Chris@152
|
1 /* -*- c-basic-offset: 4 indent-tabs-mode: nil -*- vi:set ts=8 sts=4 sw=4: */
|
Chris@152
|
2
|
Chris@152
|
3 /*
|
Chris@152
|
4 Sonic Visualiser
|
Chris@152
|
5 An audio file viewer and annotation editor.
|
Chris@152
|
6 Centre for Digital Music, Queen Mary, University of London.
|
Chris@152
|
7 This file copyright 2006 Chris Cannam.
|
Chris@152
|
8
|
Chris@152
|
9 This program is free software; you can redistribute it and/or
|
Chris@152
|
10 modify it under the terms of the GNU General Public License as
|
Chris@152
|
11 published by the Free Software Foundation; either version 2 of the
|
Chris@152
|
12 License, or (at your option) any later version. See the file
|
Chris@152
|
13 COPYING included with this distribution for more information.
|
Chris@152
|
14 */
|
Chris@152
|
15
|
Chris@152
|
16 #include "FFTModel.h"
|
Chris@152
|
17 #include "DenseTimeValueModel.h"
|
Chris@152
|
18
|
Chris@183
|
19 #include "base/Profiler.h"
|
Chris@275
|
20 #include "base/Pitch.h"
|
Chris@183
|
21
|
Chris@402
|
22 #include <algorithm>
|
Chris@402
|
23
|
Chris@152
|
24 #include <cassert>
|
Chris@1090
|
25 #include <deque>
|
Chris@152
|
26
|
Chris@608
|
27 #ifndef __GNUC__
|
Chris@608
|
28 #include <alloca.h>
|
Chris@608
|
29 #endif
|
Chris@608
|
30
|
Chris@1090
|
31 using namespace std;
|
Chris@1090
|
32
|
Chris@152
|
33 FFTModel::FFTModel(const DenseTimeValueModel *model,
|
Chris@152
|
34 int channel,
|
Chris@152
|
35 WindowType windowType,
|
Chris@929
|
36 int windowSize,
|
Chris@929
|
37 int windowIncrement,
|
Chris@1090
|
38 int fftSize) :
|
Chris@1090
|
39 m_model(model),
|
Chris@1090
|
40 m_channel(channel),
|
Chris@1090
|
41 m_windowType(windowType),
|
Chris@1090
|
42 m_windowSize(windowSize),
|
Chris@1090
|
43 m_windowIncrement(windowIncrement),
|
Chris@1090
|
44 m_fftSize(fftSize),
|
Chris@1091
|
45 m_windower(windowType, windowSize),
|
Chris@1093
|
46 m_fft(fftSize),
|
Chris@1093
|
47 m_cacheSize(3)
|
Chris@152
|
48 {
|
Chris@1091
|
49 if (m_windowSize > m_fftSize) {
|
Chris@1091
|
50 cerr << "ERROR: FFTModel::FFTModel: window size (" << m_windowSize
|
Chris@1091
|
51 << ") must be at least FFT size (" << m_fftSize << ")" << endl;
|
Chris@1091
|
52 throw invalid_argument("FFTModel window size must be at least FFT size");
|
Chris@1091
|
53 }
|
Chris@1133
|
54
|
Chris@1133
|
55 connect(model, SIGNAL(modelChanged()), this, SIGNAL(modelChanged()));
|
Chris@1133
|
56 connect(model, SIGNAL(modelChangedWithin(sv_frame_t, sv_frame_t)),
|
Chris@1133
|
57 this, SIGNAL(modelChangedWithin(sv_frame_t, sv_frame_t)));
|
Chris@152
|
58 }
|
Chris@152
|
59
|
Chris@152
|
