2 * Copyright (c) 2018 The FFmpeg Project
4 * This file is part of FFmpeg.
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with FFmpeg; if not, write to the Free Software
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 #include "libavutil/audio_fifo.h"
24 #include "libavutil/avstring.h"
25 #include "libavutil/channel_layout.h"
26 #include "libavutil/opt.h"
27 #include "libavcodec/avfft.h"
32 #define C (M_LN10 * 0.1)
34 #define RRATIO (1.0 - RATIO)
51 typedef struct DeNoiseChannel {
53 double noise_band_auto_var[15];
54 double noise_band_sample[15];
60 double *prior_band_excit;
64 double *spread_function;
69 FFTContext *fft, *ifft;
71 double noise_band_norm[15];
72 double noise_band_avr[15];
73 double noise_band_avi[15];
74 double noise_band_var[15];
78 double sfm_results[3];
79 int sfm_fail_flags[512];
83 typedef struct AudioFFTDeNoiseContext {
86 float noise_reduction;
95 float last_residual_floor;
96 float last_noise_floor;
97 float last_noise_reduction;
98 float last_noise_balance;
104 int sample_noise_start;
105 int sample_noise_end;
122 DeNoiseChannel *dnch;
127 double window_weight;
132 int noise_band_edge[17];
133 int noise_band_count;
140 } AudioFFTDeNoiseContext;
142 #define OFFSET(x) offsetof(AudioFFTDeNoiseContext, x)
143 #define A AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
145 static const AVOption afftdn_options[] = {
146 { "nr", "set the noise reduction", OFFSET(noise_reduction), AV_OPT_TYPE_FLOAT, {.dbl = 12}, .01, 97, A },
147 { "nf", "set the noise floor", OFFSET(noise_floor), AV_OPT_TYPE_FLOAT, {.dbl =-50}, -80,-20, A },
148 { "nt", "set the noise type", OFFSET(noise_type), AV_OPT_TYPE_INT, {.i64 = WHITE_NOISE}, WHITE_NOISE, NB_NOISE-1, A, "type" },
149 { "w", "white noise", 0, AV_OPT_TYPE_CONST, {.i64 = WHITE_NOISE}, 0, 0, A, "type" },
150 { "v", "vinyl noise", 0, AV_OPT_TYPE_CONST, {.i64 = VINYL_NOISE}, 0, 0, A, "type" },
151 { "s", "shellac noise", 0, AV_OPT_TYPE_CONST, {.i64 = SHELLAC_NOISE}, 0, 0, A, "type" },
152 { "c", "custom noise", 0, AV_OPT_TYPE_CONST, {.i64 = CUSTOM_NOISE}, 0, 0, A, "type" },
153 { "bn", "set the custom bands noise", OFFSET(band_noise_str), AV_OPT_TYPE_STRING, {.str = 0}, 0, 0, A },
154 { "rf", "set the residual floor", OFFSET(residual_floor), AV_OPT_TYPE_FLOAT, {.dbl =-38}, -80,-20, A },
155 { "tn", "track noise", OFFSET(track_noise), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, A },
156 { "tr", "track residual", OFFSET(track_residual), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, A },
157 { "om", "set output mode", OFFSET(output_mode), AV_OPT_TYPE_INT, {.i64 = OUT_MODE}, 0, NB_MODES-1, A, "mode" },
158 { "i", "input", 0, AV_OPT_TYPE_CONST, {.i64 = IN_MODE}, 0, 0, A, "mode" },
159 { "o", "output", 0, AV_OPT_TYPE_CONST, {.i64 = OUT_MODE}, 0, 0, A, "mode" },
160 { "n", "noise", 0, AV_OPT_TYPE_CONST, {.i64 = NOISE_MODE}, 0, 0, A, "mode" },
164 AVFILTER_DEFINE_CLASS(afftdn);
166 static int get_band_noise(AudioFFTDeNoiseContext *s,
172 d1 = a / s->band_centre[band];
173 d1 = 10.0 * log(1.0 + d1 * d1) / M_LN10;
174 d2 = b / s->band_centre[band];
175 d2 = 10.0 * log(1.0 + d2 * d2) / M_LN10;
176 d3 = s->band_centre[band] / c;
177 d3 = 10.0 * log(1.0 + d3 * d3) / M_LN10;
179 return lrint(-d1 + d2 - d3);
182 static void factor(double *array, int size)
184 for (int i = 0; i < size - 1; i++) {
185 for (int j = i + 1; j < size; j++) {
186 double d = array[j + i * size] / array[i + i * size];
188 array[j + i * size] = d;
189 for (int k = i + 1; k < size; k++) {
190 array[j + k * size] -= d * array[i + k * size];
196 static void solve(double *matrix, double *vector, int size)
198 for (int i = 0; i < size - 1; i++) {
199 for (int j = i + 1; j < size; j++) {
200 double d = matrix[j + i * size];
201 vector[j] -= d * vector[i];
205 vector[size - 1] /= matrix[size * size - 1];
207 for (int i = size - 2; i >= 0; i--) {
208 double d = vector[i];
209 for (int j = i + 1; j < size; j++)
210 d -= matrix[i + j * size] * vector[j];
211 vector[i] = d / matrix[i + i * size];
215 static int process_get_band_noise(AudioFFTDeNoiseContext *s,
