void Sound_playPart (Sound me, double tmin, double tmax,
int (*callback) (void *closure, int phase, double tmin, double tmax, double t), void *closure)
{
try {
long ifsamp = lround (1.0 / my dx), bestSampleRate = MelderAudio_getOutputBestSampleRate (ifsamp);
if (ifsamp == bestSampleRate) {
struct SoundPlay *thee = (struct SoundPlay *) & thePlayingSound;
double *fromLeft = my z [1], *fromRight = my ny > 1 ? my z [2] : NULL;
MelderAudio_stopPlaying (MelderAudio_IMPLICIT);
long i1, i2;
if ((thy numberOfSamples = Matrix_getWindowSamplesX (me, tmin, tmax, & i1, & i2)) < 1) return;
thy tmin = tmin;
thy tmax = tmax;
thy dt = my dx;
thy t1 = my x1;
thy callback = callback;
thy closure = closure;
thy silenceBefore = (long) (ifsamp * MelderAudio_getOutputSilenceBefore ());
thy silenceAfter = (long) (ifsamp * MelderAudio_getOutputSilenceAfter ());
int numberOfChannels = my ny;
NUMvector_free (thy buffer, 1); // just in case
thy buffer = NUMvector <short> (1, (i2 - i1 + 1 + thy silenceBefore + thy silenceAfter) * numberOfChannels);
thy i1 = i1;
thy i2 = i2;
short *to = thy buffer + thy silenceBefore * numberOfChannels;
if (numberOfChannels > 2) {
for (long i = i1; i <= i2; i ++) {
for (long chan = 1; chan <= my ny; chan ++) {
long value = (long) round (my z [chan] [i] * 32768.0);
* ++ to = value < -32768 ? -32768 : value > 32767 ? 32767 : value;
}
}
} else if (numberOfChannels == 2) {
for (long i = i1; i <= i2; i ++) {
long valueLeft = (long) round (fromLeft [i] * 32768.0);
* ++ to = valueLeft < -32768 ? -32768 : valueLeft > 32767 ? 32767 : valueLeft;
long valueRight = (long) round (fromRight [i] * 32768.0);
* ++ to = valueRight < -32768 ? -32768 : valueRight > 32767 ? 32767 : valueRight;
}
} else {
for (long i = i1; i <= i2; i ++) {
long value = (long) round (fromLeft [i] * 32768.0);
* ++ to = value < -32768 ? -32768 : value > 32767 ? 32767 : value;
}
}
if (thy callback) thy callback (thy closure, 1, tmin, tmax, tmin);
MelderAudio_play16 (thy buffer + 1, ifsamp,
thy silenceBefore + thy numberOfSamples + thy silenceAfter, numberOfChannels, melderPlayCallback, thee);
} else {
autoSound resampled = Sound_resample (me, bestSampleRate, 1);
Sound_playPart (resampled.peek(), tmin, tmax, callback, closure); // recursively
}
} catch (MelderError) {
Melder_throw (me, U": not played.");
}
}
autoFormant Sound_to_Formant_any (Sound me, double dt, int numberOfPoles, double maximumFrequency,
double halfdt_window, int which, double preemphasisFrequency, double safetyMargin)
{
double nyquist = 0.5 / my dx;
autoSound sound;
if (maximumFrequency <= 0.0 || fabs (maximumFrequency / nyquist - 1) < 1.0e-12) {
sound = Data_copy (me); // will be modified
} else {
sound = Sound_resample (me, maximumFrequency * 2, 50);
}
return Sound_to_Formant_any_inline (sound.peek(), dt, numberOfPoles, halfdt_window, which, preemphasisFrequency, safetyMargin);
}
autoSpectrum Spectrum_resample (Spectrum me, long numberOfFrequencies) {
try {
double newSamplingFrequency = (1 / my dx) * numberOfFrequencies / my nx;
// resample real and imaginary part !