60 FFTModel::~FFTModel()
|
Chris@152
|
61 {
|
Chris@152
|
62 }
|
Chris@152
|
63
|
Chris@360
|
64 void
|
Chris@360
|
65 FFTModel::sourceModelAboutToBeDeleted()
|
Chris@360
|
66 {
|
Chris@1090
|
67 if (m_model) {
|
Chris@1090
|
68 cerr << "FFTModel[" << this << "]::sourceModelAboutToBeDeleted(" << m_model << ")" << endl;
|
Chris@1090
|
69 m_model = 0;
|
Chris@360
|
70 }
|
Chris@360
|
71 }
|
Chris@360
|
72
|
Chris@1091
|
73 int
|
Chris@1091
|
74 FFTModel::getWidth() const
|
Chris@1091
|
75 {
|
Chris@1091
|
76 if (!m_model) return 0;
|
Chris@1091
|
77 return int((m_model->getEndFrame() - m_model->getStartFrame())
|
Chris@1091
|
78 / m_windowIncrement) + 1;
|
Chris@1091
|
79 }
|
Chris@1091
|
80
|
Chris@1091
|
81 int
|
Chris@1091
|
82 FFTModel::getHeight() const
|
Chris@1091
|
83 {
|
Chris@1091
|
84 return m_fftSize / 2 + 1;
|
Chris@1091
|
85 }
|
Chris@1091
|
86
|
Chris@152
|
87 QString
|
Chris@929
|
88 FFTModel::getBinName(int n) const
|
Chris@152
|
89 {
|
Chris@1040
|
90 sv_samplerate_t sr = getSampleRate();
|
Chris@152
|
91 if (!sr) return "";
|
Chris@204
|
92 QString name = tr("%1 Hz").arg((n * sr) / ((getHeight()-1) * 2));
|
Chris@152
|
93 return name;
|
Chris@152
|
94 }
|
Chris@152
|
95
|
Chris@1091
|
96 FFTModel::Column
|
Chris@1091
|
97 FFTModel::getColumn(int x) const
|
Chris@1091
|
98 {
|
Chris@1091
|
99 auto cplx = getFFTColumn(x);
|
Chris@1091
|
100 Column col;
|
Chris@1154
|
101 col.reserve(cplx.size());
|
Chris@1091
|
102 for (auto c: cplx) col.push_back(abs(c));
|
Chris@1154
|
103 return move(col);
|
Chris@1091
|
104 }
|
Chris@1091
|
105
|
Chris@1200
|
106 FFTModel::Column
|
Chris@1200
|
107 FFTModel::getPhases(int x) const
|
Chris@1200
|
108 {
|
Chris@1200
|
109 auto cplx = getFFTColumn(x);
|
Chris@1200
|
110 Column col;
|
Chris@1200
|
111 col.reserve(cplx.size());
|
Chris@1200
|
112 for (auto c: cplx) col.push_back(arg(c));
|
Chris@1200
|
113 return move(col);
|
Chris@1200
|
114 }
|
Chris@1200
|
115
|
Chris@1091
|
116 float
|
Chris@1091
|
117 FFTModel::getMagnitudeAt(int x, int y) const
|
Chris@1091
|
118 {
|
Chris@1093
|
119 if (x < 0 || x >= getWidth() || y < 0 || y >= getHeight()) return 0.f;
|
Chris@1093
|
120 auto col = getFFTColumn(x);
|
Chris@1093
|
121 return abs(col[y]);
|
Chris@1091
|
122 }
|
Chris@1091
|
123
|
Chris@1091
|
124 float
|
Chris@1091
|
125 FFTModel::getMaximumMagnitudeAt(int x) const
|
Chris@1091
|
126 {
|
Chris@1091
|
127 Column col(getColumn(x));
|
Chris@1092
|
128 float max = 0.f;
|
Chris@1154
|
129 int n = int(col.size());
|
Chris@1154
|
130 for (int i = 0; i < n; ++i) {
|
Chris@1092
|
131 if (col[i] > max) max = col[i];
|
Chris@1092
