../src/pitch/pitchmcomb.h \
../src/pitch/pitchyin.h \
../src/pitch/pitchyinfft.h \
+ ../src/pitch/pitchyinfast.h \
../src/pitch/pitchschmitt.h \
../src/pitch/pitchfcomb.h \
../src/pitch/pitchspecacf.h \
--- /dev/null
+#! /usr/bin/env python
+# -*- coding: utf8 -*-
+
+""" Pure python implementation of the sum of squared difference
+
+ sqd_yin: original sum of squared difference [0]
+ d_t(tau) = x ⊗ kernel
+ sqd_yinfast: sum of squared diff using complex domain [0]
+ sqd_yinfftslow: tappered squared diff [1]
+ sqd_yinfft: modified squared diff using complex domain [1]
+
+[0]:http://audition.ens.fr/adc/pdf/2002_JASA_YIN.pdf
+[1]:https://aubio.org/phd/
+"""
+
+import sys
+import numpy as np
+import matplotlib.pyplot as plt
+
+def sqd_yin(samples):
+ """ compute original sum of squared difference
+
+ Brute-force computation (cost o(N**2), slow)."""
+ B = len(samples)
+ W = B//2
+ yin = np.zeros(W)
+ for j in range(W):
+ for tau in range(1, W):
+ yin[tau] += (samples[j] - samples[j+tau])**2
+ return yin
+
+def sqd_yinfast(samples):
+ """ compute approximate sum of squared difference
+
+ Using complex convolution (fast, cost o(n*log(n)) )"""
+ # yin_t(tau) = (r_t(0) + r_(t+tau)(0)) - 2r_t(tau)
+ B = len(samples)
+ W = B//2
+ yin = np.zeros(W)
+ sqdiff = np.zeros(W)
+ kernel = np.zeros(B)
+ # compute r_(t+tau)(0)
+ squares = samples**2
+ for tau in range(W):
+ sqdiff[tau] = squares[tau:tau+W].sum()
+ # add r_t(0)
+ sqdiff += sqdiff[0]
+ # compute r_t(tau) using kernel convolution in complex domain
+ samples_fft = np.fft.fft(samples)
+ kernel[1:W+1] = samples[W-1::-1] # first half, reversed
+ kernel_fft = np.fft.fft(kernel)
+ r_t_tau = np.fft.ifft(samples_fft * kernel_fft).real[W:]
+ # compute yin_t(tau)
+ yin = sqdiff - 2 * r_t_tau
+ return yin
+
+def sqd_yintapered(samples):
+ """ compute tappered sum of squared difference
+
+ Brute-force computation (cost o(N**2), slow)."""
+ B = len(samples)
+ W = B//2
+ yin = np.zeros(W)
+ for tau in range(1, W):
+ for j in range(W - tau):
+ yin[tau] += (samples[j] - samples[j+tau])**2
+ return yin
+
+def sqd_yinfft(samples):
+ """ compute yinfft modified sum of squared differences
+
+ Very fast, improved performance in transients.
+
+ FIXME: biased."""
+ B = len(samples)
+ W = B//2
+ yin = np.zeros(W)
+ def hanningz(W):
+ return .5 * (1. - np.cos(2. * np.pi * np.arange(W) / W))
+ #win = np.ones(B)
+ win = hanningz(B)
+ sqrmag = np.zeros(B)
+ fftout = np.fft.fft(win*samples)
+ sqrmag[0] = fftout[0].real**2
+ for l in range(1, W):
+ sqrmag[l] = fftout[l].real**2 + fftout[l].imag**2
+ sqrmag[B-l] = sqrmag[l]
+ sqrmag[W] = fftout[W].real**2
+ fftout = np.fft.fft(sqrmag)
+ sqrsum = 2.*sqrmag[:W + 1].sum()
+ yin[0] = 0
+ yin[1:] = sqrsum - fftout.real[1:W]
+ return yin / B
+
+def cumdiff(yin):
+ """ compute the cumulative mean normalized difference """
+ W = len(yin)
+ yin[0] = 1.
