autoPitch SPINET_to_Pitch (SPINET me, double harmonicFallOffSlope, double ceiling, int maxnCandidates) {
try {
long nPointsPerOctave = 48;
double fmin = NUMerbToHertz (Sampled2_rowToY (me, 1));
double fmax = NUMerbToHertz (Sampled2_rowToY (me, my ny));
double fminl2 = NUMlog2 (fmin), fmaxl2 = NUMlog2 (fmax);
double points = (fmaxl2 - fminl2) * nPointsPerOctave;
double dfl2 = (fmaxl2 - fminl2) / (points - 1);
long nFrequencyPoints = (long) floor (points);
long maxHarmonic = (long) floor (fmax / fmin);
double maxStrength = 0.0, unvoicedCriterium = 0.45, maxPower = 0.0;
if (nFrequencyPoints < 2) {
Melder_throw (U"Frequency range too small.");
}
if (ceiling <= fmin) {
Melder_throw (U"Ceiling is smaller than centre frequency of lowest filter.");
}
autoPitch thee = Pitch_create (my xmin, my xmax, my nx, my dx, my x1, ceiling, maxnCandidates);
autoNUMvector<double> power (1, my nx);
autoNUMvector<double> pitch (1, nFrequencyPoints);
autoNUMvector<double> sumspec (1, nFrequencyPoints);
autoNUMvector<double> y (1, my ny);
autoNUMvector<double> yv2 (1, my ny);
autoNUMvector<double> fl2 (1, my ny);
// From ERB's to log (f)
for (long i = 1; i <= my ny; i++) {
double f = NUMerbToHertz (my y1 + (i - 1) * my dy);
fl2[i] = NUMlog2 (f);
}
// Determine global maximum power in frame
for (long j = 1; j <= my nx; j++) {
double p = 0.0;
for (long i = 1; i <= my ny; i++) {
p += my s[i][j];
}
if (p > maxPower) {
maxPower = p;
}
power[j] = p;
}
if (maxPower == 0.0) {
Melder_throw (U"No power");
}
for (long j = 1; j <= my nx; j++) {
Pitch_Frame pitchFrame = &thy frame[j];
pitchFrame -> intensity = power[j] / maxPower;
for (long i = 1; i <= my ny; i++) {
y[i] = my s[i][j];
}
NUMspline (fl2.peek(), y.peek(), my ny, 1e30, 1e30, yv2.peek());
for (long k = 1; k <= nFrequencyPoints; k++) {
double f = fminl2 + (k - 1) * dfl2;
NUMsplint (fl2.peek(), y.peek(), yv2.peek(), my ny, f, & pitch[k]);
sumspec[k] = 0.0;
}
// Formula (8): weighted harmonic summation.
for (long m = 1; m <= maxHarmonic; m++) {
double hm = 1 - harmonicFallOffSlope * NUMlog2 (m);
long kb = 1 + (long) floor (nPointsPerOctave * NUMlog2 (m));
for (long k = kb; k <= nFrequencyPoints; k++) {
if (pitch[k] > 0.0) {
sumspec[k - kb + 1] += pitch[k] * hm;
}
}
}
// into Pitch object
Pitch_Frame_init (pitchFrame, maxnCandidates);
pitchFrame -> nCandidates = 0; /* !!!!! */
Pitch_Frame_addPitch (pitchFrame, 0, 0, maxnCandidates); /* unvoiced */
for (long k = 2; k <= nFrequencyPoints - 1; k++) {
double y1 = sumspec[k - 1], y2 = sumspec[k], y3 = sumspec[k + 1];
if (y2 > y1 && y2 >= y3) {
double denum = y1 - 2.0 * y2 + y3, tmp = y3 - 4.0 * y2;
double x = dfl2 * (y1 - y3) / (2 * denum);
double f = pow (2.0, fminl2 + (k - 1) * dfl2 + x);
double strength = (2.0 * y1 * (4.0 * y2 + y3) - y1 * y1 - tmp * tmp) / (8.0 * denum);
if (strength > maxStrength) {
maxStrength = strength;
}
Pitch_Frame_addPitch (pitchFrame, f, strength, maxnCandidates);
}
}
}
// Scale the pitch strengths
for (long j = 1; j <= my nx; j++) {
double f0, localStrength;
Pitch_Frame_getPitch (&thy frame[j], &f0, &localStrength);
Pitch_Frame_resizeStrengths (&thy frame[j], localStrength / maxStrength, unvoicedCriterium);
}
return thee;
} catch (MelderError) {
Melder_throw (me, U": no Pitch created.");
}
}
Pitch Sound_to_Pitch_shs (Sound me, double timeStep, double minimumPitch,
double maximumFrequency, double ceiling, long maxnSubharmonics, long maxnCandidates,
double compressionFactor, long nPointsPerOctave) {
try {
double firstTime, newSamplingFrequency = 2 * maximumFrequency;
double windowDuration = 2 / minimumPitch, halfWindow = windowDuration / 2;
double atans = nPointsPerOctave * NUMlog2 (65.0 / 50.0) - 1;
// Number of speech samples in the downsampled signal in each frame:
// 100 for windowDuration == 0.04 and newSamplingFrequency == 2500
long nx = lround (windowDuration * newSamplingFrequency);
// The minimum number of points for the fft is 256.