|
132 }
|
Chris@1092
|
133 return max;
|
Chris@1091
|
134 }
|
Chris@1091
|
135
|
Chris@1091
|
136 float
|
Chris@1091
|
137 FFTModel::getPhaseAt(int x, int y) const
|
Chris@1091
|
138 {
|
Chris@1093
|
139 if (x < 0 || x >= getWidth() || y < 0 || y >= getHeight()) return 0.f;
|
Chris@1091
|
140 return arg(getFFTColumn(x)[y]);
|
Chris@1091
|
141 }
|
Chris@1091
|
142
|
Chris@1091
|
143 void
|
Chris@1091
|
144 FFTModel::getValuesAt(int x, int y, float &re, float &im) const
|
Chris@1091
|
145 {
|
Chris@1091
|
146 auto col = getFFTColumn(x);
|
Chris@1091
|
147 re = col[y].real();
|
Chris@1091
|
148 im = col[y].imag();
|
Chris@1091
|
149 }
|
Chris@1091
|
150
|
Chris@1091
|
151 bool
|
Chris@1091
|
152 FFTModel::getMagnitudesAt(int x, float *values, int minbin, int count) const
|
Chris@1091
|
153 {
|
Chris@1091
|
154 if (count == 0) count = getHeight();
|
Chris@1091
|
155 auto col = getFFTColumn(x);
|
Chris@1091
|
156 for (int i = 0; i < count; ++i) {
|
Chris@1091
|
157 values[i] = abs(col[minbin + i]);
|
Chris@1091
|
158 }
|
Chris@1091
|
159 return true;
|
Chris@1091
|
160 }
|
Chris@1091
|
161
|
Chris@1091
|
162 bool
|
Chris@1091
|
163 FFTModel::getPhasesAt(int x, float *values, int minbin, int count) const
|
Chris@1091
|
164 {
|
Chris@1091
|
165 if (count == 0) count = getHeight();
|
Chris@1091
|
166 auto col = getFFTColumn(x);
|
Chris@1091
|
167 for (int i = 0; i < count; ++i) {
|
Chris@1091
|
168 values[i] = arg(col[minbin + i]);
|
Chris@1091
|
169 }
|
Chris@1091
|
170 return true;
|
Chris@1091
|
171 }
|
Chris@1091
|
172
|
Chris@1091
|
173 bool
|
Chris@1091
|
174 FFTModel::getValuesAt(int x, float *reals, float *imags, int minbin, int count) const
|
Chris@1091
|
175 {
|
Chris@1091
|
176 if (count == 0) count = getHeight();
|
Chris@1091
|
177 auto col = getFFTColumn(x);
|
Chris@1091
|
178 for (int i = 0; i < count; ++i) {
|
Chris@1091
|
179 reals[i] = col[minbin + i].real();
|
Chris@1091
|
180 }
|
Chris@1091
|
181 for (int i = 0; i < count; ++i) {
|
Chris@1091
|
182 imags[i] = col[minbin + i].imag();
|
Chris@1091
|
183 }
|
Chris@1091
|
184 return true;
|
Chris@1091
|
185 }
|
Chris@1091
|
186
|
Chris@1091
|
187 vector<float>
|
Chris@1091
|
188 FFTModel::getSourceSamples(int column) const
|
Chris@1091
|
189 {
|
Chris@1094
|
190 // m_fftSize may be greater than m_windowSize, but not the reverse
|
Chris@1094
|
191
|
Chris@1094
|
192 // cerr << "getSourceSamples(" << column << ")" << endl;
|
Chris@1094
|
193
|
Chris@1091
|
194 auto range = getSourceSampleRange(column);
|
Chris@1094
|
195 auto data = getSourceData(range);
|
Chris@1094
|
196
|
Chris@1091
|