+ cumsum = 0.
+ for tau in range(1, W):
+ cumsum += yin[tau]
+ if cumsum != 0:
+ yin[tau] *= tau/cumsum
+ else:
+ yin[tau] = 1
+ return yin
+
+def compute_all(x):
+ import time
+ now = time.time()
+
+ yin = sqd_yin(x)
+ t1 = time.time()
+ print ("yin took %.2fms" % ((t1-now) * 1000.))
+
+ yinfast = sqd_yinfast(x)
+ t2 = time.time()
+ print ("yinfast took: %.2fms" % ((t2-t1) * 1000.))
+
+ yintapered = sqd_yintapered(x)
+ t3 = time.time()
+ print ("yintapered took: %.2fms" % ((t3-t2) * 1000.))
+
+ yinfft = sqd_yinfft(x)
+ t4 = time.time()
+ print ("yinfft took: %.2fms" % ((t4-t3) * 1000.))
+
+ return yin, yinfast, yintapered, yinfft
+
+def plot_all(yin, yinfast, yintapered, yinfft):
+ fig, axes = plt.subplots(nrows=2, ncols=2, sharex=True, sharey='col')
+
+ axes[0, 0].plot(yin, label='yin')
+ axes[0, 0].plot(yintapered, label='yintapered')
+ axes[0, 0].set_ylim(bottom=0)
+ axes[0, 0].legend()
+ axes[1, 0].plot(yinfast, '-', label='yinfast')
+ axes[1, 0].plot(yinfft, label='yinfft')
+ axes[1, 0].legend()
+
+ axes[0, 1].plot(cumdiff(yin), label='yin')
+ axes[0, 1].plot(cumdiff(yintapered), label='yin tapered')
+ axes[0, 1].set_ylim(bottom=0)
+ axes[0, 1].legend()
+ axes[1, 1].plot(cumdiff(yinfast), '-', label='yinfast')
+ axes[1, 1].plot(cumdiff(yinfft), label='yinfft')
+ axes[1, 1].legend()
+
+ fig.tight_layout()
+
+testfreqs = [441., 800., 10000., 40.]
+
+if len(sys.argv) > 1:
+ testfreqs = map(float,sys.argv[1:])
+
+for f in testfreqs:
+ print ("Comparing yin implementations for sine wave at %.fHz" % f)
+ samplerate = 44100.
+ win_s = 4096
+
+ x = np.cos(2.*np.pi * np.arange(win_s) * f / samplerate)
+
+ n_times = 1#00
+ for n in range(n_times):
+ yin, yinfast, yinfftslow, yinfft = compute_all(x)
+ if 0: # plot difference
+ plt.plot(yin-yinfast)
+ plt.tight_layout()
+ plt.show()
+ if 1:
+ plt.plot(yinfftslow-yinfft)
+ plt.tight_layout()
+ plt.show()
+ plot_all(yin, yinfast, yinfftslow, yinfft)
+plt.show()
'pitchspecacf',
'pitchyin',
'pitchyinfft',
+ 'pitchyinfast',
'sink',
'sink_apple_audio',
'sink_sndfile',
#print 'len(pitches), cut:', len(pitches), cut
#print 'median errors: ', median(errors), 'median pitches: ', median(pitches)
-pitch_algorithms = [ "default", "yinfft", "yin", "schmitt", "mcomb", "fcomb" , "specacf" ]
-pitch_algorithms = [ "default", "yinfft", "yin", "schmitt", "mcomb", "fcomb" ]
+pitch_algorithms = [ "default", "yinfft", "yin", "yinfast", "schmitt", "mcomb", "fcomb" , "specacf" ]
+pitch_algorithms = [ "default", "yinfft", "yin", "yinfast", "schmitt", "mcomb", "fcomb" ]
#freqs = [ 27.5, 55., 110., 220., 440., 880., 1760., 3520. ]
freqs = [ 110., 220., 440., 880., 1760., 3520. ]
#include "pitch/pitchmcomb.h"
#include "pitch/pitchyin.h"
#include "pitch/pitchyinfft.h"
+#include "pitch/pitchyinfast.h"
#include "pitch/pitchschmitt.h"
#include "pitch/pitchfcomb.h"
#include "pitch/pitchspecacf.h"
#include "pitch/pitchfcomb.h"
#include "pitch/pitchschmitt.h"
#include "pitch/pitchyinfft.h"
+#include "pitch/pitchyinfast.h"
#include "pitch/pitchspecacf.h"
#include "pitch/pitch.h"
