0); } }
static autoSpectrum Spectrum_shiftFrequencies2 (Spectrum me, double shiftBy, bool changeMaximumFrequency) { try { double xmax = my xmax; long numberOfFrequencies = my nx, interpolationDepth = 50; if (changeMaximumFrequency) { xmax += shiftBy; numberOfFrequencies += (xmax - my xmax) / my dx; } autoSpectrum thee = Spectrum_create (xmax, numberOfFrequencies); // shiftBy >= 0 for (long i = 1; i <= thy nx; i++) { double thyf = thy x1 + (i - 1) * thy dx; double myf = thyf - shiftBy; if (myf >= my xmin && myf <= my xmax) { double index = Sampled_xToIndex (me, myf); thy z[1][i] = NUM_interpolate_sinc (my z[1], my nx, index, interpolationDepth); thy z[2][i] = NUM_interpolate_sinc (my z[2], my nx, index, interpolationDepth); } } return thee; } catch (MelderError) { Melder_throw (me, U": not shifted."); } }
static void Sound_into_PitchFrame (Sound me, Pitch_Frame pitchFrame, double t, double minimumPitch, int maxnCandidates, int method, double voicingThreshold, double octaveCost, NUMfft_Table fftTable, double dt_window, long nsamp_window, long halfnsamp_window, long maximumLag, long nsampFFT, long nsamp_period, long halfnsamp_period, long brent_ixmax, long brent_depth, double globalPeak, double **frame, double *ac, double *window, double *windowR, double *r, long *imax, double *localMean) { double localPeak; long leftSample = Sampled_xToLowIndex (me, t), rightSample = leftSample + 1; long startSample, endSample; for (long channel = 1; channel <= my ny; channel ++) { /* * Compute the local mean; look one longest period to both sides. */ startSample = rightSample - nsamp_period; endSample = leftSample + nsamp_period; Melder_assert (startSample >= 1); Melder_assert (endSample <= my nx); localMean [channel] = 0.0; for (long i = startSample; i <= endSample; i ++) { localMean [channel] += my z [channel] [i]; } localMean [channel] /= 2 * nsamp_period; /* * Copy a window to a frame and subtract the local mean. * We are going to kill the DC component before windowing. */ startSample = rightSample - halfnsamp_window; endSample = leftSample + halfnsamp_window; Melder_assert (startSample >= 1); Melder_assert (endSample <= my nx); if (method < FCC_NORMAL) { for (long j = 1, i = startSample; j <= nsamp_window; j ++) frame [channel] [j] = (my z [channel] [i ++] - localMean [channel]) * window [j]; for (long j = nsamp_window + 1; j <= nsampFFT; j ++) frame [channel] [j] = 0.0; } else { for (long j = 1, i = startSample; j <= nsamp_window; j ++) frame [channel] [j] = my z [channel] [i ++] - localMean [channel]; } } /* * Compute the local peak; look half a longest period to both sides. */ localPeak = 0.0; if ((startSample = halfnsamp_window + 1 - halfnsamp_period) < 1) startSample = 1; if ((endSample = halfnsamp_window + halfnsamp_period) > nsamp_window) endSample = nsamp_window; for (long channel = 1; channel <= my ny; channel ++) { for (long j = startSample; j <= endSample; j ++) { double value = fabs (frame [channel] [j]); if (value > localPeak) localPeak = value; } } pitchFrame->intensity = localPeak > globalPeak ? 1.0 : localPeak / globalPeak; /* * Compute the correlation into the array 'r'. */ if (method >= FCC_NORMAL) { double startTime = t - 0.5 * (1.0 / minimumPitch + dt_window); long localSpan = maximumLag + nsamp_window, localMaximumLag, offset; if ((startSample = Sampled_xToLowIndex (me, startTime)) < 1) startSample = 1; if (localSpan > my nx + 1 - startSample) localSpan = my nx + 1 - startSample; localMaximumLag = localSpan - nsamp_window; offset = startSample - 1; double sumx2 = 0; // sum of squares for (long channel = 1; channel <= my ny; channel ++) { double *amp = my z [channel] + offset; for (long i = 1; i <= nsamp_window; i ++) { double x = amp [i] - localMean [channel]; sumx2 += x * x; } } double sumy2 = sumx2; // at zero lag, these are still equal r [0] = 1.0; for (long i = 1; i <= localMaximumLag; i ++) { double