197 int off = (m_fftSize - m_windowSize) / 2;
|
Chris@1094
|
198
|
Chris@1094
|
199 if (off == 0) {
|
Chris@1094
|
200 return data;
|
Chris@1094
|
201 } else {
|
Chris@1094
|
202 vector<float> pad(off, 0.f);
|
Chris@1094
|
203 vector<float> padded;
|
Chris@1094
|
204 padded.reserve(m_fftSize);
|
Chris@1094
|
205 padded.insert(padded.end(), pad.begin(), pad.end());
|
Chris@1094
|
206 padded.insert(padded.end(), data.begin(), data.end());
|
Chris@1094
|
207 padded.insert(padded.end(), pad.begin(), pad.end());
|
Chris@1094
|
208 return padded;
|
Chris@1094
|
209 }
|
Chris@1094
|
210 }
|
Chris@1094
|
211
|
Chris@1094
|
212 vector<float>
|
Chris@1094
|
213 FFTModel::getSourceData(pair<sv_frame_t, sv_frame_t> range) const
|
Chris@1094
|
214 {
|
Chris@1094
|
215 // cerr << "getSourceData(" << range.first << "," << range.second
|
Chris@1094
|
216 // << "): saved range is (" << m_savedData.range.first
|
Chris@1094
|
217 // << "," << m_savedData.range.second << ")" << endl;
|
Chris@1094
|
218
|
Chris@1100
|
219 if (m_savedData.range == range) {
|
Chris@1100
|
220 return m_savedData.data;
|
Chris@1100
|
221 }
|
Chris@1094
|
222
|
Chris@1094
|
223 if (range.first < m_savedData.range.second &&
|
Chris@1094
|
224 range.first >= m_savedData.range.first &&
|
Chris@1094
|
225 range.second > m_savedData.range.second) {
|
Chris@1094
|
226
|
Chris@1100
|
227 sv_frame_t discard = range.first - m_savedData.range.first;
|
Chris@1100
|
228
|
Chris@1100
|
229 vector<float> acc(m_savedData.data.begin() + discard,
|
Chris@1100
|
230 m_savedData.data.end());
|
Chris@1094
|
231
|
Chris@1095
|
232 vector<float> rest =
|
Chris@1095
|
233 getSourceDataUncached({ m_savedData.range.second, range.second });
|
Chris@1100
|
234
|
Chris@1100
|
235 acc.insert(acc.end(), rest.begin(), rest.end());
|
Chris@1094
|
236
|
Chris@1095
|
237 m_savedData = { range, acc };
|
Chris@1095
|
238 return acc;
|
Chris@1095
|
239
|
Chris@1095
|
240 } else {
|
Chris@1095
|
241
|
Chris@1095
|
242 auto data = getSourceDataUncached(range);
|
Chris@1095
|
243 m_savedData = { range, data };
|
Chris@1095
|
244 return data;
|
Chris@1094
|
245 }
|
Chris@1095
|
246 }
|
Chris@1094
|
247
|
Chris@1095
|
248 vector<float>
|
Chris@1095
|
249 FFTModel::getSourceDataUncached(pair<sv_frame_t, sv_frame_t> range) const
|
Chris@1095
|
250 {
|
Chris@1091
|
251 decltype(range.first) pfx = 0;
|
Chris@1091
|
252 if (range.first < 0) {
|
Chris@1091
|
253 pfx = -range.first;
|
Chris@1091
|
254 range = { 0, range.second };
|
Chris@1091
|
255 }
|
Chris@1096
|
256
|
Chris@1096
|
257 auto data = m_model->getData(m_channel,
|
Chris@1096
|
258 range.first,
|
Chris@1096
|