216 DeNoiseChannel *dnch,
219 double product, sum, f;
223 return dnch->band_noise[band];
225 for (int j = 0; j < 5; j++) {
227 for (int k = 0; k < 15; k++)
228 sum += s->matrix_b[i++] * dnch->band_noise[k];
229 s->vector_b[j] = sum;
232 solve(s->matrix_a, s->vector_b, 5);
233 f = (0.5 * s->sample_rate) / s->band_centre[14];
234 f = 15.0 + log(f / 1.5) / log(1.5);
237 for (int j = 0; j < 5; j++) {
238 sum += product * s->vector_b[j];
245 static void calculate_sfm(AudioFFTDeNoiseContext *s,
246 DeNoiseChannel *dnch,
249 double d1 = 0.0, d2 = 1.0;
252 for (int k = start; k < end; k++) {
253 if (dnch->noisy_data[k] > s->sample_floor) {
255 d1 += dnch->noisy_data[k];
256 d2 *= dnch->noisy_data[k];
260 } else if (d2 < 1.0E-100) {
268 dnch->sfm_results[0] = d1;
269 d2 = log(d2) + 230.2585 * i;
272 dnch->sfm_results[1] = d1;
273 dnch->sfm_results[2] = d1 - d2;
275 dnch->sfm_results[0] = s->auto_floor;
276 dnch->sfm_results[1] = dnch->sfm_threshold;
277 dnch->sfm_results[2] = dnch->sfm_threshold;
281 static double limit_gain(double a, double b)
284 return (b * a - 1.0) / (b + a - 2.0);
286 return (b * a - 2.0 * a + 1.0) / (b - a);
290 static void process_frame(AudioFFTDeNoiseContext *s, DeNoiseChannel *dnch,
291 FFTComplex *fft_data,
292 double *prior, double *prior_band_excit, int track_noise)
294 double d1, d2, d3, gain;
297 d1 = fft_data[0].re * fft_data[0].re;
298 dnch->noisy_data[0] = d1;
299 d2 = d1 / dnch->abs_var[0];
300 d3 = RATIO * prior[0] + RRATIO * fmax(d2 - 1.0, 0.0);
301 gain = d3 / (1.0 + d3);
302 gain *= (gain + M_PI_4 / fmax(d2, 1.0E-6));
303 prior[0] = (d2 * gain);
304 dnch->clean_data[0] = (d1 * gain);
306 dnch->gain[0] = gain;
308 for (int i = 1; i < s->fft_length2; i++) {
309 d1 = fft_data[i].re * fft_data[i].re + fft_data[i].im * fft_data[i].im;
310 if (d1 > s->sample_floor)
313 dnch->noisy_data[i] = d1;
314 d2 = d1 / dnch->abs_var[i];
315 d3 = RATIO * prior[i] + RRATIO * fmax(d2 - 1.0, 0.0);
316 gain = d3 / (1.0 + d3);
317 gain *= (gain + M_PI_4 / fmax(d2, 1.0E-6));
318 prior[i] = d2 * gain;
319 dnch->clean_data[i] = d1 * gain;
321 dnch->gain[i] = gain;
323 d1 = fft_data[0].im * fft_data[0].im;
324 if (d1 > s->sample_floor)
327 dnch->noisy_data[s->fft_length2] = d1;
328 d2 = d1 / dnch->abs_var[s->fft_length2];
329 d3 = RATIO * prior[s->fft_length2] + RRATIO * fmax(d2 - 1.0, 0.0);
330 gain = d3 / (1.0 + d3);
331 gain *= gain + M_PI_4 / fmax(d2, 1.0E-6);
332 prior[s->fft_length2] = d2 * gain;
333 dnch->clean_data[s->fft_length2] = d1 * gain;
335 dnch->gain[s->fft_length2] = gain;
336 if (n > s->fft_length2 - 2) {
338 i1 = s->noise_band_count;
341 for (int i = 0; i <= s->noise_band_count; i++) {
342 if (n > 1.1 * s->noise_band_edge[i]) {
348 if (track_noise && (i1 > s->noise_band_count / 2)) {
349 int j = FFMIN(n, s->noise_band_edge[i1]);
352 for (k = i1 - 1; k >= 0; k--) {
353 int i = s->noise_band_edge[k];
354 calculate_sfm(s, dnch, i, j);
355 dnch->noise_band_sample[k] = dnch->sfm_results[0];
356 if (dnch->sfm_results[2] + 0.013 * m * fmax(0.0, dnch->sfm_results[1] - 20.53) >= dnch->sfm_threshold) {
364 double sum = 0.0, min, max;
367 for (i = i1 - 1; i > k; i--) {
368 min = log(dnch->noise_band_sample[i] / dnch->noise_band_auto_var[i]);
374 min = 3.0E-4 * i * i;
376 min = 3.0E-4 * (8 * i - 16);
379 max = 2.0E-4 * i * i;
381 max = 2.0E-4 * (4 * i - 4);
384 if (s->track_residual) {
385 if (s->last_noise_floor > s->last_residual_floor + 9) {
388 } else if (s->last_noise_floor > s->last_residual_floor + 6) {
391 } else if (s->last_noise_floor > s->last_residual_floor + 4) {
394 } else if (s->last_noise_floor > s->last_residual_floor + 2) {
397 } else if (s->last_noise_floor > s->last_residual_floor) {
406 sum = av_clipd(sum, -min, max);
408 for (int i = 0; i < 15; i++)
409 dnch->noise_band_auto_var[i] *= sum;
410 } else if (dnch->sfm_results[2] >= dnch->sfm_threshold) {
411 dnch->sfm_fail_flags[s->block_count & 0x1FF] = 1;
412 dnch->sfm_fail_total += 1;
416 for (int i = 0; i < s->number_of_bands; i++) {
417 dnch->band_excit[i] = 0.0;