long nfft = 1;
while ( (nfft *= 2) < nx || nfft <= 128) {
;
}
long nfft2 = nfft / 2 + 1;
double frameDuration = nfft / newSamplingFrequency;
double df = newSamplingFrequency / nfft;
// The number of points on the octave scale
double fminl2 = NUMlog2 (minimumPitch), fmaxl2 = NUMlog2 (maximumFrequency);
long nFrequencyPoints = (long) floor ((fmaxl2 - fminl2) * nPointsPerOctave);
double dfl2 = (fmaxl2 - fminl2) / (nFrequencyPoints - 1);
autoSound sound = Sound_resample (me, newSamplingFrequency, 50);
long numberOfFrames;
Sampled_shortTermAnalysis (sound.peek(), windowDuration, timeStep, &numberOfFrames, &firstTime);
autoSound frame = Sound_createSimple (1, frameDuration, newSamplingFrequency);
autoSound hamming = Sound_createHamming (nx / newSamplingFrequency, newSamplingFrequency);
autoPitch thee = Pitch_create (my xmin, my xmax, numberOfFrames, timeStep, firstTime,
ceiling, maxnCandidates);
autoNUMvector<double> cc (1, numberOfFrames);
autoNUMvector<double> specAmp (1, nfft2);
autoNUMvector<double> fl2 (1, nfft2);
autoNUMvector<double> yv2 (1, nfft2);
autoNUMvector<double> arctg (1, nFrequencyPoints);
autoNUMvector<double> al2 (1, nFrequencyPoints);
Melder_assert (frame->nx >= nx);
Melder_assert (hamming->nx == nx);
// Compute the absolute value of the globally largest amplitude w.r.t. the global mean.
double globalMean, globalPeak;
Sound_localMean (sound.peek(), sound -> xmin, sound -> xmax, &globalMean);
Sound_localPeak (sound.peek(), sound -> xmin, sound -> xmax, globalMean, &globalPeak);
/*
For the cubic spline interpolation we need the frequencies on an octave
scale, i.e., a log2 scale. All frequencies must be DIFFERENT, otherwise
the cubic spline interpolation will give corrupt results.
Because log2(f==0) is not defined, we use the heuristic: f[2]-f[1] == f[3]-f[2].
*/
for (long i = 2; i <= nfft2; i++) {
fl2[i] = NUMlog2 ( (i - 1) * df);
}
fl2[1] = 2 * fl2[2] - fl2[3];
// Calculate frequencies regularly spaced on a log2-scale and
// the frequency weighting function.
for (long i = 1; i <= nFrequencyPoints; i++) {
arctg[i] = 0.5 + atan (3 * (i - atans) / nPointsPerOctave) / NUMpi;
}
// Perform the analysis on all frames.
for (long i = 1; i <= numberOfFrames; i++) {
Pitch_Frame pitchFrame = &thy frame[i];
double hm = 1, f0, pitch_strength, localMean, localPeak;
double tmid = Sampled_indexToX (thee.peek(), i); /* The center of this frame */
long nx_tmp = frame -> nx;
// Copy a frame from the sound, apply a hamming window. Get local 'intensity'
frame -> nx = nx; /*begin vies */
Sound_into_Sound (sound.peek(), frame.peek(), tmid - halfWindow);
Sounds_multiply (frame.peek(), hamming.peek());
Sound_localMean (sound.peek(), tmid - 3 * halfWindow, tmid + 3 * halfWindow, &localMean);
Sound_localPeak (sound.peek(), tmid - halfWindow, tmid + halfWindow, localMean, &localPeak);
pitchFrame -> intensity = localPeak > globalPeak ? 1 : localPeak / globalPeak;
frame -> nx = nx_tmp; /* einde vies */
// Get the Fourier spectrum.
