10 #include <linux/soundcard.h>
12 #define SAMPLE_RATE 22050
13 #define FFT_LENGTH 4096 /* in samples */
14 #define PAD_FACTOR 2 /* 1/pf of the FFT samples are real samples, the rest are padding */
15 #define OVERLAP 4 /* 1/ol samples will be replaced in the buffer every frame. Should be
16 * a multiple of 2 for the Hamming window (see
17 * http://www-ccrma.stanford.edu/~jos/parshl/Choice_Hop_Size.html).
20 #define EQUAL_TEMPERAMENT 0
21 #define WELL_TEMPERED_GUITAR 1
23 #define TUNING WELL_TEMPERED_GUITAR
26 void read_chunk(int fd, double *in, unsigned num_samples);
27 void apply_window(double *in, double *out, unsigned num_samples);
28 std::pair<double, double> find_peak(double *in, unsigned num_samples);
29 void find_peak_magnitudes(std::complex<double> *in, double *out, unsigned num_samples);
30 std::pair<double, double> adjust_for_overtones(std::pair<double, double> base, double *in, unsigned num_samples);
31 double bin_to_freq(double bin, unsigned num_samples);
32 double freq_to_bin(double freq, unsigned num_samples);
33 std::string freq_to_tonename(double freq);
34 std::pair<double, double> interpolate_peak(double ym1, double y0, double y1);
35 void print_spectrogram(double freq, double amp);
36 void write_sine(int dsp_fd, double freq, unsigned num_samples);
40 double *in, *in_windowed;
41 std::complex<double> *out;
46 in = reinterpret_cast<double *> (fftw_malloc(sizeof(double) * FFT_LENGTH / PAD_FACTOR));
47 in_windowed = reinterpret_cast<double *> (fftw_malloc(sizeof(double) * FFT_LENGTH));
48 out = reinterpret_cast<std::complex<double> *> (fftw_malloc(sizeof(std::complex<double>) * (FFT_LENGTH / 2 + 1)));
49 bins = reinterpret_cast<double *> (fftw_malloc(sizeof(double) * (FFT_LENGTH / 2 + 1)));
51 memset(in, 0, sizeof(double) * FFT_LENGTH / PAD_FACTOR);
54 p = fftw_plan_dft_r2c_1d(FFT_LENGTH, in_windowed, reinterpret_cast<fftw_complex *> (out), FFTW_ESTIMATE);
56 int fd = get_dsp_fd();
58 read_chunk(fd, in, FFT_LENGTH);
59 apply_window(in, in_windowed, FFT_LENGTH);
61 find_peak_magnitudes(out, bins, FFT_LENGTH);
62 std::pair<double, double> peak = find_peak(bins, FFT_LENGTH);
64 peak = adjust_for_overtones(peak, bins, FFT_LENGTH);
66 if (peak.first < 50.0 || peak.second - log10(FFT_LENGTH) < 0.0) {
67 #if TUNING == WELL_TEMPERED_GUITAR
70 printf("............\n");
73 print_spectrogram(peak.first, peak.second - log10(FFT_LENGTH));
80 int fd = open("/dev/dsp", O_RDWR);
86 ioctl(3, SNDCTL_DSP_RESET, 0);
88 int fmt = AFMT_S16_LE; // FIXME
89 ioctl(fd, SNDCTL_DSP_SETFMT, &fmt);
92 ioctl(fd, SOUND_PCM_WRITE_CHANNELS, &chan);
94 int rate = SAMPLE_RATE;
95 ioctl(fd, SOUND_PCM_WRITE_RATE, &rate);
97 int max_fragments = 2;
98 int frag_shift = ffs(FFT_LENGTH / OVERLAP) - 1;
99 int fragments = (max_fragments << 16) | frag_shift;
100 ioctl(fd, SNDCTL_DSP_SETFRAGMENT, &fragments);
102 ioctl(3, SNDCTL_DSP_SYNC, 0);
108 void read_chunk(int fd, double *in, unsigned num_samples)
111 unsigned to_read = num_samples / PAD_FACTOR / OVERLAP;
114 memmove(in, in + to_read, (num_samples / PAD_FACTOR - to_read) * sizeof(double));
116 ret = read(fd, buf, to_read * sizeof(short));
122 if (ret != int(to_read * sizeof(short))) {
128 for (unsigned i = 0; i < to_read; ++i)
129 in[i + (num_samples / PAD_FACTOR - to_read)] = double(buf[i]);
132 // make a pure 440hz sine for testing
133 void read_chunk(int fd, double *in, unsigned num_samples)
135 static double theta = 0.0;
136 for (unsigned i = 0; i < num_samples; ++i) {
138 theta += 2.0 * M_PI * 440.0 / double(SAMPLE_RATE);
143 void write_sine(int dsp_fd, double freq, unsigned num_samples)
145 static double theta = 0.0;
146 short buf[num_samples];
148 for (unsigned i = 0; i < num_samples; ++i) {
149 buf[i] = short(cos(theta) * 16384.0);
150 theta += 2.0 * M_PI * freq / double(SAMPLE_RATE);
153 write(dsp_fd, buf, num_samples * sizeof(short));
156 // Apply a standard Hamming window to our input data.