259 range.second - range.first);
|
Chris@1096
|
260
|
Chris@1096
|
261 // don't return a partial frame
|
Chris@1096
|
262 data.resize(range.second - range.first, 0.f);
|
Chris@1096
|
263
|
Chris@1096
|
264 if (pfx > 0) {
|
Chris@1096
|
265 vector<float> pad(pfx, 0.f);
|
Chris@1096
|
266 data.insert(data.begin(), pad.begin(), pad.end());
|
Chris@1096
|
267 }
|
Chris@1096
|
268
|
Chris@1091
|
269 if (m_channel == -1) {
|
Chris@1091
|
270 int channels = m_model->getChannelCount();
|
Chris@1091
|
271 if (channels > 1) {
|
Chris@1096
|
272 int n = int(data.size());
|
Chris@1096
|
273 float factor = 1.f / float(channels);
|
Chris@1100
|
274 // use mean instead of sum for fft model input
|
Chris@1096
|
275 for (int i = 0; i < n; ++i) {
|
Chris@1096
|
276 data[i] *= factor;
|
Chris@1091
|
277 }
|
Chris@1091
|
278 }
|
Chris@1091
|
279 }
|
Chris@1094
|
280
|
Chris@1094
|
281 return data;
|
Chris@1091
|
282 }
|
Chris@1091
|
283
|
Chris@1091
|
284 vector<complex<float>>
|
Chris@1093
|
285 FFTModel::getFFTColumn(int n) const
|
Chris@1091
|
286 {
|
Chris@1093
|
287 for (auto &incache : m_cached) {
|
Chris@1093
|
288 if (incache.n == n) {
|
Chris@1093
|
289 return incache.col;
|
Chris@1093
|
290 }
|
Chris@1093
|
291 }
|
Chris@1093
|
292
|
Chris@1093
|
293 auto samples = getSourceSamples(n);
|
Chris@1100
|
294 m_windower.cut(samples.data());
|
Chris@1093
|
295 auto col = m_fft.process(samples);
|
Chris@1093
|
296
|
Chris@1093
|
297 SavedColumn sc { n, col };
|
Chris@1093
|
298 if (m_cached.size() >= m_cacheSize) {
|
Chris@1093
|
299 m_cached.pop_front();
|
Chris@1093
|
300 }
|
Chris@1093
|
301 m_cached.push_back(sc);
|
Chris@1093
|
302
|
Chris@1154
|
303 return move(col);
|
Chris@1091
|
304 }
|
Chris@1091
|
305
|
Chris@275
|
306 bool
|
Chris@1045
|
307 FFTModel::estimateStableFrequency(int x, int y, double &frequency)
|
Chris@275
|
308 {
|
Chris@275
|
309 if (!isOK()) return false;
|
Chris@275
|
310
|
Chris@1090
|
311 frequency = double(y * getSampleRate()) / m_fftSize;
|
Chris@275
|
312
|
Chris@275
|
313 if (x+1 >= getWidth()) return false;
|
Chris@275
|
314
|
Chris@275
|
315 // At frequency f, a phase shift of 2pi (one cycle) happens in 1/f sec.
|
Chris@275
|
316 // At hopsize h and sample rate sr, one hop happens in h/sr sec.
|
Chris@275
|
317 // At window size w, for bin b, f is b*sr/w.
|
Chris@275
|
318 // thus 2pi phase shift happens in w/(b*sr) sec.
|
Chris@275
|
319 // We need to know what phase shift we expect from h/sr sec.
|
Chris@275
|
320 // -> 2pi * ((h/sr) / (w/(b*sr)))
|
Chris@275
|
321 // = 2pi * ((h * b * sr) / (w * sr))
|
Chris@275
|
322 // = 2pi * (h * b) / w.
