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).
21 void read_chunk(int fd, double *in, unsigned num_samples);
22 void apply_window(double *in, double *out, unsigned num_samples);
23 std::pair<double, double> find_peak(double *in, unsigned num_samples);
24 void find_peak_magnitudes(std::complex<double> *in, double *out, unsigned num_samples);
25 double bin_to_freq(double bin, unsigned num_samples);
26 std::string freq_to_tonename(double freq);
27 std::pair<double, double> interpolate_peak(double ym1, double y0, double y1);
28 void print_spectrogram(double freq, double amp);
29 void write_sine(int dsp_fd, double freq, unsigned num_samples);
33 double *in, *in_windowed;
34 std::complex<double> *out;
39 in = reinterpret_cast<double *> (fftw_malloc(sizeof(double) * FFT_LENGTH / PAD_FACTOR));
40 in_windowed = reinterpret_cast<double *> (fftw_malloc(sizeof(double) * FFT_LENGTH));
41 out = reinterpret_cast<std::complex<double> *> (fftw_malloc(sizeof(std::complex<double>) * (FFT_LENGTH / 2 + 1)));
42 bins = reinterpret_cast<double *> (fftw_malloc(sizeof(double) * FFT_LENGTH / 2 + 1));
44 memset(in, 0, sizeof(double) * FFT_LENGTH / PAD_FACTOR);
47 p = fftw_plan_dft_r2c_1d(FFT_LENGTH, in_windowed, reinterpret_cast<fftw_complex *> (out), FFTW_ESTIMATE);
49 int fd = get_dsp_fd();
51 read_chunk(fd, in, FFT_LENGTH);
52 apply_window(in, in_windowed, FFT_LENGTH);
54 find_peak_magnitudes(out, bins, FFT_LENGTH);
55 std::pair<double, double> peak = find_peak(bins, FFT_LENGTH);
57 if (peak.first < 50.0 || peak.second - log10(FFT_LENGTH) < 0.0) {
58 printf("............\n");
60 print_spectrogram(peak.first, peak.second - log10(FFT_LENGTH));
67 int fd = open("/dev/dsp", O_RDWR);
73 ioctl(3, SNDCTL_DSP_RESET, 0);
75 int fmt = AFMT_S16_LE; // FIXME
76 ioctl(fd, SNDCTL_DSP_SETFMT, &fmt);
79 ioctl(fd, SOUND_PCM_WRITE_CHANNELS, &chan);
82 ioctl(fd, SOUND_PCM_WRITE_RATE, &rate);
84 ioctl(3, SNDCTL_DSP_SYNC, 0);
90 void read_chunk(int fd, double *in, unsigned num_samples)
93 unsigned to_read = num_samples / PAD_FACTOR / OVERLAP;
96 memmove(in, in + to_read, (num_samples / PAD_FACTOR - to_read) * sizeof(double));
98 ret = read(fd, buf, to_read * sizeof(short));
104 if (ret != int(to_read * sizeof(short))) {
110 for (unsigned i = 0; i < to_read; ++i)
111 in[i + (num_samples / PAD_FACTOR - to_read)] = double(buf[i]);
114 // make a pure 440hz sine for testing
115 void read_chunk(int fd, double *in, unsigned num_samples)
117 static double theta = 0.0;
118 for (unsigned i = 0; i < num_samples; ++i) {
120 theta += 2.0 * M_PI * 440.0 / double(SAMPLE_RATE);
125 void write_sine(int dsp_fd, double freq, unsigned num_samples)
127 static double theta = 0.0;
128 short buf[num_samples];
130 for (unsigned i = 0; i < num_samples; ++i) {
131 buf[i] = short(cos(theta) * 16384.0);
132 theta += 2.0 * M_PI * freq / double(SAMPLE_RATE);
135 write(dsp_fd, buf, num_samples * sizeof(short));
138 // Apply a standard Hamming window to our input data.
