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 = lrint((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_array(size, sizeof(double));
174 static void cqueue_free(cqueue *q)
177 av_free(q->elements);
181 static int cqueue_size(cqueue *q)
183 return q->nb_elements;
186 static int cqueue_empty(cqueue *q)
188 return !q->nb_elements;
191 static int cqueue_enqueue(cqueue *q, double element)
195 av_assert2(q->nb_elements != q->size);
197 i = (q->first + q->nb_elements) % q->size;
198 q->elements[i] = element;
204 static double cqueue_peek(cqueue *q, int index)
206 av_assert2(index < q->nb_elements);
207 return q->elements[(q->first + index) % q->size];
210 static int cqueue_dequeue(cqueue *q, double *element)
212 av_assert2(!cqueue_empty(q));
214 *element = q->elements[q->first];
215 q->first = (q->first + 1) % q->size;
221 static int cqueue_pop(cqueue *q)
223 av_assert2(!cqueue_empty(q));
225 q->first = (q->first + 1) % q->size;
231 static void init_gaussian_filter(DynamicAudioNormalizerContext *s)
233 double total_weight = 0.0;
234 const double sigma = (((s->filter_size / 2.0) - 1.0) / 3.0) + (1.0 / 3.0);
238 // Pre-compute constants
239 const int offset = s->filter_size / 2;
240 const double c1 = 1.0 / (sigma * sqrt(2.0 * M_PI));
241 const double c2 = 2.0 * sigma * sigma;
244 for (i = 0; i < s->filter_size; i++) {
245 const int x = i - offset;
247 s->weights[i] = c1 * exp(-x * x / c2);
248 total_weight += s->weights[i];
252 adjust = 1.0 / total_weight;
253 for (i = 0; i < s->filter_size; i++) {
254 s->weights[i] *= adjust;
258 static av_cold void uninit(AVFilterContext *ctx)
260 DynamicAudioNormalizerContext *s = ctx->priv;
263 av_freep(&s->prev_amplification_factor);
264 av_freep(&s->dc_correction_value);
265 av_freep(&s->compress_threshold);
266 av_freep(&s->fade_factors[0]);
267 av_freep(&s->fade_factors[1]);
269 for (c = 0; c < s->channels; c++) {
270 if (s->gain_history_original)
271 cqueue_free(s->gain_history_original[c]);
272 if (s->gain_history_minimum)
273 cqueue_free(s->gain_history_minimum[c]);
274 if (s->gain_history_smoothed)
275 cqueue_free(s->gain_history_smoothed[c]);
278 av_freep(&s->gain_history_original);
279 av_freep(&s->gain_history_minimum);
280 av_freep(&s->gain_history_smoothed);
282 av_freep(&s->weights);
284 ff_bufqueue_discard_all(&s->queue);
287 static int config_input(AVFilterLink *inlink)
289 AVFilterContext *ctx = inlink->dst;
290 DynamicAudioNormalizerContext *s = ctx->priv;
296 inlink->min_samples =
297 inlink->max_samples =
298 inlink->partial_buf_size = frame_size(inlink->sample_rate, s->frame_len_msec);
299 av_log(ctx, AV_LOG_DEBUG, "frame len %d\n", s->frame_len);
301 s->fade_factors[0] = av_malloc_array(s->frame_len, sizeof(*s->fade_factors[0]));
302 s->fade_factors[1] = av_malloc_array(s->frame_len, sizeof(*s->fade_factors[1]));
304 s->prev_amplification_factor = av_malloc_array(inlink->channels, sizeof(*s->prev_amplification_factor));
305 s->dc_correction_value = av_calloc(inlink->channels, sizeof(*s->dc_correction_value));
306 s->compress_threshold = av_calloc(inlink->channels, sizeof(*s->compress_threshold));
307 s->gain_history_original = av_calloc(inlink->channels, sizeof(*s->gain_history_original));
308 s->gain_history_minimum = av_calloc(inlink->channels, sizeof(*s->gain_history_minimum));
309 s->gain_history_smoothed = av_calloc(inlink->channels, sizeof(*s->gain_history_smoothed));
