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", "enable channel coupling", OFFSET(channels_coupled), AV_OPT_TYPE_INT, {.i64 = 1}, 0, 1, FLAGS },
86 { "c", "enable DC correction", OFFSET(dc_correction), AV_OPT_TYPE_INT, {.i64 = 0}, 0, 1, FLAGS },
87 { "b", "enable alternative boundary mode", OFFSET(alt_boundary_mode), AV_OPT_TYPE_INT, {.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_layouts();
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 const double s_pi = 3.1415926535897932384626433832795028841971693993751058209749445923078164062862089986280348253421170679;
232 static void init_gaussian_filter(DynamicAudioNormalizerContext *s)
234 double total_weight = 0.0;
235 const double sigma = (((s->filter_size / 2.0) - 1.0) / 3.0) + (1.0 / 3.0);
239 // Pre-compute constants
240 const int offset = s->filter_size / 2;
241 const double c1 = 1.0 / (sigma * sqrt(2.0 * s_pi));
242 const double c2 = 2.0 * pow(sigma, 2.0);
245 for (i = 0; i < s->filter_size; i++) {
246 const int x = i - offset;
248 s->weights[i] = c1 * exp(-(pow(x, 2.0) / c2));
249 total_weight += s->weights[i];
253 adjust = 1.0 / total_weight;
254 for (i = 0; i < s->filter_size; i++) {
255 s->weights[i] *= adjust;
259 static int config_input(AVFilterLink *inlink)
261 AVFilterContext *ctx = inlink->dst;
262 DynamicAudioNormalizerContext *s = ctx->priv;
266 inlink->min_samples =
267 inlink->max_samples =
268 inlink->partial_buf_size = frame_size(inlink->sample_rate, s->frame_len_msec);
269 av_log(ctx, AV_LOG_DEBUG, "frame len %d\n", s->frame_len);
271 s->fade_factors[0] = av_malloc(s->frame_len * sizeof(*s->fade_factors[0]));
272 s->fade_factors[1] = av_malloc(s->frame_len * sizeof(*s->fade_factors[1]));
274 s->prev_amplification_factor = av_malloc(inlink->channels * sizeof(*s->prev_amplification_factor));
275 s->dc_correction_value = av_calloc(inlink->channels, sizeof(*s->dc_correction_value));
276 s->compress_threshold = av_calloc(inlink->channels, sizeof(*s->compress_threshold));
277 s->gain_history_original = av_calloc(inlink->channels, sizeof(*s->gain_history_original));
278 s->gain_history_minimum = av_calloc(inlink->channels, sizeof(*s->gain_history_minimum));
279 s->gain_history_smoothed = av_calloc(inlink->channels, sizeof(*s->gain_history_smoothed));
280 s->weights = av_malloc(s->filter_size * sizeof(*s->weights));
281 if (!s->prev_amplification_factor || !s->dc_correction_value ||
282 !s->compress_threshold || !s->fade_factors[0] || !s->fade_factors[1] ||
283 !s->gain_history_original || !s->gain_history_minimum ||
284 !s->gain_history_smoothed || !s->weights)
285 return AVERROR(ENOMEM);
287 for (c = 0; c < inlink->channels; c++) {
288 s->prev_amplification_factor[c] = 1.0;
290 s->gain_history_original[c] = cqueue_create(s->filter_size);
291 s->gain_history_minimum[c] = cqueue_create(s->filter_size);
292 s->gain_history_smoothed[c] = cqueue_create(s->filter_size);
294 if (!s->gain_history_original[c] || !s->gain_history_minimum[c] ||
295 !s->gain_history_smoothed[c])
296 return AVERROR(ENOMEM);
299 precalculate_fade_factors(s->fade_factors, s->frame_len);
300 init_gaussian_filter(s);
302 s->channels = inlink->channels;
303 s->delay = s->filter_size;
308 static int config_output(AVFilterLink *outlink)
310 outlink->flags |= FF_LINK_FLAG_REQUEST_LOOP;
314 static inline double fade(double prev, double next, int pos,
315 double *fade_factors[2])
317 return fade_factors[0][pos] * prev + fade_factors[1][pos] * next;
320 static inline double pow2(const double value)
322 return value * value;
325 static inline double bound(const double threshold, const double val)
327 const double CONST = 0.8862269254527580136490837416705725913987747280611935; //sqrt(PI) / 2.0
