2 * Dynamic Audio Normalizer
3 * Copyright (c) 2015 LoRd_MuldeR <mulder2@gmx.de>. Some rights reserved.
5 * This file is part of FFmpeg.
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * Dynamic Audio Normalizer
29 #include "libavutil/avassert.h"
30 #include "libavutil/opt.h"
32 #define FF_BUFQUEUE_SIZE 302
33 #include "libavfilter/bufferqueue.h"
39 typedef struct cqueue {
46 typedef struct DynamicAudioNormalizerContext {
49 struct FFBufQueue queue;
56 int alt_boundary_mode;
59 double max_amplification;
61 double compress_factor;
62 double *prev_amplification_factor;
63 double *dc_correction_value;
64 double *compress_threshold;
65 double *fade_factors[2];
71 cqueue **gain_history_original;
72 cqueue **gain_history_minimum;
73 cqueue **gain_history_smoothed;
74 } DynamicAudioNormalizerContext;
76 #define OFFSET(x) offsetof(DynamicAudioNormalizerContext, x)
77 #define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
79 static const AVOption dynaudnorm_options[] = {
80 { "f", "set the frame length in msec", OFFSET(frame_len_msec), AV_OPT_TYPE_INT, {.i64 = 500}, 10, 8000, FLAGS },
81 { "g", "set the filter size", OFFSET(filter_size), AV_OPT_TYPE_INT, {.i64 = 31}, 3, 301, FLAGS },
82 { "p", "set the peak value", OFFSET(peak_value), AV_OPT_TYPE_DOUBLE, {.dbl = 0.95}, 0.0, 1.0, FLAGS },
83 { "m", "set the max amplification", OFFSET(max_amplification), AV_OPT_TYPE_DOUBLE, {.dbl = 10.0}, 1.0, 100.0, FLAGS },
84 { "r", "set the target RMS", OFFSET(target_rms), AV_OPT_TYPE_DOUBLE, {.dbl = 0.0}, 0.0, 1.0, FLAGS },
85 { "n", "set channel coupling", OFFSET(channels_coupled), AV_OPT_TYPE_BOOL, {.i64 = 1}, 0, 1, FLAGS },
86 { "c", "set DC correction", OFFSET(dc_correction), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, FLAGS },
87 { "b", "set alternative boundary mode", OFFSET(alt_boundary_mode), AV_OPT_TYPE_BOOL, {.i64 = 0}, 0, 1, FLAGS },
88 { "s", "set the compress factor", OFFSET(compress_factor), AV_OPT_TYPE_DOUBLE, {.dbl = 0.0}, 0.0, 30.0, FLAGS },
92 AVFILTER_DEFINE_CLASS(dynaudnorm);
94 static av_cold int init(AVFilterContext *ctx)
96 DynamicAudioNormalizerContext *s = ctx->priv;
98 if (!(s->filter_size & 1)) {
99 av_log(ctx, AV_LOG_ERROR, "filter size %d is invalid. Must be an odd value.\n", s->filter_size);
100 return AVERROR(EINVAL);
106 static int query_formats(AVFilterContext *ctx)
108 AVFilterFormats *formats;
109 AVFilterChannelLayouts *layouts;
110 static const enum AVSampleFormat sample_fmts[] = {
116 layouts = ff_all_channel_counts();
118 return AVERROR(ENOMEM);
119 ret = ff_set_common_channel_layouts(ctx, layouts);
123 formats = ff_make_format_list(sample_fmts);
125 return AVERROR(ENOMEM);
126 ret = ff_set_common_formats(ctx, formats);
130 formats = ff_all_samplerates();
132 return AVERROR(ENOMEM);
133 return ff_set_common_samplerates(ctx, formats);
136 static inline int frame_size(int sample_rate, int frame_len_msec)
138 const int frame_size = round((double)sample_rate * (frame_len_msec / 1000.0));
139 return frame_size + (frame_size % 2);
142 static void precalculate_fade_factors(double *fade_factors[2], int frame_len)
144 const double step_size = 1.0 / frame_len;
147 for (pos = 0; pos < frame_len; pos++) {
148 fade_factors[0][pos] = 1.0 - (step_size * (pos + 1.0));
149 fade_factors[1][pos] = 1.0 - fade_factors[0][pos];
153 static cqueue *cqueue_create(int size)
157 q = av_malloc(sizeof(cqueue));
165 q->elements = av_malloc(sizeof(double) * size);
174 static void cqueue_free(cqueue *q)
176 av_free(q->elements);
180 static int cqueue_size(cqueue *q)
182 return q->nb_elements;
185 static int cqueue_empty(cqueue *q)
187 return !q->nb_elements;