|
Chris@275
|
323
|
Chris@1038
|
324 double oldPhase = getPhaseAt(x, y);
|
Chris@1038
|
325 double newPhase = getPhaseAt(x+1, y);
|
Chris@275
|
326
|
Chris@929
|
327 int incr = getResolution();
|
Chris@275
|
328
|
Chris@1090
|
329 double expectedPhase = oldPhase + (2.0 * M_PI * y * incr) / m_fftSize;
|
Chris@275
|
330
|
Chris@1038
|
331 double phaseError = princarg(newPhase - expectedPhase);
|
Chris@275
|
332
|
Chris@275
|
333 // The new frequency estimate based on the phase error resulting
|
Chris@275
|
334 // from assuming the "native" frequency of this bin
|
Chris@275
|
335
|
Chris@275
|
336 frequency =
|
Chris@1090
|
337 (getSampleRate() * (expectedPhase + phaseError - oldPhase)) /
|
Chris@1045
|
338 (2.0 * M_PI * incr);
|
Chris@275
|
339
|
Chris@275
|
340 return true;
|
Chris@275
|
341 }
|
Chris@275
|
342
|
Chris@275
|
343 FFTModel::PeakLocationSet
|
Chris@1191
|
344 FFTModel::getPeaks(PeakPickType type, int x, int ymin, int ymax) const
|
Chris@275
|
345 {
|
Chris@551
|
346 Profiler profiler("FFTModel::getPeaks");
|
Chris@551
|
347
|
Chris@275
|
348 FFTModel::PeakLocationSet peaks;
|
Chris@275
|
349 if (!isOK()) return peaks;
|
Chris@275
|
350
|
Chris@275
|
351 if (ymax == 0 || ymax > getHeight() - 1) {
|
Chris@275
|
352 ymax = getHeight() - 1;
|
Chris@275
|
353 }
|
Chris@275
|
354
|
Chris@275
|
355 if (type == AllPeaks) {
|
Chris@551
|
356 int minbin = ymin;
|
Chris@551
|
357 if (minbin > 0) minbin = minbin - 1;
|
Chris@551
|
358 int maxbin = ymax;
|
Chris@551
|
359 if (maxbin < getHeight() - 1) maxbin = maxbin + 1;
|
Chris@551
|
360 const int n = maxbin - minbin + 1;
|
Chris@608
|
361 #ifdef __GNUC__
|
Chris@551
|
362 float values[n];
|
Chris@608
|
363 #else
|
Chris@608
|
364 float *values = (float *)alloca(n * sizeof(float));
|
Chris@608
|
365 #endif
|
Chris@551
|
366 getMagnitudesAt(x, values, minbin, maxbin - minbin + 1);
|
Chris@929
|
367 for (int bin = ymin; bin <= ymax; ++bin) {
|
Chris@551
|
368 if (bin == minbin || bin == maxbin) continue;
|
Chris@551
|
369 if (values[bin - minbin] > values[bin - minbin - 1] &&
|
Chris@551
|
370 values[bin - minbin] > values[bin - minbin + 1]) {
|
Chris@275
|
371 peaks.insert(bin);
|
Chris@275
|
372 }
|
Chris@275
|
373 }
|
Chris@275
|
374 return peaks;
|
Chris@275
|
375 }
|
Chris@275
|
376
|
Chris@551
|
377 Column values = getColumn(x);
|
Chris@1154
|
378 int nv = int(values.size());
|
Chris@275
|
379
|
Chris@500
|
380 float mean = 0.f;
|
Chris@1154
|
381 for (int i = 0; i < nv; ++i) mean += values[i];
|
Chris@1154
|
382 if (nv > 0) mean = mean / float(values.size());
|
Chris@1038
|
383
|
Chris@275
|
384 // For peak picking we use a moving median window, picking the
|
Chris@275
|
385 // highest value within each continuous region of values that
|
Chris@275
|
386 // exceed the median. For pitch adaptivity, we adjust the window
|
Chris@275
|
387 // size to a roughly constant pitch range (about four tones).