autoSound thee = Sound_resample ((Sound) me, newSamplingFrequency, 50);
autoSpectrum him = Spectrum_create (my xmax, numberOfFrequencies);
NUMmatrix_copyElements<double> (thy z, his z, 1, 2, 1, numberOfFrequencies);
return him;
} catch (MelderError) {
Melder_throw (me, U": not resampled.");
}
}
PowerCepstrogram Sound_to_PowerCepstrogram (Sound me, double pitchFloor, double dt, double maximumFrequency, double preEmphasisFrequency) {
try {
// minimum analysis window has 3 periods of lowest pitch
double analysisWidth = 3 / pitchFloor;
double windowDuration = 2 * analysisWidth; /* gaussian window */
long nFrames;
// Convenience: analyse the whole sound into one Cepstrogram_frame
if (windowDuration > my dx * my nx) {
windowDuration = my dx * my nx;
}
double t1, samplingFrequency = 2 * maximumFrequency;
autoSound sound = Sound_resample (me, samplingFrequency, 50);
Sound_preEmphasis (sound.peek(), preEmphasisFrequency);
Sampled_shortTermAnalysis (me, windowDuration, dt, & nFrames, & t1);
autoSound sframe = Sound_createSimple (1, windowDuration, samplingFrequency);
autoSound window = Sound_createGaussian (windowDuration, samplingFrequency);
// find out the size of the FFT
long nfft = 2;
while (nfft < sframe -> nx) nfft *= 2;
long nq = nfft / 2 + 1;
double qmax = 0.5 * nfft / samplingFrequency, dq = qmax / (nq - 1);
autoPowerCepstrogram thee = PowerCepstrogram_create (my xmin, my xmax, nFrames, dt, t1, 0, qmax, nq, dq, 0);
autoMelderProgress progress (L"Cepstrogram analysis");
for (long iframe = 1; iframe <= nFrames; iframe++) {
double t = Sampled_indexToX (thee.peek(), iframe);
Sound_into_Sound (sound.peek(), sframe.peek(), t - windowDuration / 2);
Vector_subtractMean (sframe.peek());
Sounds_multiply (sframe.peek(), window.peek());
autoSpectrum spec = Sound_to_Spectrum (sframe.peek(), 1); // FFT yes
autoPowerCepstrum cepstrum = Spectrum_to_PowerCepstrum (spec.peek());
for (long i = 1; i <= nq; i++) {
thy z[i][iframe] = cepstrum -> z[1][i];
}
if ((iframe % 10) == 1) {
Melder_progress ((double) iframe / nFrames, L"PowerCepstrogram analysis of frame ",
Melder_integer (iframe), L" out of ", Melder_integer (nFrames), L".");
}
}
return thee.transfer();
} catch (MelderError) {
Melder_throw (me, ": no PowerCepstrogram created.");
}
}
Cepstrogram Sound_to_Cepstrogram (Sound me, double analysisWidth, double dt, double maximumFrequency) {
try {
double windowDuration = 2 * analysisWidth; /* gaussian window */
long nFrames;
// Convenience: analyse the whole sound into one Cepstrogram_frame
if (windowDuration > my dx * my nx) {
windowDuration = my dx * my nx;
}
double t1, samplingFrequency = 2 * maximumFrequency;
autoSound sound = Sound_resample (me, samplingFrequency, 50);
Sampled_shortTermAnalysis (me, windowDuration, dt, & nFrames, & t1);
autoSound sframe = Sound_createSimple (1, windowDuration, samplingFrequency);
autoSound window = Sound_createGaussian (windowDuration, samplingFrequency);
double qmin, qmax, dq, q1;
long nq;
{ // laziness: find out the proper dimensions
autoSpectrum spec = Sound_to_Spectrum (sframe.peek(), 1);
autoCepstrum cepstrum = Spectrum_to_Cepstrum (spec.peek());
qmin = cepstrum -> xmin; qmax = cepstrum -> xmax; dq = cepstrum -> dx;
q1 = cepstrum -> x1; nq = cepstrum -> nx;
}
autoCepstrogram thee = Cepstrogram_create (my xmin, my xmax, nFrames, dt, t1, qmin, qmax, nq, dq, q1);
autoMelderProgress progress (L"Cepstrogram analysis");
for (long iframe = 1; iframe <= nFrames; iframe++) {
double t = Sampled_indexToX (thee.peek(), iframe);
Sound_into_Sound (sound.peek(), sframe.peek(), t - windowDuration / 2);
Vector_subtractMean (sframe.peek());
Sounds_multiply (sframe.peek(), window.peek());