328 return erf(CONST * (val / threshold)) * threshold;
331 static double find_peak_magnitude(AVFrame *frame, int channel)
333 double max = DBL_EPSILON;
337 for (c = 0; c < av_frame_get_channels(frame); c++) {
338 double *data_ptr = (double *)frame->extended_data[c];
340 for (i = 0; i < frame->nb_samples; i++)
341 max = FFMAX(max, fabs(data_ptr[i]));
344 double *data_ptr = (double *)frame->extended_data[channel];
346 for (i = 0; i < frame->nb_samples; i++)
347 max = FFMAX(max, fabs(data_ptr[i]));
353 static double compute_frame_rms(AVFrame *frame, int channel)
355 double rms_value = 0.0;
359 for (c = 0; c < av_frame_get_channels(frame); c++) {
360 const double *data_ptr = (double *)frame->extended_data[c];
362 for (i = 0; i < frame->nb_samples; i++) {
363 rms_value += pow2(data_ptr[i]);
367 rms_value /= frame->nb_samples * av_frame_get_channels(frame);
369 const double *data_ptr = (double *)frame->extended_data[channel];
370 for (i = 0; i < frame->nb_samples; i++) {
371 rms_value += pow2(data_ptr[i]);
374 rms_value /= frame->nb_samples;
377 return FFMAX(sqrt(rms_value), DBL_EPSILON);
380 static double get_max_local_gain(DynamicAudioNormalizerContext *s, AVFrame *frame,
383 const double maximum_gain = s->peak_value / find_peak_magnitude(frame, channel);
384 const double rms_gain = s->target_rms > DBL_EPSILON ? (s->target_rms / compute_frame_rms(frame, channel)) : DBL_MAX;
385 return bound(s->max_amplification, FFMIN(maximum_gain, rms_gain));
388 static double minimum_filter(cqueue *q)
390 double min = DBL_MAX;
393 for (i = 0; i < cqueue_size(q); i++) {
394 min = FFMIN(min, cqueue_peek(q, i));
400 static double gaussian_filter(DynamicAudioNormalizerContext *s, cqueue *q)
405 for (i = 0; i < cqueue_size(q); i++) {
406 result += cqueue_peek(q, i) * s->weights[i];
412 static void update_gain_history(DynamicAudioNormalizerContext *s, int channel,
413 double current_gain_factor)
415 if (cqueue_empty(s->gain_history_original[channel]) ||
416 cqueue_empty(s->gain_history_minimum[channel])) {
417 const int pre_fill_size = s->filter_size / 2;
419 s->prev_amplification_factor[channel] = s->alt_boundary_mode ? current_gain_factor : 1.0;
421 while (cqueue_size(s->gain_history_original[channel]) < pre_fill_size) {
422 cqueue_enqueue(s->gain_history_original[channel], s->alt_boundary_mode ? current_gain_factor : 1.0);
425 while (cqueue_size(s->gain_history_minimum[channel]) < pre_fill_size) {
426 cqueue_enqueue(s->gain_history_minimum[channel], s->alt_boundary_mode ? current_gain_factor : 1.0);
430 cqueue_enqueue(s->gain_history_original[channel], current_gain_factor);
432 while (cqueue_size(s->gain_history_original[channel]) >= s->filter_size) {
434 av_assert0(cqueue_size(s->gain_history_original[channel]) == s->filter_size);
435 minimum = minimum_filter(s->gain_history_original[channel]);
437 cqueue_enqueue(s->gain_history_minimum[channel], minimum);
439 cqueue_pop(s->gain_history_original[channel]);
442 while (cqueue_size(s->gain_history_minimum[channel]) >= s->filter_size) {
444 av_assert0(cqueue_size(s->gain_history_minimum[channel]) == s->filter_size);
445 smoothed = gaussian_filter(s, s->gain_history_minimum[channel]);
447 cqueue_enqueue(s->gain_history_smoothed[channel], smoothed);
449 cqueue_pop(s->gain_history_minimum[channel]);
453 static inline double update_value(double new, double old, double aggressiveness)
455 av_assert0((aggressiveness >= 0.0) && (aggressiveness <= 1.0));
456 return aggressiveness * new + (1.0 - aggressiveness) * old;
459 static void perform_dc_correction(DynamicAudioNormalizerContext *s, AVFrame *frame)
461 const double diff = 1.0 / frame->nb_samples;
462 int is_first_frame = cqueue_empty(s->gain_history_original[0]);