aubio_pitcht_schmitt, /**< `schmitt`, Schmitt trigger */
aubio_pitcht_fcomb, /**< `fcomb`, Fast comb filter */
aubio_pitcht_yinfft, /**< `yinfft`, Spectral YIN */
+ aubio_pitcht_yinfast, /**< `yinfast`, YIN fast */
aubio_pitcht_specacf, /**< `specacf`, Spectral autocorrelation */
aubio_pitcht_default
= aubio_pitcht_yinfft, /**< `default` */
static void aubio_pitch_do_schmitt (aubio_pitch_t * p, const fvec_t * ibuf, fvec_t * obuf);
static void aubio_pitch_do_fcomb (aubio_pitch_t * p, const fvec_t * ibuf, fvec_t * obuf);
static void aubio_pitch_do_yinfft (aubio_pitch_t * p, const fvec_t * ibuf, fvec_t * obuf);
+static void aubio_pitch_do_yinfast (aubio_pitch_t * p, const fvec_t * ibuf, fvec_t * obuf);
static void aubio_pitch_do_specacf (aubio_pitch_t * p, const fvec_t * ibuf, fvec_t * obuf);
/* internal functions for frequency conversion */
}
if (strcmp (pitch_mode, "mcomb") == 0)
pitch_type = aubio_pitcht_mcomb;
+ else if (strcmp (pitch_mode, "yinfast") == 0)
+ pitch_type = aubio_pitcht_yinfast;
else if (strcmp (pitch_mode, "yinfft") == 0)
pitch_type = aubio_pitcht_yinfft;
else if (strcmp (pitch_mode, "yin") == 0)
p->conf_cb = (aubio_pitch_get_conf_t)aubio_pitchyinfft_get_confidence;
aubio_pitchyinfft_set_tolerance (p->p_object, 0.85);
break;
+ case aubio_pitcht_yinfast:
+ p->buf = new_fvec (bufsize);
+ p->p_object = new_aubio_pitchyinfast (bufsize);
+ if (!p->p_object) goto beach;
+ p->detect_cb = aubio_pitch_do_yinfast;
+ p->conf_cb = (aubio_pitch_get_conf_t)aubio_pitchyinfast_get_confidence;
+ aubio_pitchyinfast_set_tolerance (p->p_object, 0.15);
+ break;
case aubio_pitcht_specacf:
p->buf = new_fvec (bufsize);
p->p_object = new_aubio_pitchspecacf (bufsize);
del_fvec (p->buf);
del_aubio_pitchyinfft (p->p_object);
break;
+ case aubio_pitcht_yinfast:
+ del_fvec (p->buf);
+ del_aubio_pitchyinfast (p->p_object);
+ break;
case aubio_pitcht_specacf:
del_fvec (p->buf);
del_aubio_pitchspecacf (p->p_object);
case aubio_pitcht_yinfft:
aubio_pitchyinfft_set_tolerance (p->p_object, tol);
break;
+ case aubio_pitcht_yinfast:
+ aubio_pitchyinfast_set_tolerance (p->p_object, tol);
+ break;
default:
break;
}
case aubio_pitcht_yinfft:
tolerance = aubio_pitchyinfft_get_tolerance (p->p_object);
break;
+ case aubio_pitcht_yinfast:
+ tolerance = aubio_pitchyinfast_get_tolerance (p->p_object);
+ break;
default:
break;
}
}
void
+aubio_pitch_do_yinfast (aubio_pitch_t * p, const fvec_t * ibuf, fvec_t * obuf)
+{
+ smpl_t pitch = 0.;
+ aubio_pitch_slideblock (p, ibuf);
+ aubio_pitchyinfast_do (p->p_object, p->buf, obuf);
+ pitch = obuf->data[0];
+ if (pitch > 0) {
+ pitch = p->samplerate / (pitch + 0.);
+ } else {
+ pitch = 0.;
+ }
+ obuf->data[0] = pitch;
+}
+
+void
aubio_pitch_do_specacf (aubio_pitch_t * p, const fvec_t * ibuf, fvec_t * out)
{
smpl_t pitch = 0., period;
see http://recherche.ircam.fr/equipes/pcm/pub/people/cheveign.html
+ \b \p yinfast : Yinfast algorithm
+
+ This algorithm is equivalent to the YIN algorithm, but computed in the
+ spectral domain for efficiency. See also `python/demos/demo_yin_compare.py`.