190 static int cqueue_enqueue(cqueue *q, double element)
194 av_assert2(q->nb_elements != q->size);
196 i = (q->first + q->nb_elements) % q->size;
197 q->elements[i] = element;
203 static double cqueue_peek(cqueue *q, int index)
205 av_assert2(index < q->nb_elements);
206 return q->elements[(q->first + index) % q->size];
209 static int cqueue_dequeue(cqueue *q, double *element)
211 av_assert2(!cqueue_empty(q));
213 *element = q->elements[q->first];
214 q->first = (q->first + 1) % q->size;
220 static int cqueue_pop(cqueue *q)
222 av_assert2(!cqueue_empty(q));
224 q->first = (q->first + 1) % q->size;
230 static void init_gaussian_filter(DynamicAudioNormalizerContext *s)
232 double total_weight = 0.0;
233 const double sigma = (((s->filter_size / 2.0) - 1.0) / 3.0) + (1.0 / 3.0);
237 // Pre-compute constants
238 const int offset = s->filter_size / 2;
239 const double c1 = 1.0 / (sigma * sqrt(2.0 * M_PI));
240 const double c2 = 2.0 * sigma * sigma;
243 for (i = 0; i < s->filter_size; i++) {
244 const int x = i - offset;
246 s->weights[i] = c1 * exp(-x * x / c2);
247 total_weight += s->weights[i];
251 adjust = 1.0 / total_weight;
252 for (i = 0; i < s->filter_size; i++) {
253 s->weights[i] *= adjust;
257 static int config_input(AVFilterLink *inlink)
259 AVFilterContext *ctx = inlink->dst;
260 DynamicAudioNormalizerContext *s = ctx->priv;
264 inlink->min_samples =
265 inlink->max_samples =
266 inlink->partial_buf_size = frame_size(inlink->sample_rate, s->frame_len_msec);
267 av_log(ctx, AV_LOG_DEBUG, "frame len %d\n", s->frame_len);
269 s->fade_factors[0] = av_malloc(s->frame_len * sizeof(*s->fade_factors[0]));
270 s->fade_factors[1] = av_malloc(s->frame_len * sizeof(*s->fade_factors[1]));
272 s->prev_amplification_factor = av_malloc(inlink->channels * sizeof(*s->prev_amplification_factor));
273 s->dc_correction_value = av_calloc(inlink->channels, sizeof(*s->dc_correction_value));
274 s->compress_threshold = av_calloc(inlink->channels, sizeof(*s->compress_threshold));
275 s->gain_history_original = av_calloc(inlink->channels, sizeof(*s->gain_history_original));
276 s->gain_history_minimum = av_calloc(inlink->channels, sizeof(*s->gain_history_minimum));
277 s->gain_history_smoothed = av_calloc(inlink->channels, sizeof(*s->gain_history_smoothed));
278 s->weights = av_malloc(s->filter_size * sizeof(*s->weights));
279 if (!s->prev_amplification_factor || !s->dc_correction_value ||
280 !s->compress_threshold || !s->fade_factors[0] || !s->fade_factors[1] ||
281 !s->gain_history_original || !s->gain_history_minimum ||
282 !s->gain_history_smoothed || !s->weights)
283 return AVERROR(ENOMEM);
285 for (c = 0; c < inlink->channels; c++) {
286 s->prev_amplification_factor[c] = 1.0;
288 s->gain_history_original[c] = cqueue_create(s->filter_size);
289 s->gain_history_minimum[c] = cqueue_create(s->filter_size);
290 s->gain_history_smoothed[c] = cqueue_create(s->filter_size);
292 if (!s->gain_history_original[c] || !s->gain_history_minimum[c] ||
293 !s->gain_history_smoothed[c])
294 return AVERROR(ENOMEM);
297 precalculate_fade_factors(s->fade_factors, s->frame_len);
298 init_gaussian_filter(s);
300 s->channels = inlink->channels;
301 s->delay = s->filter_size;
306 static inline double fade(double prev, double next, int pos,
307 double *fade_factors[2])
309 return fade_factors[0][pos] * prev + fade_factors[1][pos] * next;
312 static inline double pow2(const double value)
314 return value * value;
317 static inline double bound(const double threshold, const double val)
319 const double CONST = 0.8862269254527580136490837416705725913987747280611935; //sqrt(PI) / 2.0
320 return erf(CONST * (val / threshold)) * threshold;