autoSpectrum spec = Sound_to_Spectrum (sframe.peek(), 1);
autoCepstrum cepstrum = Spectrum_to_Cepstrum (spec.peek());
for (long i = 1; i <= nq; i++) {
thy z[i][iframe] = cepstrum -> z[1][i];
}
if ((iframe % 10) == 1) {
Melder_progress ((double) iframe / nFrames, L"Cepstrogram analysis of frame ",
Melder_integer (iframe), L" out of ", Melder_integer (nFrames), L".");
}
}
return thee.transfer();
} catch (MelderError) {
Melder_throw (me, ": no Cepstrogram created.");
}
}
static autoSound synthesize_pulses_lpc (Manipulation me) {
try {
if (! my lpc) {
if (! my sound) Melder_throw (U"Missing original sound.");
autoSound sound10k = Sound_resample (my sound.get(), 10000.0, 50);
my lpc = Sound_to_LPC_burg (sound10k.get(), 20, 0.025, 0.01, 50.0);
}
if (! my pulses) Melder_throw (U"Missing pulses analysis.");
autoSound train = PointProcess_to_Sound_pulseTrain (my pulses.get(), 1.0 / my lpc -> samplingPeriod, 0.7, 0.05, 30);
train -> dx = my lpc -> samplingPeriod; // to be exact
Sound_PointProcess_fillVoiceless (train.get(), my pulses.get());
autoSound result = LPC_and_Sound_filter (my lpc.get(), train.get(), true);
NUMdeemphasize_f (result -> z [1], result -> nx, result -> dx, 50.0);
Vector_scale (result.get(), 0.99);
return result;
} catch (MelderError) {
Melder_throw (me, U": LPC synthesis not performed.");
}
}
Formant Sound_to_Formant_robust (Sound me, double dt_in, int numberOfPoles, double maximumFrequency,
double halfdt_window, double preEmphasisFrequency, double safetyMargin, double k, int itermax, double tol, int wantlocation)
{
double dt = dt_in > 0.0 ? dt_in : halfdt_window / 4.0;
double nyquist = 0.5 / my dx;
long predictionOrder = 2 * numberOfPoles;
try {
autoSound sound = NULL;
if (maximumFrequency <= 0.0 || fabs (maximumFrequency / nyquist - 1) < 1.0e-12) {
sound.reset (Data_copy (me)); // will be modified
} else {
sound.reset (Sound_resample (me, maximumFrequency * 2, 50));
}
autoLPC lpc = Sound_to_LPC_auto (sound.peek(), predictionOrder, halfdt_window, dt, preEmphasisFrequency);
autoLPC lpcr = LPC_and_Sound_to_LPC_robust (lpc.peek(), sound.peek(), halfdt_window, preEmphasisFrequency, k, itermax, tol, wantlocation);
autoFormant thee = LPC_to_Formant (lpcr.peek(), safetyMargin);
return thee.transfer();
} catch (MelderError) {
Melder_throw (me, ": no robust Formant created.");
}
}
/*
gain used as a constant amplitude multiplyer within a frame of duration my dx.
future alternative: convolve gain with a smoother.
*/
autoSound LPC_and_Sound_filter (LPC me, Sound thee, int useGain) {
try {
double xmin = my xmin > thy xmin ? my xmin : thy xmin;
double xmax = my xmax < thy xmax ? my xmax : thy xmax;
if (xmin >= xmax) {
Melder_throw (U"Domains of Sound [", thy xmin, U",", thy xmax, U"] and LPC [",
my xmin, U",", my xmax, U"] do not overlap.");
}
// resample sound if samplings don't match
autoSound source;
if (my samplingPeriod != thy dx) {
source = Sound_resample (thee, 1.0 / my samplingPeriod, 50);
thee = source.get(); // reference copy; remove at end
}
autoSound him = Data_copy (thee);
double *x = his z[1];
long ifirst = Sampled_xToHighIndex (thee, xmin);
long ilast = Sampled_xToLowIndex (thee, xmax);
for (long i = ifirst; i <= ilast; i++) {
double t = his x1 + (i - 1) * his dx; /* Sampled_indexToX (him, i) */
long iFrame = lround ( (t - my x1) / my dx + 1.0); /* Sampled_xToNearestIndex (me, t) */
if (iFrame < 1) {
continue;
}
if (iFrame > my nx) {
break;
}
double *a = my d_frames[iFrame].a;
long m = i > my d_frames[iFrame].nCoefficients ? my d_frames[iFrame].nCoefficients : i - 1;
for (long j = 1; j <= m; j++) {
x[i] -= a[j] * x[i - j];
}
}
// Make samples before first frame and after last frame zero.