+
\b \p yinfft : Yinfft algorithm
This algorithm was derived from the YIN algorithm. In this implementation, a
--- /dev/null
+/*
+ Copyright (C) 2003-2017 Paul Brossier <piem@aubio.org>
+
+ This file is part of aubio.
+
+ aubio is free software: you can redistribute it and/or modify
+ it under the terms of the GNU General Public License as published by
+ the Free Software Foundation, either version 3 of the License, or
+ (at your option) any later version.
+
+ aubio is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with aubio. If not, see <http://www.gnu.org/licenses/>.
+
+*/
+
+/* This algorithm was developed by A. de Cheveigné and H. Kawahara and
+ * published in:
+ *
+ * de Cheveigné, A., Kawahara, H. (2002) "YIN, a fundamental frequency
+ * estimator for speech and music", J. Acoust. Soc. Am. 111, 1917-1930.
+ *
+ * see http://recherche.ircam.fr/equipes/pcm/pub/people/cheveign.html
+ */
+
+#include "aubio_priv.h"
+#include "fvec.h"
+#include "mathutils.h"
+#include "cvec.h"
+#include "spectral/fft.h"
+#include "pitch/pitchyinfast.h"
+
+struct _aubio_pitchyinfast_t
+{
+ fvec_t *yin;
+ smpl_t tol;
+ smpl_t confidence;
+ fvec_t *tmpdata;
+ fvec_t *sqdiff;
+ fvec_t *kernel;
+ fvec_t *samples_fft;
+ fvec_t *kernel_fft;
+ aubio_fft_t *fft;
+};
+
+aubio_pitchyinfast_t *
+new_aubio_pitchyinfast (uint_t bufsize)
+{
+ aubio_pitchyinfast_t *o = AUBIO_NEW (aubio_pitchyinfast_t);
+ o->yin = new_fvec (bufsize / 2);
+ o->tmpdata = new_fvec (bufsize);
+ o->sqdiff = new_fvec (bufsize / 2);
+ o->kernel = new_fvec (bufsize);
+ o->samples_fft = new_fvec (bufsize);
+ o->kernel_fft = new_fvec (bufsize);
+ o->fft = new_aubio_fft (bufsize);
+ o->tol = 0.15;
+ return o;
+}
+
+void
+del_aubio_pitchyinfast (aubio_pitchyinfast_t * o)
+{
+ del_fvec (o->yin);
+ del_fvec (o->tmpdata);
+ del_fvec (o->sqdiff);
+ del_fvec (o->kernel);
+ del_fvec (o->samples_fft);
+ del_fvec (o->kernel_fft);
+ del_aubio_fft (o->fft);
+ AUBIO_FREE (o);
+}
+
+/* all the above in one */
+void
+aubio_pitchyinfast_do (aubio_pitchyinfast_t * o, const fvec_t * input, fvec_t * out)
+{
+ const smpl_t tol = o->tol;
+ fvec_t* yin = o->yin;
+ const uint_t length = yin->length;
+ uint_t B = o->tmpdata->length;
+ uint_t W = o->yin->length; // B / 2
+ fvec_t tmp_slice, kernel_ptr;
+ smpl_t *yin_data = yin->data;
+ uint_t tau;
+ sint_t period;
+ smpl_t tmp2 = 0.;
+
+ // compute r_t(0) + r_t+tau(0)
+ {
+ fvec_t *squares = o->tmpdata;
+ fvec_weighted_copy(input, input, squares);
+#if 0
+ for (tau = 0; tau < W; tau++) {
+ tmp_slice.data = squares->data + tau;
+ tmp_slice.length = W;
+ o->sqdiff->data[tau] = fvec_sum(&tmp_slice);
+ }
+#else
+ tmp_slice.data = squares->data;