323 static double find_peak_magnitude(AVFrame *frame, int channel)
325 double max = DBL_EPSILON;
329 for (c = 0; c < av_frame_get_channels(frame); c++) {
330 double *data_ptr = (double *)frame->extended_data[c];
332 for (i = 0; i < frame->nb_samples; i++)
333 max = FFMAX(max, fabs(data_ptr[i]));
336 double *data_ptr = (double *)frame->extended_data[channel];
338 for (i = 0; i < frame->nb_samples; i++)
339 max = FFMAX(max, fabs(data_ptr[i]));
345 static double compute_frame_rms(AVFrame *frame, int channel)
347 double rms_value = 0.0;
351 for (c = 0; c < av_frame_get_channels(frame); c++) {
352 const double *data_ptr = (double *)frame->extended_data[c];
354 for (i = 0; i < frame->nb_samples; i++) {
355 rms_value += pow2(data_ptr[i]);
359 rms_value /= frame->nb_samples * av_frame_get_channels(frame);
361 const double *data_ptr = (double *)frame->extended_data[channel];
362 for (i = 0; i < frame->nb_samples; i++) {
363 rms_value += pow2(data_ptr[i]);
366 rms_value /= frame->nb_samples;
369 return FFMAX(sqrt(rms_value), DBL_EPSILON);
372 static double get_max_local_gain(DynamicAudioNormalizerContext *s, AVFrame *frame,
375 const double maximum_gain = s->peak_value / find_peak_magnitude(frame, channel);
376 const double rms_gain = s->target_rms > DBL_EPSILON ? (s->target_rms / compute_frame_rms(frame, channel)) : DBL_MAX;
377 return bound(s->max_amplification, FFMIN(maximum_gain, rms_gain));
380 static double minimum_filter(cqueue *q)
382 double min = DBL_MAX;
385 for (i = 0; i < cqueue_size(q); i++) {
386 min = FFMIN(min, cqueue_peek(q, i));
392 static double gaussian_filter(DynamicAudioNormalizerContext *s, cqueue *q)
397 for (i = 0; i < cqueue_size(q); i++) {
398 result += cqueue_peek(q, i) * s->weights[i];
404 static void update_gain_history(DynamicAudioNormalizerContext *s, int channel,
405 double current_gain_factor)
407 if (cqueue_empty(s->gain_history_original[channel]) ||
408 cqueue_empty(s->gain_history_minimum[channel])) {
409 const int pre_fill_size = s->filter_size / 2;
411 s->prev_amplification_factor[channel] = s->alt_boundary_mode ? current_gain_factor : 1.0;
413 while (cqueue_size(s->gain_history_original[channel]) < pre_fill_size) {
414 cqueue_enqueue(s->gain_history_original[channel], s->alt_boundary_mode ? current_gain_factor : 1.0);
417 while (cqueue_size(s->gain_history_minimum[channel]) < pre_fill_size) {
418 cqueue_enqueue(s->gain_history_minimum[channel], s->alt_boundary_mode ? current_gain_factor : 1.0);
422 cqueue_enqueue(s->gain_history_original[channel], current_gain_factor);
424 while (cqueue_size(s->gain_history_original[channel]) >= s->filter_size) {
426 av_assert0(cqueue_size(s->gain_history_original[channel]) == s->filter_size);
427 minimum = minimum_filter(s->gain_history_original[channel]);
429 cqueue_enqueue(s->gain_history_minimum[channel], minimum);
431 cqueue_pop(s->gain_history_original[channel]);
434 while (cqueue_size(s->gain_history_minimum[channel]) >= s->filter_size) {
436 av_assert0(cqueue_size(s->gain_history_minimum[channel]) == s->filter_size);
437 smoothed = gaussian_filter(s, s->gain_history_minimum[channel]);
439 cqueue_enqueue(s->gain_history_smoothed[channel], smoothed);
441 cqueue_pop(s->gain_history_minimum[channel]);
445 static inline double update_value(double new, double old, double aggressiveness)
447 av_assert0((aggressiveness >= 0.0) && (aggressiveness <= 1.0));
448 return aggressiveness * new + (1.0 - aggressiveness) * old;
451 static void perform_dc_correction(DynamicAudioNormalizerContext *s, AVFrame *frame)
453 const double diff = 1.0 / frame->nb_samples;
454 int is_first_frame = cqueue_empty(s->gain_history_original[0]);