product = 0.0; for (long channel = 1; channel <= my ny; channel ++) { double *amp = my z [channel] + offset; double y0 = amp [i] - localMean [channel]; double yZ = amp [i + nsamp_window] - localMean [channel]; sumy2 += yZ * yZ - y0 * y0; for (long j = 1; j <= nsamp_window; j ++) { double x = amp [j] - localMean [channel]; double y = amp [i + j] - localMean [channel]; product += x * y; } } r [- i] = r [i] = product / sqrt (sumx2 * sumy2); } } else { /* * The FFT of the autocorrelation is the power spectrum. */ for (long i = 1; i <= nsampFFT; i ++) { ac [i] = 0.0; } for (long channel = 1; channel <= my ny; channel ++) { NUMfft_forward (fftTable, frame [channel]); // complex spectrum ac [1] += frame [channel] [1] * frame [channel] [1]; // DC component for (long i = 2; i < nsampFFT; i += 2) { ac [i] += frame [channel] [i] * frame [channel] [i] + frame [channel] [i+1] * frame [channel] [i+1]; // power spectrum } ac [nsampFFT] += frame [channel] [nsampFFT] * frame [channel] [nsampFFT]; // Nyquist frequency } NUMfft_backward (fftTable, ac); /* Autocorrelation. */ /* * Normalize the autocorrelation to the value with zero lag, * and divide it by the normalized autocorrelation of the window. */ r [0] = 1.0; for (long i = 1; i <= brent_ixmax; i ++) r [- i] = r [i] = ac [i + 1] / (ac [1] * windowR [i + 1]); } /* * Register the first candidate, which is always present: voicelessness. */ pitchFrame->nCandidates = 1; pitchFrame->candidate[1].frequency = 0.0; // voiceless: always present pitchFrame->candidate[1].strength = 0.0; /* * Shortcut: absolute silence is always voiceless. * We are done for this frame. */ if (localPeak == 0) return; /* * Find the strongest maxima of the correlation of this frame, * and register them as candidates. */ imax [1] = 0; for (long i = 2; i < maximumLag && i < brent_ixmax; i ++) if (r [i] > 0.5 * voicingThreshold && // not too unvoiced? r [i] > r [i-1] && r [i] >= r [i+1]) // maximum? { int place = 0; /* * Use parabolic interpolation for first estimate of frequency, * and sin(x)/x interpolation to compute the strength of this frequency. */ double dr = 0.5 * (r [i+1] - r [i-1]), d2r = 2 * r [i] - r [i-1] - r [i+1]; double frequencyOfMaximum = 1 / my dx / (i + dr / d2r); long offset = - brent_ixmax - 1; double strengthOfMaximum = /* method & 1 ? */ NUM_interpolate_sinc (& r [offset], brent_ixmax - offset, 1 / my dx / frequencyOfMaximum - offset, 30) /* : r [i] + 0.5 * dr * dr / d2r */; /* High values due to short windows are to be reflected around 1. */ if (strengthOfMaximum > 1.0) strengthOfMaximum = 1.0 / strengthOfMaximum; /* * Find a place for this maximum. */ if (pitchFrame->nCandidates < maxnCandidates) { // is there still a free place? place = ++ pitchFrame->nCandidates; } else { /* Try the place of the weakest candidate so far. */ double weakest = 2; for (int iweak = 2; iweak <= maxnCandidates; iweak ++) { /* High frequencies are to be favoured */ /* if we want to analyze a perfectly periodic signal correctly. */ double localStrength = pitchFrame->candidate[iweak].strength - octaveCost * NUMlog2 (minimumPitch / pitchFrame->candidate[iweak].frequency); if (localStrength < weakest) { weakest = localStrength; place = iweak; } } /* If this maximum is weaker than the weakest candidate so far, give it no place. */ if (strengthOfMaximum - octaveCost * NUMlog2 (minimumPitch / frequencyOfMaximum) <= weakest) place = 0; } if (place) { // have we found a place for this candidate? pitchFrame->candidate[place].frequency = frequencyOfMaximum; pitchFrame->candidate[place].strength = strengthOfMaximum; imax [place] = i; } } /* * Second pass: for extra precision, maximize sin(x)/x interpolation ('sinc'). */ for (long i = 2; i <= pitchFrame->nCandidates; i ++) { if (method != AC_HANNING || pitchFrame->candidate[i].frequency > 0.0 / my dx) { double xmid, ymid; long offset = - brent_ixmax - 1; ymid = NUMimproveMaximum (& r [offset], brent_ixmax - offset, imax [i] - offset, pitchFrame->candidate[i].frequency > 0.3 / my dx ? NUM_PEAK_INTERPOLATE_SINC700 : brent_depth, & xmid); xmid += offset; pitchFrame->candidate[i].frequency = 1.0 / my dx / xmid; if (ymid > 1.0) ymid = 1.0 / ymid; pitchFrame->candidate[i].strength = ymid; } } }