465 for (c = 0; c < s->channels; c++) {
466 double *dst_ptr = (double *)frame->extended_data[c];
467 double current_average_value = 0.0;
470 for (i = 0; i < frame->nb_samples; i++)
471 current_average_value += dst_ptr[i] * diff;
473 prev_value = is_first_frame ? current_average_value : s->dc_correction_value[c];
474 s->dc_correction_value[c] = is_first_frame ? current_average_value : update_value(current_average_value, s->dc_correction_value[c], 0.1);
476 for (i = 0; i < frame->nb_samples; i++) {
477 dst_ptr[i] -= fade(prev_value, s->dc_correction_value[c], i, s->fade_factors);
482 static double setup_compress_thresh(double threshold)
484 if ((threshold > DBL_EPSILON) && (threshold < (1.0 - DBL_EPSILON))) {
485 double current_threshold = threshold;
486 double step_size = 1.0;
488 while (step_size > DBL_EPSILON) {
489 while ((current_threshold + step_size > current_threshold) &&
490 (bound(current_threshold + step_size, 1.0) <= threshold)) {
491 current_threshold += step_size;
497 return current_threshold;
503 static double compute_frame_std_dev(DynamicAudioNormalizerContext *s,
504 AVFrame *frame, int channel)
506 double variance = 0.0;
510 for (c = 0; c < s->channels; c++) {
511 const double *data_ptr = (double *)frame->extended_data[c];
513 for (i = 0; i < frame->nb_samples; i++) {
514 variance += pow2(data_ptr[i]); // Assume that MEAN is *zero*
517 variance /= (s->channels * frame->nb_samples) - 1;
519 const double *data_ptr = (double *)frame->extended_data[channel];
521 for (i = 0; i < frame->nb_samples; i++) {
522 variance += pow2(data_ptr[i]); // Assume that MEAN is *zero*
524 variance /= frame->nb_samples - 1;
527 return FFMAX(sqrt(variance), DBL_EPSILON);
530 static void perform_compression(DynamicAudioNormalizerContext *s, AVFrame *frame)
532 int is_first_frame = cqueue_empty(s->gain_history_original[0]);
535 if (s->channels_coupled) {
536 const double standard_deviation = compute_frame_std_dev(s, frame, -1);
537 const double current_threshold = FFMIN(1.0, s->compress_factor * standard_deviation);
539 const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[0];
540 double prev_actual_thresh, curr_actual_thresh;
541 s->compress_threshold[0] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[0], (1.0/3.0));
543 prev_actual_thresh = setup_compress_thresh(prev_value);
544 curr_actual_thresh = setup_compress_thresh(s->compress_threshold[0]);
546 for (c = 0; c < s->channels; c++) {
547 double *const dst_ptr = (double *)frame->extended_data[c];
548 for (i = 0; i < frame->nb_samples; i++) {
549 const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
550 dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
554 for (c = 0; c < s->channels; c++) {
555 const double standard_deviation = compute_frame_std_dev(s, frame, c);
556 const double current_threshold = setup_compress_thresh(FFMIN(1.0, s->compress_factor * standard_deviation));
558 const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[c];
559 double prev_actual_thresh, curr_actual_thresh;
561 s->compress_threshold[c] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[c], 1.0/3.0);
563 prev_actual_thresh = setup_compress_thresh(prev_value);
564 curr_actual_thresh = setup_compress_thresh(s->compress_threshold[c]);
566 dst_ptr = (double *)frame->extended_data[c];
567 for (i = 0; i < frame->nb_samples; i++) {
568 const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
569 dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
575 static void analyze_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
577 if (s->dc_correction) {
578 perform_dc_correction(s, frame);
581 if (s->compress_factor > DBL_EPSILON) {