+ tmp_slice.length = W;
+ o->sqdiff->data[0] = fvec_sum(&tmp_slice);
+ for (tau = 1; tau < W; tau++) {
+ o->sqdiff->data[tau] = o->sqdiff->data[tau-1];
+ o->sqdiff->data[tau] -= squares->data[tau-1];
+ o->sqdiff->data[tau] += squares->data[W+tau-1];
+ }
+#endif
+ fvec_add(o->sqdiff, o->sqdiff->data[0]);
+ }
+ // compute r_t(tau) = -2.*ifft(fft(samples)*fft(samples[W-1::-1]))
+ {
+ fvec_t *compmul = o->tmpdata;
+ fvec_t *rt_of_tau = o->samples_fft;
+ aubio_fft_do_complex(o->fft, input, o->samples_fft);
+ // build kernel, take a copy of first half of samples
+ tmp_slice.data = input->data;
+ tmp_slice.length = W;
+ kernel_ptr.data = o->kernel->data + 1;
+ kernel_ptr.length = W;
+ fvec_copy(&tmp_slice, &kernel_ptr);
+ // reverse them
+ fvec_rev(&kernel_ptr);
+ // compute fft(kernel)
+ aubio_fft_do_complex(o->fft, o->kernel, o->kernel_fft);
+ // compute complex product
+ compmul->data[0] = o->kernel_fft->data[0] * o->samples_fft->data[0];
+ for (tau = 1; tau < W; tau++) {
+ compmul->data[tau] = o->kernel_fft->data[tau] * o->samples_fft->data[tau];
+ compmul->data[tau] -= o->kernel_fft->data[B-tau] * o->samples_fft->data[B-tau];
+ }
+ compmul->data[W] = o->kernel_fft->data[W] * o->samples_fft->data[W];
+ for (tau = 1; tau < W; tau++) {
+ compmul->data[B-tau] = o->kernel_fft->data[B-tau] * o->samples_fft->data[tau];
+ compmul->data[B-tau] += o->kernel_fft->data[tau] * o->samples_fft->data[B-tau];
+ }
+ // compute inverse fft
+ aubio_fft_rdo_complex(o->fft, compmul, rt_of_tau);
+ // compute square difference r_t(tau) = sqdiff - 2 * r_t_tau[W-1:-1]
+ for (tau = 0; tau < W; tau++) {
+ yin_data[tau] = o->sqdiff->data[tau] - 2. * rt_of_tau->data[tau+W];
+ }
+ }
+
+ // now build yin and look for first minimum
+ fvec_set_all(out, 0.);
+ yin_data[0] = 1.;
+ for (tau = 1; tau < length; tau++) {
+ tmp2 += yin_data[tau];
+ if (tmp2 != 0) {
+ yin->data[tau] *= tau / tmp2;
+ } else {
+ yin->data[tau] = 1.;
+ }
+ period = tau - 3;
+ if (tau > 4 && (yin_data[period] < tol) &&
+ (yin_data[period] < yin_data[period + 1])) {
+ out->data[0] = fvec_quadratic_peak_pos (yin, period);
+ goto beach;
+ }
+ }
+ out->data[0] = fvec_quadratic_peak_pos (yin, fvec_min_elem (yin) );
+beach:
+ return;
+}
+
+smpl_t
+aubio_pitchyinfast_get_confidence (aubio_pitchyinfast_t * o) {
+ o->confidence = 1. - fvec_min (o->yin);
+ return o->confidence;
+}
+
+uint_t
+aubio_pitchyinfast_set_tolerance (aubio_pitchyinfast_t * o, smpl_t tol)
+{
+ o->tol = tol;
+ return 0;
+}
+
+smpl_t
+aubio_pitchyinfast_get_tolerance (aubio_pitchyinfast_t * o)
+{
+ return o->tol;
+}
--- /dev/null
+/*
+ Copyright (C) 2003-2017 Paul Brossier <piem@aubio.org>
+
+ This file is part of aubio.