autoSpectrum spec = Sound_to_Spectrum (frame.peek(), 1);
Melder_assert (spec->nx == nfft2);
// From complex spectrum to amplitude spectrum.
for (long j = 1; j <= nfft2; j++) {
double rs = spec -> z[1][j], is = spec -> z[2][j];
specAmp[j] = sqrt (rs * rs + is * is);
}
// Enhance the peaks in the spectrum.
spec_enhance_SHS (specAmp.peek(), nfft2);
// Smooth the enhanced spectrum.
spec_smoooth_SHS (specAmp.peek(), nfft2);
// Go to a logarithmic scale and perform cubic spline interpolation to get
// spectral values for the increased number of frequency points.
NUMspline (fl2.peek(), specAmp.peek(), nfft2, 1e30, 1e30, yv2.peek());
for (long j = 1; j <= nFrequencyPoints; j++) {
double f = fminl2 + (j - 1) * dfl2;
NUMsplint (fl2.peek(), specAmp.peek(), yv2.peek(), nfft2, f, &al2[j]);
}
// Multiply by frequency selectivity of the auditory system.
for (long j = 1; j <= nFrequencyPoints; j++) al2[j] = al2[j] > 0 ?
al2[j] * arctg[j] : 0;
// The subharmonic summation. Shift spectra in octaves and sum.
Pitch_Frame_init (pitchFrame, maxnCandidates);
autoNUMvector<double> sumspec (1, nFrequencyPoints);
pitchFrame -> nCandidates = 0; /* !!!!! */
for (long m = 1; m <= maxnSubharmonics + 1; m++) {
long kb = 1 + (long) floor (nPointsPerOctave * NUMlog2 (m));
for (long k = kb; k <= nFrequencyPoints; k++) {
sumspec[k - kb + 1] += al2[k] * hm;
}
hm *= compressionFactor;
}
// First register the voiceless candidate (always present).
Pitch_Frame_addPitch (pitchFrame, 0, 0, maxnCandidates);
/*
Get the best local estimates for the pitch as the maxima of the
subharmonic sum spectrum by parabolic interpolation on three points:
The formula for a parabole with a maximum is:
y(x) = a - b (x - c)^2 with a, b, c >= 0
The three points are (-x, y1), (0, y2) and (x, y3).
The solution for a (the maximum) and c (the position) is:
a = (2 y1 (4 y2 + y3) - y1^2 - (y3 - 4 y2)^2)/( 8 (y1 - 2 y2 + y3)
c = dx (y1 - y3) / (2 (y1 - 2 y2 + y3))
(b = (2 y2 - y1 - y3) / (2 dx^2) )
*/
for (long k = 2; k <= nFrequencyPoints - 1; k++) {
double y1 = sumspec[k - 1], y2 = sumspec[k], y3 = sumspec[k + 1];
if (y2 > y1 && y2 >= y3) {
double denum = y1 - 2 * y2 + y3, tmp = y3 - 4 * y2;
double x = dfl2 * (y1 - y3) / (2 * denum);
double f = pow (2, fminl2 + (k - 1) * dfl2 + x);
double strength = (2 * y1 * (4 * y2 + y3) - y1 * y1 - tmp * tmp) / (8 * denum);
Pitch_Frame_addPitch (pitchFrame, f, strength, maxnCandidates);
}
}
/*
Check whether f0 corresponds to an actual periodicity T = 1 / f0:
correlate two signal periods of duration T, one starting at the
middle of the interval and one starting T seconds before.
If there is periodicity the correlation coefficient should be high.
However, some sounds do not show any regularity, or very low
frequency and regularity, and nevertheless have a definite
pitch, e.g. Shepard sounds.
*/
Pitch_Frame_getPitch (pitchFrame, &f0, &pitch_strength);
if (f0 > 0) {
cc[i] = Sound_correlateParts (sound.peek(), tmid - 1.0 / f0, tmid, 1.0 / f0);
}
}
// Base V/UV decision on correlation coefficients.
// Resize the pitch strengths w.r.t. the cc.
double vuvCriterium = 0.52;
for (long i = 1; i <= numberOfFrames; i++) {
Pitch_Frame_resizeStrengths (& thy frame[i], cc[i], vuvCriterium);
}
return thee.transfer();
} catch (MelderError) {
Melder_throw (me, U": no Pitch (shs) created.");
}
}