139 void apply_window(double *in, double *out, unsigned num_samples)
141 static double *win_data;
142 static unsigned win_len;
143 static bool win_inited = false;
145 // Initialize the window for the first time
147 win_len = num_samples / PAD_FACTOR;
148 win_data = new double[win_len];
150 for (unsigned i = 0; i < win_len; ++i) {
151 win_data[i] = 0.54 - 0.46 * cos(2.0 * M_PI * double(i) / double(win_len - 1));
157 assert(win_len == num_samples / PAD_FACTOR);
159 for (unsigned i = 0; i < win_len; ++i) {
160 out[i] = in[i] * win_data[i];
162 for (unsigned i = win_len; i < num_samples; ++i) {
167 void find_peak_magnitudes(std::complex<double> *in, double *out, unsigned num_samples)
169 for (unsigned i = 0; i < num_samples / 2 + 1; ++i)
173 std::pair<double, double> find_peak(double *in, unsigned num_samples)
175 double best_peak = in[0];
176 unsigned best_bin = 0;
178 for (unsigned i = 1; i < num_samples / 2 + 1; ++i) {
179 if (in[i] > best_peak) {
185 if (best_bin == 0 || best_bin == num_samples / 2) {
186 return std::make_pair(-1.0, 0.0);
190 printf("undertone strength: %+4.2f %+4.2f %+4.2f [%+4.2f] %+4.2f %+4.2f %+4.2f\n",
191 20.0 * log10(in[best_bin*4] / FFT_LENGTH),
192 20.0 * log10(in[best_bin*3] / FFT_LENGTH),
193 20.0 * log10(in[best_bin*2] / FFT_LENGTH),
194 20.0 * log10(in[best_bin] / FFT_LENGTH),
195 20.0 * log10(in[best_bin/2] / FFT_LENGTH),
196 20.0 * log10(in[best_bin/3] / FFT_LENGTH),
197 20.0 * log10(in[best_bin/4] / FFT_LENGTH));
200 // see if we might have hit an overtone (set a limit of 5dB)
201 for (unsigned i = 4; i >= 1; --i) {
202 if (best_bin != best_bin / i &&
203 20.0 * log10(in[best_bin] / in[best_bin / i]) < 5.0f) {
205 printf("Overtone of degree %u!\n", i);
209 // consider sliding one bin up or down
210 if (best_bin > 0 && in[best_bin - 1] > in[best_bin] && in[best_bin - 1] > in[best_bin - 2]) {
212 } else if (best_bin < num_samples / 2 && in[best_bin + 1] > in[best_bin] && in[best_bin + 1] > in[best_bin + 2]) {
220 std::pair<double, double> peak =
221 interpolate_peak(in[best_bin - 1],
225 return std::make_pair(bin_to_freq(double(best_bin) + peak.first, num_samples), peak.second);
228 double bin_to_freq(double bin, unsigned num_samples)
230 return bin * SAMPLE_RATE / double(num_samples);
234 * Given three bins, find the interpolated real peak based
235 * on their magnitudes. To do this, we execute the following
238 * Fit a polynomial of the form ax^2 + bx + c = 0 to the data
239 * we have. Maple readily yields our coefficients, assuming
240 * that we have the values at x=-1, x=0 and x=1:
242 * > f := x -> a*x^2 + b*x + c;
245 * f := x -> a x + b x + c
247 * > cf := solve({ f(-1) = ym1, f(0) = y0, f(1) = y1 }, { a, b, c });
250 * cf := {c = y0, b = ---- - ---, a = ---- + --- - y0}
253 * Now let's find the maximum point for the polynomial (it has to be
254 * a maximum, since y0 is the greatest value):
256 * > xmax := solve(subs(cf, diff(f(x), x)) = 0, x);
259 * xmax := -------------------
260 * 2 (y1 + ym1 - 2 y0)
262 * We could go further and insert {fmax,a,b,c} into the original
263 * polynomial, but it just gets hairy. We calculate a, b and c separately
266 * http://www-ccrma.stanford.edu/~jos/parshl/Peak_Detection_Steps_3.html
267 * claims that detection is almost twice as good when using dB scale instead
268 * of linear scale for the input values, so we use that. (As a tiny bonus,
269 * we get back dB scale from the function.)
271 std::pair<double, double> interpolate_peak(double ym1, double y0, double y1)
282 double a = 0.5 * y1 + 0.5 * ym1 - y0;
283 double b = 0.5 * y1 - 0.5 * ym1;
286 double xmax = (ym1 - y1) / (2.0 * (y1 + ym1 - 2.0 * y0));
287 double ymax = 20.0 * (a * xmax * xmax + b * xmax + c) - 90.0;
289 return std::make_pair(xmax, ymax);
292 std::string freq_to_tonename(double freq)
294 std::string notenames[] = { "C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B" };
295 double half_notes_away = 12.0 * log2(freq / 440.0) - 3.0;
296 int hnai = int(floor(half_notes_away + 0.5));
297 int octave = (hnai + 48) / 12;
300 sprintf(buf, "%s%d + %.2f [%d]", notenames[((hnai % 12) + 12) % 12].c_str(), octave, half_notes_away - hnai, hnai);
304 void print_spectrogram(double freq, double amp)
306 std::string notenames[] = { "C", "C#", "D", "D#", "E", "F", "F#", "G", "G#", "A", "A#", "B" };
307 double half_notes_away = 12.0 * log2(freq / 440.0) - 3.0;
308 int hnai = int(floor(half_notes_away + 0.5));
309 int octave = (hnai + 48) / 12;
311 for (int i = 0; i < 12; ++i)
312 if (i == ((hnai % 12) + 12) % 12)
317 printf(" (%-2s%d %+.2f, %5.2fHz) [%5.2fdB] [", notenames[((hnai % 12) + 12) % 12].c_str(), octave, half_notes_away - hnai,
320 double off = half_notes_away - hnai;
321 for (int i = -10; i <= 10; ++i) {
322 if (off >= (i-0.5) * 0.05 && off < (i+0.5) * 0.05) {