457 for (c = 0; c < s->channels; c++) {
458 double *dst_ptr = (double *)frame->extended_data[c];
459 double current_average_value = 0.0;
462 for (i = 0; i < frame->nb_samples; i++)
463 current_average_value += dst_ptr[i] * diff;
465 prev_value = is_first_frame ? current_average_value : s->dc_correction_value[c];
466 s->dc_correction_value[c] = is_first_frame ? current_average_value : update_value(current_average_value, s->dc_correction_value[c], 0.1);
468 for (i = 0; i < frame->nb_samples; i++) {
469 dst_ptr[i] -= fade(prev_value, s->dc_correction_value[c], i, s->fade_factors);
474 static double setup_compress_thresh(double threshold)
476 if ((threshold > DBL_EPSILON) && (threshold < (1.0 - DBL_EPSILON))) {
477 double current_threshold = threshold;
478 double step_size = 1.0;
480 while (step_size > DBL_EPSILON) {
481 while ((current_threshold + step_size > current_threshold) &&
482 (bound(current_threshold + step_size, 1.0) <= threshold)) {
483 current_threshold += step_size;
489 return current_threshold;
495 static double compute_frame_std_dev(DynamicAudioNormalizerContext *s,
496 AVFrame *frame, int channel)
498 double variance = 0.0;
502 for (c = 0; c < s->channels; c++) {
503 const double *data_ptr = (double *)frame->extended_data[c];
505 for (i = 0; i < frame->nb_samples; i++) {
506 variance += pow2(data_ptr[i]); // Assume that MEAN is *zero*
509 variance /= (s->channels * frame->nb_samples) - 1;
511 const double *data_ptr = (double *)frame->extended_data[channel];
513 for (i = 0; i < frame->nb_samples; i++) {
514 variance += pow2(data_ptr[i]); // Assume that MEAN is *zero*
516 variance /= frame->nb_samples - 1;
519 return FFMAX(sqrt(variance), DBL_EPSILON);
522 static void perform_compression(DynamicAudioNormalizerContext *s, AVFrame *frame)
524 int is_first_frame = cqueue_empty(s->gain_history_original[0]);
527 if (s->channels_coupled) {
528 const double standard_deviation = compute_frame_std_dev(s, frame, -1);
529 const double current_threshold = FFMIN(1.0, s->compress_factor * standard_deviation);
531 const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[0];
532 double prev_actual_thresh, curr_actual_thresh;
533 s->compress_threshold[0] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[0], (1.0/3.0));
535 prev_actual_thresh = setup_compress_thresh(prev_value);
536 curr_actual_thresh = setup_compress_thresh(s->compress_threshold[0]);
538 for (c = 0; c < s->channels; c++) {
539 double *const dst_ptr = (double *)frame->extended_data[c];
540 for (i = 0; i < frame->nb_samples; i++) {
541 const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
542 dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
546 for (c = 0; c < s->channels; c++) {
547 const double standard_deviation = compute_frame_std_dev(s, frame, c);
548 const double current_threshold = setup_compress_thresh(FFMIN(1.0, s->compress_factor * standard_deviation));
550 const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[c];
551 double prev_actual_thresh, curr_actual_thresh;
553 s->compress_threshold[c] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[c], 1.0/3.0);
555 prev_actual_thresh = setup_compress_thresh(prev_value);
556 curr_actual_thresh = setup_compress_thresh(s->compress_threshold[c]);
558 dst_ptr = (double *)frame->extended_data[c];
559 for (i = 0; i < frame->nb_samples; i++) {
560 const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
561 dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
567 static void analyze_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
569 if (s->dc_correction) {
570 perform_dc_correction(s, frame);
573 if (s->compress_factor > DBL_EPSILON) {
574 perform_compression(s, frame);