|
Chris@275
|
388
|
Chris@1040
|
389 sv_samplerate_t sampleRate = getSampleRate();
|
Chris@275
|
390
|
Chris@1090
|
391 deque<float> window;
|
Chris@1090
|
392 vector<int> inrange;
|
Chris@280
|
393 float dist = 0.5;
|
Chris@500
|
394
|
Chris@929
|
395 int medianWinSize = getPeakPickWindowSize(type, sampleRate, ymin, dist);
|
Chris@929
|
396 int halfWin = medianWinSize/2;
|
Chris@275
|
397
|
Chris@929
|
398 int binmin;
|
Chris@275
|
399 if (ymin > halfWin) binmin = ymin - halfWin;
|
Chris@275
|
400 else binmin = 0;
|
Chris@275
|
401
|
Chris@929
|
402 int binmax;
|
Chris@1154
|
403 if (ymax + halfWin < nv) binmax = ymax + halfWin;
|
Chris@1154
|
404 else binmax = nv - 1;
|
Chris@275
|
405
|
Chris@929
|
406 int prevcentre = 0;
|
Chris@500
|
407
|
Chris@929
|
408 for (int bin = binmin; bin <= binmax; ++bin) {
|
Chris@275
|
409
|
Chris@275
|
410 float value = values[bin];
|
Chris@275
|
411
|
Chris@275
|
412 window.push_back(value);
|
Chris@275
|
413
|
Chris@280
|
414 // so-called median will actually be the dist*100'th percentile
|
Chris@280
|
415 medianWinSize = getPeakPickWindowSize(type, sampleRate, bin, dist);
|
Chris@275
|
416 halfWin = medianWinSize/2;
|
Chris@275
|
417
|
Chris@929
|
418 while ((int)window.size() > medianWinSize) {
|
Chris@500
|
419 window.pop_front();
|
Chris@500
|
420 }
|
Chris@500
|
421
|
Chris@1038
|
422 int actualSize = int(window.size());
|
Chris@275
|
423
|
Chris@275
|
424 if (type == MajorPitchAdaptivePeaks) {
|
Chris@1154
|
425 if (ymax + halfWin < nv) binmax = ymax + halfWin;
|
Chris@1154
|
426 else binmax = nv - 1;
|
Chris@275
|
427 }
|
Chris@275
|
428
|
Chris@1090
|
429 deque<float> sorted(window);
|
Chris@1090
|
430 sort(sorted.begin(), sorted.end());
|
Chris@1038
|
431 float median = sorted[int(float(sorted.size()) * dist)];
|
Chris@275
|
432
|
Chris@929
|
433 int centrebin = 0;
|
Chris@500
|
434 if (bin > actualSize/2) centrebin = bin - actualSize/2;
|
Chris@500
|
435
|
Chris@500
|
436 while (centrebin > prevcentre || bin == binmin) {
|
Chris@275
|
437
|
Chris@500
|
438 if (centrebin > prevcentre) ++prevcentre;
|
Chris@500
|
439
|
Chris@500
|
440 float centre = values[prevcentre];
|
Chris@500
|
441
|
Chris@500
|
442 if (centre > median) {
|
Chris@500
|
443 inrange.push_back(centrebin);
|
Chris@500
|
444 }
|
Chris@500
|
445
|
Chris@1154
|
446 if (centre <= median || centrebin+1 == nv) {
|
Chris@500
|
447 if (!inrange.empty()) {
|
Chris@929
|
448 int peakbin = 0;
|
Chris@500
|
449 float peakval = 0.f;
|
Chris@929
|
450 for (int i = 0; i < (int)inrange.size(); ++i) {
|
Chris@500
|
451 if (i == 0 || values[inrange[i]] > peakval) {
|
Chris@500
|
452 peakval = values[inrange[i]];
|
Chris@500
|
453 peakbin = inrange[i];
|
Chris@500
|
454 }
|
Chris@500
|
455 }
|
Chris@500
|
456 inrange.clear();
|
Chris@500
|
457 if (peakbin >= ymin && peakbin <= ymax) {
|
Chris@500
|
458 peaks.insert(peakbin);
|
Chris@275
|
459 }
|
Chris@275
|
460 }
|
Chris@275
|
461 }
|
Chris@500
|
462
|
Chris@500
|