418 dnch->band_amt[i] = 0.0;
421 for (int i = 0; i < s->bin_count; i++) {
422 dnch->band_excit[s->bin2band[i]] += dnch->clean_data[i];
425 for (int i = 0; i < s->number_of_bands; i++) {
426 dnch->band_excit[i] = fmax(dnch->band_excit[i],
427 s->band_alpha[i] * dnch->band_excit[i] +
428 s->band_beta[i] * prior_band_excit[i]);
429 prior_band_excit[i] = dnch->band_excit[i];
432 for (int j = 0, i = 0; j < s->number_of_bands; j++) {
433 for (int k = 0; k < s->number_of_bands; k++) {
434 dnch->band_amt[j] += dnch->spread_function[i++] * dnch->band_excit[k];
438 for (int i = 0; i < s->bin_count; i++)
439 dnch->amt[i] = dnch->band_amt[s->bin2band[i]];
441 if (dnch->amt[0] > dnch->abs_var[0]) {
443 } else if (dnch->amt[0] > dnch->min_abs_var[0]) {
444 double limit = sqrt(dnch->abs_var[0] / dnch->amt[0]);
445 dnch->gain[0] = limit_gain(dnch->gain[0], limit);
447 dnch->gain[0] = limit_gain(dnch->gain[0], s->max_gain);
449 if (dnch->amt[s->fft_length2] > dnch->abs_var[s->fft_length2]) {
450 dnch->gain[s->fft_length2] = 1.0;
451 } else if (dnch->amt[s->fft_length2] > dnch->min_abs_var[s->fft_length2]) {
452 double limit = sqrt(dnch->abs_var[s->fft_length2] / dnch->amt[s->fft_length2]);
453 dnch->gain[s->fft_length2] = limit_gain(dnch->gain[s->fft_length2], limit);
455 dnch->gain[s->fft_length2] = limit_gain(dnch->gain[s->fft_length2], s->max_gain);
458 for (int i = 1; i < s->fft_length2; i++) {
459 if (dnch->amt[i] > dnch->abs_var[i]) {
461 } else if (dnch->amt[i] > dnch->min_abs_var[i]) {
462 double limit = sqrt(dnch->abs_var[i] / dnch->amt[i]);
463 dnch->gain[i] = limit_gain(dnch->gain[i], limit);
465 dnch->gain[i] = limit_gain(dnch->gain[i], s->max_gain);
469 gain = dnch->gain[0];
470 dnch->clean_data[0] = (gain * gain * dnch->noisy_data[0]);
471 fft_data[0].re *= gain;
472 gain = dnch->gain[s->fft_length2];
473 dnch->clean_data[s->fft_length2] = (gain * gain * dnch->noisy_data[s->fft_length2]);
474 fft_data[0].im *= gain;
475 for (int i = 1; i < s->fft_length2; i++) {
476 gain = dnch->gain[i];
477 dnch->clean_data[i] = (gain * gain * dnch->noisy_data[i]);
478 fft_data[i].re *= gain;
479 fft_data[i].im *= gain;
483 static double freq2bark(double x)
485 double d = x / 7500.0;
487 return 13.0 * atan(7.6E-4 * x) + 3.5 * atan(d * d);
490 static int get_band_centre(AudioFFTDeNoiseContext *s, int band)
493 return lrint(s->band_centre[0] / 1.5);
495 return s->band_centre[band];
498 static int get_band_edge(AudioFFTDeNoiseContext *s, int band)
503 i = lrint(s->band_centre[14] * 1.224745);
505 i = lrint(s->band_centre[band] / 1.224745);
508 return FFMIN(i, s->sample_rate / 2);
511 static void set_band_parameters(AudioFFTDeNoiseContext *s,
512 DeNoiseChannel *dnch)
514 double band_noise, d2, d3, d4, d5;
515 int i = 0, j = 0, k = 0;
518 band_noise = process_get_band_noise(s, dnch, 0);
519 for (int m = j; m <= s->fft_length2; m++) {
526 j = s->fft_length * get_band_centre(s, k) / s->sample_rate;
529 band_noise = process_get_band_noise(s, dnch, k);
534 dnch->rel_var[m] = exp((d5 * d3 + band_noise * d4) * C);
536 dnch->rel_var[s->fft_length2] = exp(band_noise * C);
538 for (i = 0; i < 15; i++)
539 dnch->noise_band_auto_var[i] = s->max_var * exp((process_get_band_noise(s, dnch, i) - 2.0) * C);
541 for (i = 0; i <= s->fft_length2; i++) {
542 dnch->abs_var[i] = fmax(s->max_var * dnch->rel_var[i], 1.0);
543 dnch->min_abs_var[i] = s->gain_scale * dnch->abs_var[i];
547 static void read_custom_noise(AudioFFTDeNoiseContext *s, int ch)
549 DeNoiseChannel *dnch = &s->dnch[ch];
550 char *p, *arg, *saveptr = NULL;
551 int i, ret, band_noise[15] = { 0 };
553 if (!s->band_noise_str)
556 p = av_strdup(s->band_noise_str);
560 for (i = 0; i < 15; i++) {
561 if (!(arg = av_strtok(p, "| ", &saveptr)))
566 ret = sscanf(arg, "%d", &band_noise[i]);
568 av_log(s, AV_LOG_ERROR, "Custom band noise must be integer.\n");
572 band_noise[i] = av_clip(band_noise[i], -24, 24);
576 memcpy(dnch->band_noise, band_noise, sizeof(band_noise));
579 static void set_parameters(AudioFFTDeNoiseContext *s)
581 if (s->last_noise_floor != s->noise_floor)