157 void apply_window(double *in, double *out, unsigned num_samples)
159 static double *win_data;
160 static unsigned win_len;
161 static bool win_inited = false;
163 // Initialize the window for the first time
165 win_len = num_samples / PAD_FACTOR;
166 win_data = new double[win_len];
168 for (unsigned i = 0; i < win_len; ++i) {
169 win_data[i] = 0.54 - 0.46 * cos(2.0 * M_PI * double(i) / double(win_len - 1));
175 assert(win_len == num_samples / PAD_FACTOR);
177 for (unsigned i = 0; i < win_len; ++i) {
178 out[i] = in[i] * win_data[i];
180 for (unsigned i = win_len; i < num_samples; ++i) {
185 void find_peak_magnitudes(std::complex<double> *in, double *out, unsigned num_samples)
187 for (unsigned i = 0; i < num_samples / 2 + 1; ++i)
191 std::pair<double, double> find_peak(double *in, unsigned num_samples)
193 double best_peak = in[0];
194 unsigned best_bin = 0;
196 for (unsigned i = 1; i < num_samples / 2 + 1; ++i) {
197 if (in[i] > best_peak) {
203 if (best_bin == 0 || best_bin == num_samples / 2) {
204 return std::make_pair(-1.0, 0.0);
208 printf("undertone strength: %+4.2f %+4.2f %+4.2f [%+4.2f] %+4.2f %+4.2f %+4.2f\n",
209 20.0 * log10(in[best_bin*4] / FFT_LENGTH),
210 20.0 * log10(in[best_bin*3] / FFT_LENGTH),
211 20.0 * log10(in[best_bin*2] / FFT_LENGTH),
212 20.0 * log10(in[best_bin] / FFT_LENGTH),
213 20.0 * log10(in[best_bin/2] / FFT_LENGTH),
214 20.0 * log10(in[best_bin/3] / FFT_LENGTH),
215 20.0 * log10(in[best_bin/4] / FFT_LENGTH));
218 // see if we might have hit an overtone (set a limit of 5dB)
219 for (unsigned i = 4; i >= 1; --i) {
220 if (best_bin != best_bin / i &&
221 20.0 * log10(in[best_bin] / in[best_bin / i]) < 5.0f) {
223 printf("Overtone of degree %u!\n", i);
227 // consider sliding one bin up or down
228 if (best_bin > 1 && in[best_bin - 1] > in[best_bin] && in[best_bin - 1] > in[best_bin - 2]) {
230 } else if (best_bin < num_samples / 2 - 1 && in[best_bin + 1] > in[best_bin] && in[best_bin + 1] > in[best_bin + 2]) {
238 if (best_bin == 0 || best_bin == num_samples / 2) {
239 return std::make_pair(-1.0, 0.0);
241 std::pair<double, double> peak =
242 interpolate_peak(in[best_bin - 1],
246 return std::make_pair(bin_to_freq(double(best_bin) + peak.first, num_samples), peak.second);
249 // it's perhaps not ideal to _first_ find the peak and _then_ the harmonics --
250 // ideally, one should find the set of all peaks and then determine the likely
251 // base from that... something for later. :-)
252 std::pair<double, double> adjust_for_overtones(std::pair<double, double> base, double *in, unsigned num_samples)
254 double mu = base.first, var = 1.0 / (base.second * base.second);
256 //printf("Base at %.2f (amp=%5.2fdB)\n", base.first, base.second);
258 for (unsigned i = 2; i < 10; ++i) {
259 unsigned middle = unsigned(floor(freq_to_bin(base.first, num_samples) * i + 0.5));
260 unsigned lower = middle - (i+1)/2, upper = middle + (i+1)/2;
262 if (upper >= num_samples)
263 upper = num_samples - 2;
265 // printf("Searching in [%u,%u] = %f..%f\n", lower, upper, bin_to_freq(lower, num_samples), bin_to_freq(upper, num_samples));
267 // search for a peak in this interval
268 double best_harmonics_freq = -1.0;
269 double best_harmonics_amp = -1.0;
270 for (unsigned j = lower; j <= upper; ++j) {
271 if (in[j] > in[j-1] && in[j] > in[j+1]) {
272 std::pair<double, double> peak =
273 interpolate_peak(in[j - 1],
277 if (peak.second > best_harmonics_amp) {
278 best_harmonics_freq = bin_to_freq(j + peak.first, num_samples);
279 best_harmonics_amp = peak.second;
284 if (best_harmonics_amp <= 0.0)
287 //printf("Found overtone %u at %.2f (amp=%5.2fdB)\n", i, best_harmonics_freq,
288 // best_harmonics_amp);
290 double this_mu = best_harmonics_freq / double(i);
291 double this_var = 1.0 / (best_harmonics_amp * best_harmonics_amp);