463 if (bin == binmin) break;
|
Chris@275
|
464 }
|
Chris@275
|
465 }
|
Chris@275
|
466
|
Chris@275
|
467 return peaks;
|
Chris@275
|
468 }
|
Chris@275
|
469
|
Chris@929
|
470 int
|
Chris@1040
|
471 FFTModel::getPeakPickWindowSize(PeakPickType type, sv_samplerate_t sampleRate,
|
Chris@929
|
472 int bin, float &percentile) const
|
Chris@275
|
473 {
|
Chris@280
|
474 percentile = 0.5;
|
Chris@275
|
475 if (type == MajorPeaks) return 10;
|
Chris@275
|
476 if (bin == 0) return 3;
|
Chris@280
|
477
|
Chris@1091
|
478 double binfreq = (sampleRate * bin) / m_fftSize;
|
Chris@1038
|
479 double hifreq = Pitch::getFrequencyForPitch(73, 0, binfreq);
|
Chris@280
|
480
|
Chris@1091
|
481 int hibin = int(lrint((hifreq * m_fftSize) / sampleRate));
|
Chris@275
|
482 int medianWinSize = hibin - bin;
|
Chris@275
|
483 if (medianWinSize < 3) medianWinSize = 3;
|
Chris@280
|
484
|
Chris@1091
|
485 percentile = 0.5f + float(binfreq / sampleRate);
|
Chris@280
|
486
|
Chris@275
|
487 return medianWinSize;
|
Chris@275
|
488 }
|
Chris@275
|
489
|
Chris@275
|
490 FFTModel::PeakSet
|
Chris@929
|
491 FFTModel::getPeakFrequencies(PeakPickType type, int x,
|
Chris@1191
|
492 int ymin, int ymax) const
|
Chris@275
|
493 {
|
Chris@551
|
494 Profiler profiler("FFTModel::getPeakFrequencies");
|
Chris@551
|
495
|
Chris@275
|
496 PeakSet peaks;
|
Chris@275
|
497 if (!isOK()) return peaks;
|
Chris@275
|
498 PeakLocationSet locations = getPeaks(type, x, ymin, ymax);
|
Chris@275
|
499
|
Chris@1040
|
500 sv_samplerate_t sampleRate = getSampleRate();
|
Chris@929
|
501 int incr = getResolution();
|
Chris@275
|
502
|
Chris@275
|
503 // This duplicates some of the work of estimateStableFrequency to
|
Chris@275
|
504 // allow us to retrieve the phases in two separate vertical
|
Chris@275
|
505 // columns, instead of jumping back and forth between columns x and
|
Chris@275
|
506 // x+1, which may be significantly slower if re-seeking is needed
|
Chris@275
|
507
|
Chris@1090
|
508 vector<float> phases;
|
Chris@275
|
509 for (PeakLocationSet::iterator i = locations.begin();
|
Chris@275
|
510 i != locations.end(); ++i) {
|
Chris@275
|
511 phases.push_back(getPhaseAt(x, *i));
|
Chris@275
|
512 }
|
Chris@275
|
513
|
Chris@929
|
514 int phaseIndex = 0;
|
Chris@275
|
515 for (PeakLocationSet::iterator i = locations.begin();
|
Chris@275
|
516 i != locations.end(); ++i) {
|
Chris@1038
|
517 double oldPhase = phases[phaseIndex];
|
Chris@1038
|
518 double newPhase = getPhaseAt(x+1, *i);
|
Chris@1090
|
519 double expectedPhase = oldPhase + (2.0 * M_PI * *i * incr) / m_fftSize;
|
Chris@1038
|
520 double phaseError = princarg(newPhase - expectedPhase);
|
Chris@1038
|
521 double frequency =
|
Chris@275
|
522 (sampleRate * (expectedPhase + phaseError - oldPhase))
|
Chris@275
|
523 / (2 * M_PI * incr);
|
Chris@1045
|
524 peaks[*i] = frequency;
|
Chris@275
|
525 ++phaseIndex;
|
Chris@275
|
526 }
|
Chris@275
|
527
|
Chris@275
|
528 return peaks;
|
Chris@275
|
529 }
|
Chris@275
|
530
|