Pitch Sound_to_Pitch_any (Sound me, double dt, /*timeStepStradygy related*/ double minimumPitch, /*Pitch settings realted*/ double periodsPerWindow, /*kTimeSoundAnalysisEditor_pitch_analysisMethod related*/ int maxnCandidates, int method, /*method related*/ double silenceThreshold, double voicingThreshold, double octaveCost, double octaveJumpCost, double voicedUnvoicedCost, double ceiling) { NUMfft_Table fftTable = NUMfft_Table_create(); double duration, t1; double dt_window; /* Window length in seconds. */ long nsamp_window, halfnsamp_window; /* Number of samples per window. */ long nFrames, minimumLag, maximumLag; long iframe, nsampFFT; double interpolation_depth; long nsamp_period, halfnsamp_period; /* Number of samples in longest period. */ long brent_ixmax, brent_depth; double brent_accuracy; /* Obsolete. */ double globalPeak; if (maxnCandidates < 2 || method < AC_HANNING && method > FCC_ACCURATE) { std::cout<<"Error: maxnCandidates: "<<maxnCandidates<<" method: "<<method<<"."<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 13. 69"<<std::endl; return NULL; } if (maxnCandidates < ceiling / minimumPitch) maxnCandidates = ceiling / minimumPitch; if (dt <= 0.0) dt = periodsPerWindow / minimumPitch / 4.0; /* e.g. 3 periods, 75 Hz: 10 milliseconds. */ switch (method) { case AC_HANNING: brent_depth = NUM_PEAK_INTERPOLATE_SINC70; brent_accuracy = 1e-7; interpolation_depth = 0.5; break; case AC_GAUSS: periodsPerWindow *= 2; /* Because Gaussian window is twice as long. */ brent_depth = NUM_PEAK_INTERPOLATE_SINC700; brent_accuracy = 1e-11; interpolation_depth = 0.25; /* Because Gaussian window is twice as long. */ break; case FCC_NORMAL: brent_depth = NUM_PEAK_INTERPOLATE_SINC70; brent_accuracy = 1e-7; interpolation_depth = 1.0; break; case FCC_ACCURATE: brent_depth = NUM_PEAK_INTERPOLATE_SINC700; brent_accuracy = 1e-11; interpolation_depth = 1.0; break; } duration = my dx * my nx; if (minimumPitch < periodsPerWindow / duration) { std::cout<<"To analyse this Sound, minimum pitch must not be less than "<< periodsPerWindow / duration<<" Hz."<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 31.103"<<std::endl; return NULL; } /* * Determine the number of samples in the longest period. * We need this to compute the local mean of the sound (looking one period in both directions), * and to compute the local peak of the sound (looking half a period in both directions). */ nsamp_period = floor(1 / my dx / minimumPitch); halfnsamp_period = nsamp_period / 2 + 1; if (ceiling > 0.5 / my dx) ceiling = 0.5 / my dx; // Determine window length in seconds and in samples. dt_window = periodsPerWindow / minimumPitch; nsamp_window = floor (dt_window / my dx); halfnsamp_window = nsamp_window / 2 - 1; if (halfnsamp_window < 2){ std::cout<<"Analysis window too short."<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 31.123"<<std::endl; return NULL; } nsamp_window = halfnsamp_window * 2; // Determine the minimum and maximum lags. minimumLag = floor (1 / my dx / ceiling); if (minimumLag < 2) minimumLag = 2; maximumLag = floor (nsamp_window / periodsPerWindow) + 2; if (maximumLag > nsamp_window) maximumLag = nsamp_window; /* * Determine the number of frames. * Fit as many frames as possible symmetrically in the total duration. * We do this even for the forward cross-correlation method, * because that allows us to compare the two methods. */ if(!Sampled_shortTermAnalysis (me, method >= FCC_NORMAL ? 1 / minimumPitch + dt_window : dt_window, dt, & nFrames, & t1)){ std::cout<<"The pitch analysis would give zero pitch frames."