293 double k = var / (var + this_var);
294 mu = (1.0 - k) * mu + k * this_mu;
297 return std::make_pair(mu, base.second);
300 double bin_to_freq(double bin, unsigned num_samples)
302 return bin * SAMPLE_RATE / double(num_samples);
304 double freq_to_bin(double freq, unsigned num_samples)
306 return freq * double(num_samples) / double(SAMPLE_RATE);
310 * Given three bins, find the interpolated real peak based
311 * on their magnitudes. To do this, we execute the following
314 * Fit a polynomial of the form ax^2 + bx + c = 0 to the data
315 * we have. Maple readily yields our coefficients, assuming
316 * that we have the values at x=-1, x=0 and x=1:
318 * > f := x -> a*x^2 + b*x + c;
321 * f := x -> a x + b x + c
323 * > cf := solve({ f(-1) = ym1, f(0) = y0, f(1) = y1 }, { a, b, c });
326 * cf := {c = y0, b = ---- - ---, a = ---- + --- - y0}
329 * Now let's find the maximum point for the polynomial (it has to be
330 * a maximum, since y0 is the greatest value):
332 * > xmax := solve(subs(cf, diff(f(x), x)) = 0, x);
335 * xmax := -------------------
336 * 2 (y1 + ym1 - 2 y0)
338 * We could go further and insert {fmax,a,b,c} into the original
339 * polynomial, but it just gets hairy. We calculate a, b and c separately
342 * http://www-ccrma.stanford.edu/~jos/parshl/Peak_Detection_Steps_3.html
343 * claims that detection is almost twice as good when using dB scale instead
344 * of linear scale for the input values, so we use that. (As a tiny bonus,
345 * we get back dB scale from the function.)
347 std::pair<double, double> interpolate_peak(double ym1, double y0, double y1)
358 double a = 0.5 * y1 + 0.5 * ym1 - y0;
359 double b = 0.5 * y1 - 0.5 * ym1;
362 double xmax = (ym1 - y1) / (2.0 * (y1 + ym1 - 2.0 * y0));
363 double ymax = 20.0 * (a * xmax * xmax + b * xmax + c) - 90.0;
365 return std::make_pair(xmax, ymax);
368 std::string freq_to_tonename(double freq)
370 std::string notenames[] = { "C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B" };
371 double half_notes_away = 12.0 * log2(freq / 440.0) - 3.0;
372 int hnai = int(floor(half_notes_away + 0.5));
373 int octave = (hnai + 48) / 12;
376 sprintf(buf, "%s%d + %.2f [%d]", notenames[((hnai % 12) + 12) % 12].c_str(), octave, half_notes_away - hnai, hnai);
380 #if TUNING == EQUAL_TEMPERAMENT
381 void print_spectrogram(double freq, double amp)
383 std::string notenames[] = { "C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B" };
384 double half_notes_away = 12.0 * log2(freq / 440.0) - 3.0;
385 int hnai = int(floor(half_notes_away + 0.5));
386 int octave = (hnai + 48) / 12;
388 for (int i = 0; i < 12; ++i)
389 if (i == ((hnai % 12) + 12) % 12)
394 printf(" (%-2s%d %+.2f, %5.2fHz) [%5.2fdB] [", notenames[((hnai % 12) + 12) % 12].c_str(), octave, half_notes_away - hnai,
397 double off = half_notes_away - hnai;
398 for (int i = -10; i <= 10; ++i) {
399 if (off >= (i-0.5) * 0.05 && off < (i+0.5) * 0.05) {
417 static note notes[] = {
418 { "E-3", 110.0 * (3.0/4.0) },
420 { "D-4", 110.0 * (4.0/3.0) },
421 { "G-4", 110.0 * (4.0/3.0)*(4.0/3.0) },
422 { "B-4", 440.0 * (3.0/4.0)*(3.0/4.0) },
423 { "E-5", 440.0 * (3.0/4.0) }
426 void print_spectrogram(double freq, double amp)
428 double best_away = 999999999.9;
429 unsigned best_away_ind = 0;
431 for (unsigned i = 0; i < sizeof(notes)/sizeof(note); ++i) {
432 double half_notes_away = 12.0 * log2(freq / notes[i].freq);
433 if (fabs(half_notes_away) < fabs(best_away)) {
434 best_away = half_notes_away;
439 for (unsigned i = 0; i < sizeof(notes)/sizeof(note); ++i)
440 if (i == best_away_ind)
445 printf(" (%s %+.2f, %5.2fHz) [%5.2fdB] [", notes[best_away_ind].notename, best_away, freq, amp);
447 for (int i = -10; i <= 10; ++i) {
448 if (best_away >= (i-0.5) * 0.05 && best_away < (i+0.5) * 0.05) {