582 s->last_noise_floor = s->noise_floor;
584 if (s->track_residual)
585 s->last_noise_floor = fmaxf(s->last_noise_floor, s->residual_floor);
587 s->max_var = s->floor * exp((100.0 + s->last_noise_floor) * C);
589 if (s->track_residual) {
590 s->last_residual_floor = s->residual_floor;
591 s->last_noise_reduction = fmax(s->last_noise_floor - s->last_residual_floor, 0);
592 s->max_gain = exp(s->last_noise_reduction * (0.5 * C));
593 } else if (s->noise_reduction != s->last_noise_reduction) {
594 s->last_noise_reduction = s->noise_reduction;
595 s->last_residual_floor = av_clipf(s->last_noise_floor - s->last_noise_reduction, -80, -20);
596 s->max_gain = exp(s->last_noise_reduction * (0.5 * C));
599 s->gain_scale = 1.0 / (s->max_gain * s->max_gain);
601 for (int ch = 0; ch < s->channels; ch++) {
602 DeNoiseChannel *dnch = &s->dnch[ch];
604 set_band_parameters(s, dnch);
608 static int config_input(AVFilterLink *inlink)
610 AVFilterContext *ctx = inlink->dst;
611 AudioFFTDeNoiseContext *s = ctx->priv;
612 double wscale, sar, sum, sdiv;
615 s->dnch = av_calloc(inlink->channels, sizeof(*s->dnch));
617 return AVERROR(ENOMEM);
619 s->pts = AV_NOPTS_VALUE;
620 s->channels = inlink->channels;
621 s->sample_rate = inlink->sample_rate;
622 s->sample_advance = s->sample_rate / 80;
623 s->window_length = 3 * s->sample_advance;
624 s->fft_length2 = 1 << (32 - ff_clz(s->window_length));
625 s->fft_length = s->fft_length2 * 2;
626 s->buffer_length = s->fft_length * 2;
627 s->bin_count = s->fft_length2 + 1;
629 s->band_centre[0] = 80;
630 for (i = 1; i < 15; i++) {
631 s->band_centre[i] = lrint(1.5 * s->band_centre[i - 1] + 5.0);
632 if (s->band_centre[i] < 1000) {
633 s->band_centre[i] = 10 * (s->band_centre[i] / 10);
634 } else if (s->band_centre[i] < 5000) {
635 s->band_centre[i] = 50 * ((s->band_centre[i] + 20) / 50);
636 } else if (s->band_centre[i] < 15000) {
637 s->band_centre[i] = 100 * ((s->band_centre[i] + 45) / 100);
639 s->band_centre[i] = 1000 * ((s->band_centre[i] + 495) / 1000);
643 for (j = 0; j < 5; j++) {
644 for (k = 0; k < 5; k++) {
645 s->matrix_a[j + k * 5] = 0.0;
646 for (m = 0; m < 15; m++)
647 s->matrix_a[j + k * 5] += pow(m, j + k);
651 factor(s->matrix_a, 5);
654 for (j = 0; j < 5; j++)
655 for (k = 0; k < 15; k++)
656 s->matrix_b[i++] = pow(k, j);
659 for (j = 0; j < 15; j++)
660 for (k = 0; k < 5; k++)
661 s->matrix_c[i++] = pow(j, k);
663 s->window = av_calloc(s->window_length, sizeof(*s->window));
664 s->bin2band = av_calloc(s->bin_count, sizeof(*s->bin2band));
665 if (!s->window || !s->bin2band)
666 return AVERROR(ENOMEM);
668 sdiv = s->sample_rate / 17640.0;
669 for (i = 0; i <= s->fft_length2; i++)
670 s->bin2band[i] = lrint(sdiv * freq2bark((0.5 * i * s->sample_rate) / s->fft_length2));
672 s->number_of_bands = s->bin2band[s->fft_length2] + 1;
674 s->band_alpha = av_calloc(s->number_of_bands, sizeof(*s->band_alpha));
675 s->band_beta = av_calloc(s->number_of_bands, sizeof(*s->band_beta));
676 if (!s->band_alpha || !s->band_beta)
677 return AVERROR(ENOMEM);
679 for (int ch = 0; ch < inlink->channels; ch++) {
680 DeNoiseChannel *dnch = &s->dnch[ch];
682 switch (s->noise_type) {
684 for (i = 0; i < 15; i++)
685 dnch->band_noise[i] = 0;
688 for (i = 0; i < 15; i++)
689 dnch->band_noise[i] = get_band_noise(s, i, 50.0, 500.5, 2125.0) + FFMAX(i - 7, 0);
692 for (i = 0; i < 15; i++)
693 dnch->band_noise[i] = get_band_noise(s, i, 1.0, 500.0, 1.0E10) + FFMAX(i - 12, -5);
695 read_custom_noise(s, ch);
702 dnch->sfm_threshold = 0.8;
703 dnch->sfm_alpha = 0.05;
704 for (i = 0; i < 512; i++)
705 dnch->sfm_fail_flags[i] = 0;
707 dnch->sfm_fail_total = 0;
708 j = FFMAX((int)(10.0 * (1.3 - dnch->sfm_threshold)), 1);
710 for (i = 0; i < 512; i += j) {
711 dnch->sfm_fail_flags[i] = 1;
712 dnch->sfm_fail_total += 1;
715 dnch->amt = av_calloc(s->bin_count, sizeof(*dnch->amt));
716 dnch->band_amt = av_calloc(s->number_of_bands, sizeof(*dnch->band_amt));
717 dnch->band_excit = av_calloc(s->number_of_bands, sizeof(*dnch->band_excit));
718 dnch->gain = av_calloc(s->bin_count, sizeof(*dnch->gain));