<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 31.142"<<std::endl; return NULL; } // Create the resulting pitch contour. Pitch thee = Pitch_create (my xmin, my xmax, nFrames, dt, t1, ceiling, maxnCandidates); // Compute the global absolute peak for determination of silence threshold. globalPeak = 0.0; for (long channel = 1; channel <= my ny; channel ++) { double mean = 0.0; for (long i = 1; i <= my nx; i ++) { mean += my z [channel] [i]; } mean /= my nx; for (long i = 1; i <= my nx; i ++) { double value = fabs (my z [channel] [i] - mean); if (value > globalPeak) globalPeak = value; } } if (globalPeak == 0.0) return thee; double **frame, *ac, *window, *windowR; if (method >= FCC_NORMAL) { /* For cross-correlation analysis. */ // Create buffer for cross-correlation analysis. frame = (double **)malloc(sizeof(double *) * (my ny + 1)); for(long i = 1; i <= my ny; ++ i){ frame[i] = (double *)malloc(sizeof(double) * (nsamp_window + 1)); for(long j = 1; j <= nsamp_window; ++ j) frame[i][j] = 0.0; } /****frame.reset (1, my ny, 1, nsamp_window);****/ brent_ixmax = nsamp_window * interpolation_depth; } else { /* For autocorrelation analysis. */ /* * Compute the number of samples needed for doing FFT. * To avoid edge effects, we have to append zeroes to the window. * The maximum lag considered for maxima is maximumLag. * The maximum lag used in interpolation is nsamp_window * interpolation_depth. */ nsampFFT = 1; while (nsampFFT < nsamp_window * (1 + interpolation_depth)) nsampFFT *= 2; // Create buffers for autocorrelation analysis. frame = (double **)malloc(sizeof(double *) * (my ny + 1)); for(long i = 1; i <= my ny; ++ i){ frame [i] = (double *)malloc(sizeof(double) * (nsampFFT + 1)); for(long j = 0; j <= nsampFFT; ++ j) frame[i][j] = 0.0; } /****frame.reset (1, my ny, 1, nsampFFT);****/ window = (double *)malloc(sizeof(double) * (nsamp_window + 1)); for(long i = 0; i <= nsamp_window; ++ i) window[i] = 0.0; /****window.reset (1, nsamp_window);****/ windowR = (double *)malloc(sizeof(double) * (nsampFFT + 1)); ac = (double *)malloc(sizeof(double) * (nsampFFT + 1)); for(long i = 0; i <= nsampFFT; ++ i) windowR[i] = ac[i] = 0.0; /****windowR.reset (1, nsampFFT); ac.reset (1, nsampFFT); ****/ NUMfft_Table_init (fftTable, nsampFFT); /* * A Gaussian or Hanning window is applied against phase effects. * The Hanning window is 2 to 5 dB better for 3 periods/window. * The Gaussian window is 25 to 29 dB better for 6 periods/window. */ if (method == AC_GAUSS) { /* Gaussian window. */ double imid = 0.5 * (nsamp_window + 1), edge = exp (-12.0); for (long i = 1; i <= nsamp_window; i ++) window[i] = (exp(-48.0*(i-imid)*(i-imid) / (nsamp_window + 1) / (nsamp_window + 1)) - edge) / (1 - edge); } else { /* Hanning window*/ for (long i = 1; i <= nsamp_window; i ++) window [i] = 0.5 - 0.5 * cos (i * 2 * NUMpi / (nsamp_window + 1)); } // Compute the normalized autocorrelation of the window. for (long i = 1; i <= nsamp_window; i ++) windowR [i] = window [i]; NUMfft_forward (fftTable, windowR); windowR [1] *= windowR [1]; // DC component for (long i = 2; i < nsampFFT; i += 2) { windowR [i] = windowR [i] * windowR [i] + windowR [i+1] * windowR [i+1]; windowR [i + 1] = 0.0; // power spectrum: square and zero } windowR [nsampFFT] *= windowR [nsampFFT]; // Nyquist frequency NUMfft_backward (fftTable, windowR); // autocorrelation for (long i = 2; i <= nsamp_window; i ++) windowR [i] /= windowR [1]; // normalize windowR [1] = 1.0; // normalize brent_ixmax = nsamp_window * interpolation_depth; } double *r = (double *) malloc( sizeof(double) * (2 * (nsamp_window + 1) + 1) ); r += nsamp_window + 1; //make "r" become a symetrical vectr long *imax = (long *) malloc( sizeof(long) * (maxnCandidates + 1)); double *localMean = (double *) malloc( sizeof(double) * (my ny + 1)); for(iframe = 1; iframe <= nFrames; iframe ++){ Pitch_Frame