582 perform_compression(s, frame);
585 if (s->channels_coupled) {
586 const double current_gain_factor = get_max_local_gain(s, frame, -1);
589 for (c = 0; c < s->channels; c++)
590 update_gain_history(s, c, current_gain_factor);
594 for (c = 0; c < s->channels; c++)
595 update_gain_history(s, c, get_max_local_gain(s, frame, c));
599 static void amplify_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
603 for (c = 0; c < s->channels; c++) {
604 double *dst_ptr = (double *)frame->extended_data[c];
605 double current_amplification_factor;
607 cqueue_dequeue(s->gain_history_smoothed[c], ¤t_amplification_factor);
609 for (i = 0; i < frame->nb_samples; i++) {
610 const double amplification_factor = fade(s->prev_amplification_factor[c],
611 current_amplification_factor, i,
614 dst_ptr[i] *= amplification_factor;
616 if (fabs(dst_ptr[i]) > s->peak_value)
617 dst_ptr[i] = copysign(s->peak_value, dst_ptr[i]);
620 s->prev_amplification_factor[c] = current_amplification_factor;
624 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
626 AVFilterContext *ctx = inlink->dst;
627 DynamicAudioNormalizerContext *s = ctx->priv;
628 AVFilterLink *outlink = inlink->dst->outputs[0];
631 if (!cqueue_empty(s->gain_history_smoothed[0])) {
632 AVFrame *out = ff_bufqueue_get(&s->queue);
634 amplify_frame(s, out);
635 ret = ff_filter_frame(outlink, out);
638 analyze_frame(s, in);
639 ff_bufqueue_add(ctx, &s->queue, in);
644 static int flush_buffer(DynamicAudioNormalizerContext *s, AVFilterLink *inlink,
645 AVFilterLink *outlink)
647 AVFrame *out = ff_get_audio_buffer(outlink, s->frame_len);
651 return AVERROR(ENOMEM);
653 for (c = 0; c < s->channels; c++) {
654 double *dst_ptr = (double *)out->extended_data[c];
656 for (i = 0; i < out->nb_samples; i++) {
657 dst_ptr[i] = s->alt_boundary_mode ? DBL_EPSILON : ((s->target_rms > DBL_EPSILON) ? FFMIN(s->peak_value, s->target_rms) : s->peak_value);
658 if (s->dc_correction) {
659 dst_ptr[i] *= ((i % 2) == 1) ? -1 : 1;
660 dst_ptr[i] += s->dc_correction_value[c];
666 return filter_frame(inlink, out);
669 static int request_frame(AVFilterLink *outlink)
671 AVFilterContext *ctx = outlink->src;
672 DynamicAudioNormalizerContext *s = ctx->priv;
675 ret = ff_request_frame(ctx->inputs[0]);
677 if (ret == AVERROR_EOF && !ctx->is_disabled && s->delay)
678 ret = flush_buffer(s, ctx->inputs[0], outlink);
683 static av_cold void uninit(AVFilterContext *ctx)
685 DynamicAudioNormalizerContext *s = ctx->priv;
688 av_freep(&s->prev_amplification_factor);
689 av_freep(&s->dc_correction_value);
690 av_freep(&s->compress_threshold);
691 av_freep(&s->fade_factors[0]);
692 av_freep(&s->fade_factors[1]);
694 for (c = 0; c < s->channels; c++) {
695 cqueue_free(s->gain_history_original[c]);
696 cqueue_free(s->gain_history_minimum[c]);
697 cqueue_free(s->gain_history_smoothed[c]);
700 av_freep(&s->gain_history_original);
701 av_freep(&s->gain_history_minimum);
702 av_freep(&s->gain_history_smoothed);
704 av_freep(&s->weights);
706 ff_bufqueue_discard_all(&s->queue);
709 static const AVFilterPad avfilter_af_dynaudnorm_inputs[] = {
712 .type = AVMEDIA_TYPE_AUDIO,
713 .filter_frame = filter_frame,
714 .config_props = config_input,
720 static const AVFilterPad avfilter_af_dynaudnorm_outputs[] = {
723 .type = AVMEDIA_TYPE_AUDIO,
724 .config_props = config_output,
725 .request_frame = request_frame,
730 AVFilter ff_af_dynaudnorm = {
731 .name = "dynaudnorm",
732 .description = NULL_IF_CONFIG_SMALL("Dynamic Audio Normalizer."),
733 .query_formats = query_formats,
734 .priv_size = sizeof(DynamicAudioNormalizerContext),
737 .inputs = avfilter_af_dynaudnorm_inputs,
738 .outputs = avfilter_af_dynaudnorm_outputs,
739 .priv_class = &dynaudnorm_class,