719 dnch->prior = av_calloc(s->bin_count, sizeof(*dnch->prior));
720 dnch->prior_band_excit = av_calloc(s->number_of_bands, sizeof(*dnch->prior_band_excit));
721 dnch->clean_data = av_calloc(s->bin_count, sizeof(*dnch->clean_data));
722 dnch->noisy_data = av_calloc(s->bin_count, sizeof(*dnch->noisy_data));
723 dnch->out_samples = av_calloc(s->buffer_length, sizeof(*dnch->out_samples));
724 dnch->abs_var = av_calloc(s->bin_count, sizeof(*dnch->abs_var));
725 dnch->rel_var = av_calloc(s->bin_count, sizeof(*dnch->rel_var));
726 dnch->min_abs_var = av_calloc(s->bin_count, sizeof(*dnch->min_abs_var));
727 dnch->fft_data = av_calloc(s->fft_length2 + 1, sizeof(*dnch->fft_data));
728 dnch->fft = av_fft_init(av_log2(s->fft_length2), 0);
729 dnch->ifft = av_fft_init(av_log2(s->fft_length2), 1);
730 dnch->spread_function = av_calloc(s->number_of_bands * s->number_of_bands,
731 sizeof(*dnch->spread_function));
738 !dnch->prior_band_excit ||
741 !dnch->out_samples ||
745 !dnch->min_abs_var ||
746 !dnch->spread_function ||
749 return AVERROR(ENOMEM);
752 for (int ch = 0; ch < inlink->channels; ch++) {
753 DeNoiseChannel *dnch = &s->dnch[ch];
754 double *prior_band_excit = dnch->prior_band_excit;
755 double *prior = dnch->prior;
759 p1 = pow(0.1, 2.5 / sdiv);
760 p2 = pow(0.1, 1.0 / sdiv);
762 for (m = 0; m < s->number_of_bands; m++) {
763 for (n = 0; n < s->number_of_bands; n++) {
765 dnch->spread_function[j++] = pow(p2, m - n);
767 dnch->spread_function[j++] = pow(p1, n - m);
769 dnch->spread_function[j++] = 1.0;
774 for (m = 0; m < s->number_of_bands; m++) {
775 dnch->band_excit[m] = 0.0;
776 prior_band_excit[m] = 0.0;
779 for (m = 0; m <= s->fft_length2; m++)
780 dnch->band_excit[s->bin2band[m]] += 1.0;
783 for (m = 0; m < s->number_of_bands; m++) {
784 for (n = 0; n < s->number_of_bands; n++)
785 prior_band_excit[m] += dnch->spread_function[j++] * dnch->band_excit[n];
790 for (int i = 0; i < s->number_of_bands; i++) {
791 if (i < lrint(12.0 * sdiv)) {
792 dnch->band_excit[i] = pow(0.1, 1.45 + 0.1 * i / sdiv);
794 dnch->band_excit[i] = pow(0.1, 2.5 - 0.2 * (i / sdiv - 14.0));
796 dnch->band_excit[i] = av_clipd(dnch->band_excit[i], min, max);
799 for (int i = 0; i <= s->fft_length2; i++)
801 for (int i = 0; i < s->buffer_length; i++)
802 dnch->out_samples[i] = 0;
805 for (int i = 0; i < s->number_of_bands; i++)
806 for (int k = 0; k < s->number_of_bands; k++)
807 dnch->spread_function[j++] *= dnch->band_excit[i] / prior_band_excit[i];
811 sar = s->sample_advance / s->sample_rate;
812 for (int i = 0; i <= s->fft_length2; i++) {
813 if ((i == s->fft_length2) || (s->bin2band[i] > j)) {
814 double d6 = (i - 1) * s->sample_rate / s->fft_length;
815 double d7 = fmin(0.008 + 2.2 / d6, 0.03);
816 s->band_alpha[j] = exp(-sar / d7);
817 s->band_beta[j] = 1.0 - s->band_alpha[j];
822 wscale = sqrt(16.0 / (9.0 * s->fft_length));
824 for (int i = 0; i < s->window_length; i++) {
825 double d10 = sin(i * M_PI / s->window_length);
831 s->window_weight = 0.5 * sum;
832 s->floor = (1LL << 48) * exp(-23.025558369790467) * s->window_weight;
833 s->sample_floor = s->floor * exp(4.144600506562284);
834 s->auto_floor = s->floor * exp(6.907667510937141);
838 s->noise_band_edge[0] = FFMIN(s->fft_length2, s->fft_length * get_band_edge(s, 0) / s->sample_rate);
840 for (int j = 1; j < 16; j++) {
841 s->noise_band_edge[j] = FFMIN(s->fft_length2, s->fft_length * get_band_edge(s, j) / s->sample_rate);
842 if (s->noise_band_edge[j] > lrint(1.1 * s->noise_band_edge[j - 1]))
844 s->noise_band_edge[16] = i;
846 s->noise_band_count = s->noise_band_edge[16];
848 s->fifo = av_audio_fifo_alloc(inlink->format, inlink->channels, s->fft_length);
850 return AVERROR(ENOMEM);
855 static void preprocess(FFTComplex *in, int len)
857 double d1, d2, d3, d4, d5, d6, d7, d8, d9, d10;
860 d5 = 2.0 * M_PI / len;
868 for (i = 1; i < len / 4; i++) {
870 d2 = 0.5 * (in[i].re + in[k].re);
871 d1 = 0.5 * (in[i].im - in[k].im);
872 d4 = 0.5 * (in[i].im + in[k].im);
873 d3 = 0.5 * (in[k].re - in[i].re);
874 in[i].re = d2 + d9 * d4 + d6 * d3;