+
+ aubio is free software: you can redistribute it and/or modify
+ it under the terms of the GNU General Public License as published by
+ the Free Software Foundation, either version 3 of the License, or
+ (at your option) any later version.
+
+ aubio is distributed in the hope that it will be useful,
+ but WITHOUT ANY WARRANTY; without even the implied warranty of
+ MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ GNU General Public License for more details.
+
+ You should have received a copy of the GNU General Public License
+ along with aubio. If not, see <http://www.gnu.org/licenses/>.
+
+*/
+
+/** \file
+
+ Pitch detection using YIN algorithm (fast implementation)
+
+ This algorithm was developed by A. de Cheveigne and H. Kawahara and
+ published in:
+
+ De Cheveigné, A., Kawahara, H. (2002) "YIN, a fundamental frequency
+ estimator for speech and music", J. Acoust. Soc. Am. 111, 1917-1930.
+
+ This implementation compute the autocorrelation function using time domain
+ convolution computed in the spectral domain.
+
+ see http://recherche.ircam.fr/equipes/pcm/pub/people/cheveign.html
+ http://recherche.ircam.fr/equipes/pcm/cheveign/ps/2002_JASA_YIN_proof.pdf
+
+*/
+
+#ifndef AUBIO_PITCHYINFAST_H
+#define AUBIO_PITCHYINFAST_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/** pitch detection object */
+typedef struct _aubio_pitchyinfast_t aubio_pitchyinfast_t;
+
+/** creation of the pitch detection object
+
+ \param buf_size size of the input buffer to analyse
+
+*/
+aubio_pitchyinfast_t *new_aubio_pitchyinfast (uint_t buf_size);
+
+/** deletion of the pitch detection object
+
+ \param o pitch detection object as returned by new_aubio_pitchyin()
+
+*/
+void del_aubio_pitchyinfast (aubio_pitchyinfast_t * o);
+
+/** execute pitch detection an input buffer
+
+ \param o pitch detection object as returned by new_aubio_pitchyin()
+ \param samples_in input signal vector (length as specified at creation time)
+ \param cands_out pitch period candidates, in samples
+
+*/
+void aubio_pitchyinfast_do (aubio_pitchyinfast_t * o, const fvec_t * samples_in, fvec_t * cands_out);
+
+
+/** set tolerance parameter for YIN algorithm
+
+ \param o YIN pitch detection object
+ \param tol tolerance parameter for minima selection [default 0.15]
+
+*/
+uint_t aubio_pitchyinfast_set_tolerance (aubio_pitchyinfast_t * o, smpl_t tol);
+
+/** get tolerance parameter for YIN algorithm
+
+ \param o YIN pitch detection object
+ \return tolerance parameter for minima selection [default 0.15]
+
+*/
+smpl_t aubio_pitchyinfast_get_tolerance (aubio_pitchyinfast_t * o);
+
+/** get current confidence of YIN algorithm
+
+ \param o YIN pitch detection object
+ \return confidence parameter
+
+*/
+smpl_t aubio_pitchyinfast_get_confidence (aubio_pitchyinfast_t * o);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* AUBIO_PITCHYINFAST_H */
+
projects / aubio.git / commitdiff