for (long i = 1; i < ifirst; i++) {
x[i] = 0.0;
}
for (long i = ilast + 1; i <= his nx; i++) {
x[i] = 0.0;
}
if (useGain) {
for (long i = ifirst; i <= ilast; i++) {
double t = his x1 + (i - 1) * his dx; /* Sampled_indexToX (him, i) */
double riFrame = (t - my x1) / my dx + 1; /* Sampled_xToIndex (me, t); */
long iFrame = (long) floor (riFrame);
double phase = riFrame - iFrame;
if (iFrame < 0 || iFrame > my nx) {
x[i] = 0.0;
} else if (iFrame == 0) {
x[i] *= sqrt (my d_frames[1].gain) * phase;
} else if (iFrame == my nx) {
x[i] *= sqrt (my d_frames[my nx].gain) * (1.0 - phase);
} else x[i] *=
phase * sqrt (my d_frames[iFrame + 1].gain) + (1.0 - phase) * sqrt (my d_frames[iFrame].gain);
}
}
return him;
} catch (MelderError) {
Melder_throw (thee, U": not filtered.");
}
}
PowerCepstrogram Sound_to_PowerCepstrogram_hillenbrand (Sound me, double minimumPitch, double dt) {
try {
// minimum analysis window has 3 periods of lowest pitch
double analysisWidth = 3 / minimumPitch;
if (analysisWidth > my dx * my nx) {
analysisWidth = my dx * my nx;
}
double t1, samplingFrequency = 1 / my dx;
autoSound thee;
if (samplingFrequency > 30000) {
samplingFrequency = samplingFrequency / 2;
thee.reset (Sound_resample (me, samplingFrequency, 1));
} else {
thee.reset (Data_copy (me));
}
// pre-emphasis with fixed coefficient 0.9
for (long i = thy nx; i > 1; i--) {
thy z[1][i] -= 0.9 * thy z[1][i - 1];
}
long nosInWindow = analysisWidth * samplingFrequency, nFrames;
if (nosInWindow < 8) {
Melder_throw ("Analysis window too short.");
}
Sampled_shortTermAnalysis (thee.peek(), analysisWidth, dt, & nFrames, & t1);
autoNUMvector<double> hamming (1, nosInWindow);
for (long i = 1; i <= nosInWindow; i++) {
hamming[i] = 0.54 -0.46 * cos(2 * NUMpi * (i - 1) / (nosInWindow - 1));
}
long nfft = 8; // minimum possible
while (nfft < nosInWindow) { nfft *= 2; }
long nfftdiv2 = nfft / 2;
autoNUMvector<double> fftbuf (1, nfft); // "complex" array
autoNUMvector<double> spectrum (1, nfftdiv2 + 1); // +1 needed
autoNUMfft_Table fftTable;
NUMfft_Table_init (&fftTable, nfft); // sound to spectrum
double qmax = 0.5 * nfft / samplingFrequency, dq = qmax / (nfftdiv2 + 1);
autoPowerCepstrogram him = PowerCepstrogram_create (my xmin, my xmax, nFrames, dt, t1, 0, qmax, nfftdiv2+1, dq, 0);
autoMelderProgress progress (L"Cepstrogram analysis");
for (long iframe = 1; iframe <= nFrames; iframe++) {
double tbegin = t1 + (iframe - 1) * dt - analysisWidth / 2;
tbegin = tbegin < thy xmin ? thy xmin : tbegin;
long istart = Sampled_xToIndex (thee.peek(), tbegin);
istart = istart < 1 ? 1 : istart;
long iend = istart + nosInWindow - 1;
iend = iend > thy nx ? thy nx : iend;
for (long i = 1; i <= nosInWindow; i++) {
fftbuf[i] = thy z[1][istart + i - 1] * hamming[i];
}
for (long i = nosInWindow + 1; i <= nfft; i++) {
fftbuf[i] = 0;
}
NUMfft_forward (&fftTable, fftbuf.peek());
complexfftoutput_to_power (fftbuf.peek(), nfft, spectrum.peek(), true); // log10(|fft|^2)
// subtract average
double specmean = spectrum[1];
for (long i = 2; i <= nfftdiv2 + 1; i++) {
specmean += spectrum[i];
}
specmean /= nfftdiv2 + 1;
for (long i = 1; i <= nfftdiv2 + 1; i++) {
spectrum[i] -= specmean;
}
/*
* Here we diverge from Hillenbrand as he takes the fft of half of the spectral values.
* H. forgets that the actual spectrum has nfft/2+1 values. Thefore, we take the inverse
* transform because this keeps the number of samples a power of 2.
* At the same time this results in twice as much numbers in the quefrency domain, i.e. we end with nfft/2+1
* numbers while H. has only nfft/4!