875 in[i].im = d1 + d9 * d3 - d6 * d4;
876 in[k].re = d2 - d9 * d4 - d6 * d3;
877 in[k].im = -d1 + d9 * d3 - d6 * d4;
879 d9 += d9 * d8 - d6 * d7;
880 d6 += d6 * d8 + d10 * d7;
884 in[0].re = d2 + in[0].im;
885 in[0].im = d2 - in[0].im;
888 static void postprocess(FFTComplex *in, int len)
890 double d1, d2, d3, d4, d5, d6, d7, d8, d9, d10;
893 d5 = 2.0 * M_PI / len;
900 for (i = 1; i < len / 4; i++) {
902 d2 = 0.5 * (in[i].re + in[k].re);
903 d1 = 0.5 * (in[i].im - in[k].im);
904 d4 = 0.5 * (in[i].re - in[k].re);
905 d3 = 0.5 * (in[i].im + in[k].im);
906 in[i].re = d2 - d9 * d3 - d6 * d4;
907 in[i].im = d1 + d9 * d4 - d6 * d3;
908 in[k].re = d2 + d9 * d3 + d6 * d4;
909 in[k].im = -d1 + d9 * d4 - d6 * d3;
911 d9 += d9 * d8 - d6 * d7;
912 d6 += d6 * d8 + d10 * d7;
915 in[0].re = 0.5 * (d2 + in[0].im);
916 in[0].im = 0.5 * (d2 - in[0].im);
919 static void init_sample_noise(DeNoiseChannel *dnch)
921 for (int i = 0; i < 15; i++) {
922 dnch->noise_band_norm[i] = 0.0;
923 dnch->noise_band_avr[i] = 0.0;
924 dnch->noise_band_avi[i] = 0.0;
925 dnch->noise_band_var[i] = 0.0;
929 static void sample_noise_block(AudioFFTDeNoiseContext *s,
930 DeNoiseChannel *dnch,
933 float *src = (float *)in->extended_data[ch];
934 double mag2, var = 0.0, avr = 0.0, avi = 0.0;
935 int edge, j, k, n, edgemax;
937 for (int i = 0; i < s->window_length; i++) {
938 dnch->fft_data[i].re = s->window[i] * src[i] * (1LL << 24);
939 dnch->fft_data[i].im = 0.0;
942 for (int i = s->window_length; i < s->fft_length2; i++) {
943 dnch->fft_data[i].re = 0.0;
944 dnch->fft_data[i].im = 0.0;
947 av_fft_permute(dnch->fft, dnch->fft_data);
948 av_fft_calc(dnch->fft, dnch->fft_data);
950 preprocess(dnch->fft_data, s->fft_length);
952 edge = s->noise_band_edge[0];
956 edgemax = fmin(s->fft_length2, s->noise_band_edge[15]);
957 dnch->fft_data[s->fft_length2].re = dnch->fft_data[0].im;
958 dnch->fft_data[0].im = 0.0;
959 dnch->fft_data[s->fft_length2].im = 0.0;
961 for (int i = j; i <= edgemax; i++) {
962 if ((i == j) && (i < edgemax)) {
964 dnch->noise_band_norm[k - 1] += j - edge;
965 dnch->noise_band_avr[k - 1] += avr;
966 dnch->noise_band_avi[k - 1] += avi;
967 dnch->noise_band_var[k - 1] += var;
971 j = s->noise_band_edge[k];
979 avr += dnch->fft_data[n].re;
980 avi += dnch->fft_data[n].im;
981 mag2 = dnch->fft_data[n].re * dnch->fft_data[n].re +
982 dnch->fft_data[n].im * dnch->fft_data[n].im;
984 mag2 = fmax(mag2, s->sample_floor);
986 dnch->noisy_data[i] = mag2;
991 dnch->noise_band_norm[k - 1] += j - edge;
992 dnch->noise_band_avr[k - 1] += avr;
993 dnch->noise_band_avi[k - 1] += avi;
994 dnch->noise_band_var[k - 1] += var;
997 static void finish_sample_noise(AudioFFTDeNoiseContext *s,
998 DeNoiseChannel *dnch,
999 double *sample_noise)
1001 for (int i = 0; i < s->noise_band_count; i++) {
1002 dnch->noise_band_avr[i] /= dnch->noise_band_norm[i];
1003 dnch->noise_band_avi[i] /= dnch->noise_band_norm[i];
1004 dnch->noise_band_var[i] /= dnch->noise_band_norm[i];
1005 dnch->noise_band_var[i] -= dnch->noise_band_avr[i] * dnch->noise_band_avr[i] +
1006 dnch->noise_band_avi[i] * dnch->noise_band_avi[i];
1007 dnch->noise_band_auto_var[i] = dnch->noise_band_var[i];
1008 sample_noise[i] = (1.0 / C) * log(dnch->noise_band_var[i] / s->floor) - 100.0;
1010 if (s->noise_band_count < 15) {
1011 for (int i = s->noise_band_count; i < 15; i++)
1012 sample_noise[i] = sample_noise[i - 1];
1016 static void set_noise_profile(AudioFFTDeNoiseContext *s,
1017 DeNoiseChannel *dnch,
1018 double *sample_noise,
1021 int new_band_noise[15];
1023 double sum = 0.0, d1;
1024 float new_noise_floor;
1027 for (int m = 0; m < 15; m++)
1028 temp[m] = sample_noise[m];
1032 for (int m = 0; m < 5; m++) {
1034 for (n = 0; n < 15; n++)
1035 sum += s->matrix_b[i++] * temp[n];
1036 s->vector_b[m] = sum;
1038 solve(s->matrix_a, s->vector_b, 5);
1040 for (int m = 0; m < 15; m++) {
1042 for (n = 0; n < 5; n++)
1043 sum += s->matrix_c[i++] * s->vector_b[n];
1049 for (int m = 0; m < 15; m++)
1052 d1 = (int)(sum / 15.0 - 0.5);
1054 i = lrint(temp[7] - d1);