pitchFrame = & thy frame[iframe]; double t = thy x1 + (iframe - 1) *(thy dx), localPeak; long leftSample = (long) floor((t - my x1) / my dx) + 1; long rightSample = leftSample + 1; long startSample, endSample; for(long channel = 1; channel <= my ny; ++ channel){ //Compute the local mean; look one longest period to both sides. startSample = rightSample - nsamp_period; endSample = leftSample + nsamp_period; if ( startSample < 0 ) { std::cout<<"StartSample < 1"<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 31"<<std::endl; return NULL; } if (endSample > my nx){ std::cout<<"EndSample > my nx"<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 31.262"<<std::endl; return NULL; } localMean[channel] = 0.0; for (long i = startSample; i <= endSample; i ++) { localMean[channel] += my z[channel][i]; } localMean[channel] /= 2 * nsamp_period; // Copy a window to a frame and subtract the local mean. We are going to kill the DC component before windowing. startSample = rightSample - halfnsamp_window; endSample = leftSample + halfnsamp_window; if ( startSample < 1 ) { std::cout<<"StartSample < 1"<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 31.281"<<std::endl; return NULL; } if (endSample > my nx){ std::cout<<"EndSample > my nx"<<std::endl; std::cout<<"Sound_to_Pitch.cpp: Line 31.287"<<std::endl; return NULL; } if (method < FCC_NORMAL) { for (long j = 1, i = startSample; j <= nsamp_window; j ++) frame [channel] [j] = (my z [channel] [i ++] - localMean [channel]) * window [j]; for (long j = nsamp_window + 1; j <= nsampFFT; j ++) frame [channel] [j] = 0.0; } else { for (long j = 1, i = startSample; j <= nsamp_window; j ++) frame [channel] [j] = my z [channel] [i ++] - localMean [channel]; } } // Compute the local peak; look half a longest period to both sides. localPeak = 0.0; if ((startSample = halfnsamp_window + 1 - halfnsamp_period) < 1) startSample = 1; if ((endSample = halfnsamp_window + halfnsamp_period) > nsamp_window) endSample = nsamp_window; for (long channel = 1; channel <= my ny; channel ++) { for (long j = startSample; j <= endSample; j ++) { double value = fabs (frame [channel] [j]); if (value > localPeak) localPeak = value; } } pitchFrame->intensity = localPeak > globalPeak ? 1.0 : localPeak / globalPeak; // Compute the correlation into the array 'r'. if (method >= FCC_NORMAL) { double startTime = t - 0.5 * (1.0 / minimumPitch + dt_window); long localSpan = maximumLag + nsamp_window, localMaximumLag, offset; if ((startSample = (long) floor ((startTime - my x1) / my dx)) + 1 < 1) startSample = 1; if (localSpan > my nx + 1 - startSample) localSpan = my nx + 1 - startSample; localMaximumLag = localSpan - nsamp_window; offset = startSample - 1; double sumx2 = 0; /* Sum of squares. */ for (long channel = 1; channel <= my ny; channel ++) { ///channel = 1; channel <= my ny double *amp = my z[channel] + offset; for (long i = 1; i <= nsamp_window; i ++) { ///i = 1; i <= nsamp_window double x = amp[i] - localMean[channel]; sumx2 += x * x; } } double sumy2 = sumx2; /* At zero lag, these are still equal. */ r[0] = 1.0; for (long i = 1; i <= localMaximumLag; i ++) { double product = 0.0; for (long channel = 1; channel <= my ny; channel ++) { ///channel = 1; channel <= my ny double *amp = my z[channel] + offset; double y0 = amp[i] - localMean[channel]; double yZ = amp[i + nsamp_window] - localMean[channel]; sumy2 += yZ * yZ - y0 * y0; for (long j = 1; j <= nsamp_window; j ++) { ///j = 1; j <= nsamp_window double x = amp[j] - localMean[channel]; double y = amp[i + j] - localMean[channel]; product += x * y; } } r[- i] = r[i] = product / sqrt (sumx2 * sumy2); } } else { // The FFT of the autocorrelation is the power spectrum. for (long i = 1; i <= nsampFFT; i ++) ac [i] = 0.0; for (long channel = 1; channel <= my ny; channel ++) { NUMfft_forward (fftTable, frame [channel]); /* Complex spectrum. */ ac [1] += frame [channel] [1] * frame [channel] [1]; /* DC component. */ for (long i = 2; i < nsampFFT; i += 2) { ac [i] += frame [channel] [i] * frame [channel] [i] + frame [channel] [i+1] * frame [channel] [i+1]; /* Power spectrum. */ } ac [nsampFFT] += frame [channel] [nsampFFT] * frame [channel] [nsampFFT]; /* Nyquist frequency. */ } NUMfft_backward (fftTable, ac); /* Autocorrelation. */ /* * Normalize the autocorrelation to the value with zero lag, * and divide it by the normalized autocorrelation of the window. */ r [0] = 1.0; for (long i = 1; i <= brent_ixmax; i ++) r [- i] = r [i] = ac [i + 1] / (ac [1] * windowR [i + 1]); } // Create (too much) space for candidates Pitch_Frame_init (pitchFrame, maxnCandidates); // Register the first candidate, which is always present: voicelessness. pitchFrame->nCandidates = 1; pitchFrame->candidate[1].frequency = 0.0; /* Voiceless: always present. */ pitchFrame->candidate[1].strength = 0.0; /* * Shortcut: absolute silence is always voiceless. * Go to next frame. */ if (localPeak == 0) continue; /* * Find the strongest maxima of the correlation of this frame, * and register them as candidates. */ imax[1] = 0; for (long i = 2; i < maximumLag && i < brent_ixmax; i ++) if (r[i] > 0.5 * voicingThreshold && /* Not too unvoiced? */ r[i] > r[i-1] && r[i] >= r[i+1]) /* Maximum ? */ { int place = 0; // Use parabolic interpolation for first estimate of frequency,and sin(x)/x interpolation to compute the strength of this frequency. double dr = 0.5 * (r[i+1] - r[i-1]); double d2r = 2 * r[i] - r[i-1] - r[i+1]; double frequencyOfMaximum = 1 / my dx / (i + dr / d2r); long offset = - brent_ixmax - 1; double strengthOfMaximum = /* method & 1 ? */ NUM_interpolate_sinc (& r[offset], brent_ixmax - offset, 1 / my dx / frequencyOfMaximum - offset, 30) /* : r [i] + 0.5 * dr * dr / d2r */; /* High values due to short windows are to be reflected around 1. */ if (strengthOfMaximum > 1.0) strengthOfMaximum = 1.0 / strengthOfMaximum; // Find a place for this maximum. if (pitchFrame->nCandidates < thy maxnCandidates) { /* Is there still a free place? */ place = ++ pitchFrame->nCandidates; } else { /* Try the place of the weakest candidate so far. */ double weakest = 2; for (int iweak = 2; iweak <= thy maxnCandidates; iweak ++) { //iweak = 2; iweak <= thy maxnCandidates; /* High frequencies are to be favoured */ /* if we want to analyze a perfectly periodic signal correctly. */ double localStrength = pitchFrame->candidate[iweak].strength - octaveCost * NUMlog2 (minimumPitch / pitchFrame->candidate[iweak].frequency); if (localStrength < weakest) { weakest = localStrength; place = iweak; } } /* If this maximum is weaker than the weakest candidate so far, give it no place. */ if (strengthOfMaximum - octaveCost * NUMlog2 (minimumPitch / frequencyOfMaximum) <= weakest) place = 0; } if (place) { /* Have we found a place for this candidate? */ pitchFrame->candidate[place].frequency = frequencyOfMaximum; pitchFrame->candidate[place].strength = strengthOfMaximum; imax [place] = i; } } // Second pass: for extra precision, maximize sin(x)/x interpolation ('sinc'). for (long i = 2; i <= pitchFrame->nCandidates; i ++) { if (method != AC_HANNING || pitchFrame->candidate[i].frequency > 0.0 / my dx) { double xmid, ymid; long offset = - brent_ixmax - 1; ymid = NUMimproveMaximum (& r[offset], brent_ixmax - offset, imax[i] - offset, pitchFrame->candidate[i].frequency > 0.3 / my dx ? NUM_PEAK_INTERPOLATE_SINC700 : brent_depth, & xmid); xmid += offset; pitchFrame->candidate[i].frequency = 1.0 / my dx / xmid; if (ymid > 1.0) ymid = 1.0 / ymid; pitchFrame->candidate[i].strength = ymid; } } } /* Next frame. */ Pitch_pathFinder (thee, silenceThreshold, voicingThreshold,octaveCost, octaveJumpCost, voicedUnvoicedCost, ceiling, false); //false: Melder_debug == 31 ? true : false Melder_debug 31: Pitch analysis: formant pulling on return thee; }