*/
fftbuf[1] = spectrum[1];
for (long i = 2; i < nfftdiv2 + 1; i++) {
fftbuf[i+i-2] = spectrum[i];
fftbuf[i+i-1] = 0;
}
fftbuf[nfft] = spectrum[nfftdiv2 + 1];
NUMfft_backward (&fftTable, fftbuf.peek());
for (long i = 1; i <= nfftdiv2 + 1; i++) {
his z[i][iframe] = fftbuf[i] * fftbuf[i];
}
if ((iframe % 10) == 1) {
Melder_progress ((double) iframe / nFrames, L"Cepstrogram analysis of frame ",
Melder_integer (iframe), L" out of ", Melder_integer (nFrames), L".");
}
}
return him.transfer();
} catch (MelderError) {
Melder_throw (me, ": no Cepstrogram created.");
}
}
autoSound SpeechSynthesizer_to_Sound (SpeechSynthesizer me, const char32 *text, autoTextGrid *tg, autoTable *events) {
try {
int fsamp = espeak_Initialize (AUDIO_OUTPUT_SYNCHRONOUS, 0, nullptr, // 5000ms
espeakINITIALIZE_PHONEME_EVENTS|espeakINITIALIZE_PHONEME_IPA);
if (fsamp == -1) {
Melder_throw (U"Internal espeak error.");
}
int synth_flags = espeakCHARS_WCHAR;
if (my d_inputTextFormat == SpeechSynthesizer_INPUT_TAGGEDTEXT) {
synth_flags |= espeakSSML;
}
if (my d_inputTextFormat != SpeechSynthesizer_INPUT_TEXTONLY) {
synth_flags |= espeakPHONEMES;
}
option_phoneme_events = espeakINITIALIZE_PHONEME_EVENTS; // extern int option_phoneme_events;
if (my d_outputPhonemeCoding == SpeechSynthesizer_PHONEMECODINGS_IPA) {
option_phoneme_events |= espeakINITIALIZE_PHONEME_IPA;
}
espeak_SetParameter (espeakRATE, my d_wordsPerMinute, 0);
espeak_SetParameter (espeakPITCH, my d_pitchAdjustment, 0);
espeak_SetParameter (espeakRANGE, my d_pitchRange, 0);
const char32 *voiceLanguageCode = SpeechSynthesizer_getVoiceLanguageCodeFromName (me, my d_voiceLanguageName);
const char32 *voiceVariantCode = SpeechSynthesizer_getVoiceVariantCodeFromName (me, my d_voiceVariantName);
espeakdata_SetVoiceByName ((const char *) Melder_peek32to8 (voiceLanguageCode),
(const char *) Melder_peek32to8 (voiceVariantCode));
espeak_SetParameter (espeakWORDGAP, my d_wordgap * 100, 0); // espeak wordgap is in units of 10 ms
espeak_SetParameter (espeakCAPITALS, 0, 0);
espeak_SetParameter (espeakPUNCTUATION, espeakPUNCT_NONE, 0);
espeak_SetSynthCallback (synthCallback);
my d_events = Table_createWithColumnNames (0, U"time type type-t t-pos length a-pos sample id uniq");
#ifdef _WIN32
wchar_t *textW = Melder_peek32toW (text);
espeak_Synth (textW, wcslen (textW) + 1, 0, POS_CHARACTER, 0, synth_flags, nullptr, me);
#else
espeak_Synth (text, str32len (text) + 1, 0, POS_CHARACTER, 0, synth_flags, nullptr, me);
#endif
espeak_Terminate ();
autoSound thee = buffer_to_Sound (my d_wav, my d_numberOfSamples, my d_internalSamplingFrequency);
if (my d_samplingFrequency != my d_internalSamplingFrequency) {
thee = Sound_resample (thee.get(), my d_samplingFrequency, 50);
}
my d_numberOfSamples = 0; // re-use the wav-buffer
if (tg) {
double xmin = Table_getNumericValue_Assert (my d_events.get(), 1, 1);
if (xmin > thy xmin) {
xmin = thy xmin;
}
double xmax = Table_getNumericValue_Assert (my d_events.get(), my d_events -> rows.size, 1);
if (xmax < thy xmax) {
xmax = thy xmax;
}
autoTextGrid tg1 = Table_to_TextGrid (my d_events.get(), text, xmin, xmax);
*tg = TextGrid_extractPart (tg1.get(), thy xmin, thy xmax, 0);
}
if (events) {
Table_setEventTypeString (my d_events.get());
*events = my d_events.move();
}
my d_events.reset();
return thee;
} catch (MelderError) {
espeak_Terminate ();
Melder_throw (U"Text not played.");
}
}
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.");
}
}