1056 for (d1 -= dnch->band_noise[7] - i; d1 > -20.0; d1 -= 1.0)
1059 for (int m = 0; m < 15; m++)
1062 new_noise_floor = d1 + 2.5;
1065 av_log(s, AV_LOG_INFO, "bn=");
1066 for (int m = 0; m < 15; m++) {
1067 new_band_noise[m] = lrint(temp[m]);
1068 new_band_noise[m] = av_clip(new_band_noise[m], -24, 24);
1069 av_log(s, AV_LOG_INFO, "%d ", new_band_noise[m]);
1071 av_log(s, AV_LOG_INFO, "\n");
1072 memcpy(dnch->band_noise, new_band_noise, sizeof(new_band_noise));
1076 s->noise_floor = new_noise_floor;
1079 typedef struct ThreadData {
1083 static int filter_channel(AVFilterContext *ctx, void *arg, int jobnr, int nb_jobs)
1085 AudioFFTDeNoiseContext *s = ctx->priv;
1086 ThreadData *td = arg;
1087 AVFrame *in = td->in;
1088 const int start = (in->channels * jobnr) / nb_jobs;
1089 const int end = (in->channels * (jobnr+1)) / nb_jobs;
1091 for (int ch = start; ch < end; ch++) {
1092 DeNoiseChannel *dnch = &s->dnch[ch];
1093 const float *src = (const float *)in->extended_data[ch];
1094 double *dst = dnch->out_samples;
1096 if (s->track_noise) {
1097 int i = s->block_count & 0x1FF;
1099 if (dnch->sfm_fail_flags[i])
1100 dnch->sfm_fail_total--;
1101 dnch->sfm_fail_flags[i] = 0;
1102 dnch->sfm_threshold *= 1.0 - dnch->sfm_alpha;
1103 dnch->sfm_threshold += dnch->sfm_alpha * (0.5 + (1.0 / 640) * dnch->sfm_fail_total);
1106 for (int m = 0; m < s->window_length; m++) {
1107 dnch->fft_data[m].re = s->window[m] * src[m] * (1LL << 24);
1108 dnch->fft_data[m].im = 0;
1111 for (int m = s->window_length; m < s->fft_length2; m++) {
1112 dnch->fft_data[m].re = 0;
1113 dnch->fft_data[m].im = 0;
1116 av_fft_permute(dnch->fft, dnch->fft_data);
1117 av_fft_calc(dnch->fft, dnch->fft_data);
1119 preprocess(dnch->fft_data, s->fft_length);
1120 process_frame(s, dnch, dnch->fft_data,
1122 dnch->prior_band_excit,
1124 postprocess(dnch->fft_data, s->fft_length);
1126 av_fft_permute(dnch->ifft, dnch->fft_data);
1127 av_fft_calc(dnch->ifft, dnch->fft_data);
1129 for (int m = 0; m < s->window_length; m++)
1130 dst[m] += s->window[m] * dnch->fft_data[m].re / (1LL << 24);
1136 static void get_auto_noise_levels(AudioFFTDeNoiseContext *s,
1137 DeNoiseChannel *dnch,
1140 if (s->noise_band_count > 0) {
1141 for (int i = 0; i < s->noise_band_count; i++) {
1142 levels[i] = (1.0 / C) * log(dnch->noise_band_auto_var[i] / s->floor) - 100.0;
1144 if (s->noise_band_count < 15) {
1145 for (int i = s->noise_band_count; i < 15; i++)
1146 levels[i] = levels[i - 1];
1149 for (int i = 0; i < 15; i++) {
1155 static int filter_frame(AVFilterLink *inlink, AVFrame *frame)
1157 AVFilterContext *ctx = inlink->dst;
1158 AVFilterLink *outlink = ctx->outputs[0];
1159 AudioFFTDeNoiseContext *s = ctx->priv;
1160 AVFrame *out = NULL, *in = NULL;
1164 if (s->pts == AV_NOPTS_VALUE)
1165 s->pts = frame->pts;
1167 ret = av_audio_fifo_write(s->fifo, (void **)frame->extended_data, frame->nb_samples);
1168 av_frame_free(&frame);
1172 while (av_audio_fifo_size(s->fifo) >= s->window_length) {
1174 in = ff_get_audio_buffer(outlink, s->window_length);
1176 return AVERROR(ENOMEM);
1179 ret = av_audio_fifo_peek(s->fifo, (void **)in->extended_data, s->window_length);
1183 if (s->track_noise) {
1184 for (int ch = 0; ch < inlink->channels; ch++) {
1185 DeNoiseChannel *dnch = &s->dnch[ch];
1188 get_auto_noise_levels(s, dnch, levels);
1189 set_noise_profile(s, dnch, levels, 0);
1192 if (s->noise_floor != s->last_noise_floor)
1196 if (s->sample_noise_start) {
1197 for (int ch = 0; ch < inlink->channels; ch++) {
1198 DeNoiseChannel *dnch = &s->dnch[ch];
1200 init_sample_noise(dnch);
1202 s->sample_noise_start = 0;
1203 s->sample_noise = 1;
1206 if (s->sample_noise) {
1207 for (int ch = 0; ch < inlink->channels; ch++) {
1208 DeNoiseChannel *dnch = &s->dnch[ch];
1210 sample_noise_block(s, dnch, in, ch);
1214 if (s->sample_noise_end) {
1215 for (int ch = 0; ch < inlink->channels; ch++) {
1216 DeNoiseChannel *dnch = &s->dnch[ch];
1217 double sample_noise[15];
1219 finish_sample_noise(s, dnch, sample_noise);
1220 set_noise_profile(s, dnch, sample_noise, 1);
1221 set_band_parameters(s, dnch);