577 if (s->channels_coupled) {
578 const double current_gain_factor = get_max_local_gain(s, frame, -1);
581 for (c = 0; c < s->channels; c++)
582 update_gain_history(s, c, current_gain_factor);
586 for (c = 0; c < s->channels; c++)
587 update_gain_history(s, c, get_max_local_gain(s, frame, c));
591 static void amplify_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
595 for (c = 0; c < s->channels; c++) {
596 double *dst_ptr = (double *)frame->extended_data[c];
597 double current_amplification_factor;
599 cqueue_dequeue(s->gain_history_smoothed[c], ¤t_amplification_factor);
601 for (i = 0; i < frame->nb_samples; i++) {
602 const double amplification_factor = fade(s->prev_amplification_factor[c],
603 current_amplification_factor, i,
606 dst_ptr[i] *= amplification_factor;
608 if (fabs(dst_ptr[i]) > s->peak_value)
609 dst_ptr[i] = copysign(s->peak_value, dst_ptr[i]);
612 s->prev_amplification_factor[c] = current_amplification_factor;
616 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
618 AVFilterContext *ctx = inlink->dst;
619 DynamicAudioNormalizerContext *s = ctx->priv;
620 AVFilterLink *outlink = inlink->dst->outputs[0];
623 if (!cqueue_empty(s->gain_history_smoothed[0])) {
624 AVFrame *out = ff_bufqueue_get(&s->queue);
626 amplify_frame(s, out);
627 ret = ff_filter_frame(outlink, out);
630 analyze_frame(s, in);
631 ff_bufqueue_add(ctx, &s->queue, in);
636 static int flush_buffer(DynamicAudioNormalizerContext *s, AVFilterLink *inlink,
637 AVFilterLink *outlink)
639 AVFrame *out = ff_get_audio_buffer(outlink, s->frame_len);
643 return AVERROR(ENOMEM);
645 for (c = 0; c < s->channels; c++) {
646 double *dst_ptr = (double *)out->extended_data[c];
648 for (i = 0; i < out->nb_samples; i++) {
649 dst_ptr[i] = s->alt_boundary_mode ? DBL_EPSILON : ((s->target_rms > DBL_EPSILON) ? FFMIN(s->peak_value, s->target_rms) : s->peak_value);
650 if (s->dc_correction) {
651 dst_ptr[i] *= ((i % 2) == 1) ? -1 : 1;
652 dst_ptr[i] += s->dc_correction_value[c];
658 return filter_frame(inlink, out);
661 static int request_frame(AVFilterLink *outlink)
663 AVFilterContext *ctx = outlink->src;
664 DynamicAudioNormalizerContext *s = ctx->priv;
667 ret = ff_request_frame(ctx->inputs[0]);
669 if (ret == AVERROR_EOF && !ctx->is_disabled && s->delay)
670 ret = flush_buffer(s, ctx->inputs[0], outlink);
675 static av_cold void uninit(AVFilterContext *ctx)
677 DynamicAudioNormalizerContext *s = ctx->priv;
680 av_freep(&s->prev_amplification_factor);
681 av_freep(&s->dc_correction_value);
682 av_freep(&s->compress_threshold);
683 av_freep(&s->fade_factors[0]);
684 av_freep(&s->fade_factors[1]);
686 for (c = 0; c < s->channels; c++) {
687 cqueue_free(s->gain_history_original[c]);
688 cqueue_free(s->gain_history_minimum[c]);
689 cqueue_free(s->gain_history_smoothed[c]);
692 av_freep(&s->gain_history_original);
693 av_freep(&s->gain_history_minimum);
694 av_freep(&s->gain_history_smoothed);
696 av_freep(&s->weights);
698 ff_bufqueue_discard_all(&s->queue);
701 static const AVFilterPad avfilter_af_dynaudnorm_inputs[] = {
704 .type = AVMEDIA_TYPE_AUDIO,
705 .filter_frame = filter_frame,
706 .config_props = config_input,
712 static const AVFilterPad avfilter_af_dynaudnorm_outputs[] = {
715 .type = AVMEDIA_TYPE_AUDIO,
716 .request_frame = request_frame,
721 AVFilter ff_af_dynaudnorm = {
722 .name = "dynaudnorm",
723 .description = NULL_IF_CONFIG_SMALL("Dynamic Audio Normalizer."),
724 .query_formats = query_formats,
725 .priv_size = sizeof(DynamicAudioNormalizerContext),
728 .inputs = avfilter_af_dynaudnorm_inputs,
729 .outputs = avfilter_af_dynaudnorm_outputs,
730 .priv_class = &dynaudnorm_class,