1223 s->sample_noise = 0;
1224 s->sample_noise_end = 0;
1229 ctx->internal->execute(ctx, filter_channel, &td, NULL,
1230 FFMIN(outlink->channels, ff_filter_get_nb_threads(ctx)));
1232 out = ff_get_audio_buffer(outlink, s->sample_advance);
1234 ret = AVERROR(ENOMEM);
1238 for (int ch = 0; ch < inlink->channels; ch++) {
1239 DeNoiseChannel *dnch = &s->dnch[ch];
1240 double *src = dnch->out_samples;
1241 float *orig = (float *)in->extended_data[ch];
1242 float *dst = (float *)out->extended_data[ch];
1244 switch (s->output_mode) {
1246 for (int m = 0; m < s->sample_advance; m++)
1250 for (int m = 0; m < s->sample_advance; m++)
1254 for (int m = 0; m < s->sample_advance; m++)
1255 dst[m] = orig[m] - src[m];
1260 memmove(src, src + s->sample_advance, (s->window_length - s->sample_advance) * sizeof(*src));
1261 memset(src + (s->window_length - s->sample_advance), 0, s->sample_advance * sizeof(*src));
1264 av_audio_fifo_drain(s->fifo, s->sample_advance);
1267 ret = ff_filter_frame(outlink, out);
1270 s->pts += s->sample_advance;
1277 static av_cold void uninit(AVFilterContext *ctx)
1279 AudioFFTDeNoiseContext *s = ctx->priv;
1281 av_freep(&s->window);
1282 av_freep(&s->bin2band);
1283 av_freep(&s->band_alpha);
1284 av_freep(&s->band_beta);
1287 for (int ch = 0; ch < s->channels; ch++) {
1288 DeNoiseChannel *dnch = &s->dnch[ch];
1289 av_freep(&dnch->amt);
1290 av_freep(&dnch->band_amt);
1291 av_freep(&dnch->band_excit);
1292 av_freep(&dnch->gain);
1293 av_freep(&dnch->prior);
1294 av_freep(&dnch->prior_band_excit);
1295 av_freep(&dnch->clean_data);
1296 av_freep(&dnch->noisy_data);
1297 av_freep(&dnch->out_samples);
1298 av_freep(&dnch->spread_function);
1299 av_freep(&dnch->abs_var);
1300 av_freep(&dnch->rel_var);
1301 av_freep(&dnch->min_abs_var);
1302 av_freep(&dnch->fft_data);
1303 av_fft_end(dnch->fft);
1305 av_fft_end(dnch->ifft);
1311 av_audio_fifo_free(s->fifo);
1314 static int query_formats(AVFilterContext *ctx)
1316 AVFilterFormats *formats = NULL;
1317 AVFilterChannelLayouts *layouts = NULL;
1318 static const enum AVSampleFormat sample_fmts[] = {
1324 formats = ff_make_format_list(sample_fmts);
1326 return AVERROR(ENOMEM);
1327 ret = ff_set_common_formats(ctx, formats);
1331 layouts = ff_all_channel_counts();
1333 return AVERROR(ENOMEM);
1335 ret = ff_set_common_channel_layouts(ctx, layouts);
1339 formats = ff_all_samplerates();
1340 return ff_set_common_samplerates(ctx, formats);
1343 static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
1344 char *res, int res_len, int flags)
1346 AudioFFTDeNoiseContext *s = ctx->priv;
1349 if (!strcmp(cmd, "sample_noise") ||
1350 !strcmp(cmd, "sn")) {
1351 if (!strcmp(args, "start")) {
1352 s->sample_noise_start = 1;
1353 s->sample_noise_end = 0;
1354 } else if (!strcmp(args, "end")) {
1355 s->sample_noise_start = 0;
1356 s->sample_noise_end = 1;
1358 } else if (!strcmp(cmd, "nr") ||
1359 !strcmp(cmd, "noise_reduction")) {
1362 if (sscanf(args, "%f", &nr) == 1) {
1363 s->noise_reduction = av_clipf(nr, 0.01, 97);
1366 } else if (!strcmp(cmd, "nf") ||
1367 !strcmp(cmd, "noise_floor")) {
1370 if (sscanf(args, "%f", &nf) == 1) {
1371 s->noise_floor = av_clipf(nf, -80, -20);
1374 } else if (!strcmp(cmd, "output_mode") ||
1375 !strcmp(cmd, "om")) {
1376 if (!strcmp(args, "i")) {
1377 s->output_mode = IN_MODE;
1378 } else if (!strcmp(args, "o")) {
1379 s->output_mode = OUT_MODE;
1380 } else if (!strcmp(args, "n")) {
1381 s->output_mode = NOISE_MODE;
1391 static const AVFilterPad inputs[] = {
1394 .type = AVMEDIA_TYPE_AUDIO,
1395 .filter_frame = filter_frame,
1396 .config_props = config_input,
1401 static const AVFilterPad outputs[] = {
1404 .type = AVMEDIA_TYPE_AUDIO,
1409 AVFilter ff_af_afftdn = {
1411 .description = NULL_IF_CONFIG_SMALL("Denoise audio samples using FFT."),
1412 .query_formats = query_formats,
1413 .priv_size = sizeof(AudioFFTDeNoiseContext),
1414 .priv_class = &afftdn_class,
1418 .process_command = process_command,
1419 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_GENERIC |
1420 AVFILTER_FLAG_SLICE_THREADS,