310 s->weights = av_malloc_array(s->filter_size, sizeof(*s->weights));
311 if (!s->prev_amplification_factor || !s->dc_correction_value ||
312 !s->compress_threshold || !s->fade_factors[0] || !s->fade_factors[1] ||
313 !s->gain_history_original || !s->gain_history_minimum ||
314 !s->gain_history_smoothed || !s->weights)
315 return AVERROR(ENOMEM);
317 for (c = 0; c < inlink->channels; c++) {
318 s->prev_amplification_factor[c] = 1.0;
320 s->gain_history_original[c] = cqueue_create(s->filter_size);
321 s->gain_history_minimum[c] = cqueue_create(s->filter_size);
322 s->gain_history_smoothed[c] = cqueue_create(s->filter_size);
324 if (!s->gain_history_original[c] || !s->gain_history_minimum[c] ||
325 !s->gain_history_smoothed[c])
326 return AVERROR(ENOMEM);
329 precalculate_fade_factors(s->fade_factors, s->frame_len);
330 init_gaussian_filter(s);
332 s->channels = inlink->channels;
333 s->delay = s->filter_size;
338 static inline double fade(double prev, double next, int pos,
339 double *fade_factors[2])
341 return fade_factors[0][pos] * prev + fade_factors[1][pos] * next;
344 static inline double pow_2(const double value)
346 return value * value;
349 static inline double bound(const double threshold, const double val)
351 const double CONST = 0.8862269254527580136490837416705725913987747280611935; //sqrt(PI) / 2.0
352 return erf(CONST * (val / threshold)) * threshold;
355 static double find_peak_magnitude(AVFrame *frame, int channel)
357 double max = DBL_EPSILON;
361 for (c = 0; c < frame->channels; c++) {
362 double *data_ptr = (double *)frame->extended_data[c];
364 for (i = 0; i < frame->nb_samples; i++)
365 max = FFMAX(max, fabs(data_ptr[i]));
368 double *data_ptr = (double *)frame->extended_data[channel];
370 for (i = 0; i < frame->nb_samples; i++)
371 max = FFMAX(max, fabs(data_ptr[i]));
377 static double compute_frame_rms(AVFrame *frame, int channel)
379 double rms_value = 0.0;
383 for (c = 0; c < frame->channels; c++) {
384 const double *data_ptr = (double *)frame->extended_data[c];
386 for (i = 0; i < frame->nb_samples; i++) {
387 rms_value += pow_2(data_ptr[i]);
391 rms_value /= frame->nb_samples * frame->channels;
393 const double *data_ptr = (double *)frame->extended_data[channel];
394 for (i = 0; i < frame->nb_samples; i++) {
395 rms_value += pow_2(data_ptr[i]);
398 rms_value /= frame->nb_samples;
401 return FFMAX(sqrt(rms_value), DBL_EPSILON);
404 static double get_max_local_gain(DynamicAudioNormalizerContext *s, AVFrame *frame,
407 const double maximum_gain = s->peak_value / find_peak_magnitude(frame, channel);
408 const double rms_gain = s->target_rms > DBL_EPSILON ? (s->target_rms / compute_frame_rms(frame, channel)) : DBL_MAX;
409 return bound(s->max_amplification, FFMIN(maximum_gain, rms_gain));
412 static double minimum_filter(cqueue *q)
414 double min = DBL_MAX;
417 for (i = 0; i < cqueue_size(q); i++) {
418 min = FFMIN(min, cqueue_peek(q, i));
424 static double gaussian_filter(DynamicAudioNormalizerContext *s, cqueue *q)
429 for (i = 0; i < cqueue_size(q); i++) {
430 result += cqueue_peek(q, i) * s->weights[i];
436 static void update_gain_history(DynamicAudioNormalizerContext *s, int channel,
437 double current_gain_factor)
439 if (cqueue_empty(s->gain_history_original[channel]) ||
440 cqueue_empty(s->gain_history_minimum[channel])) {
441 const int pre_fill_size = s->filter_size / 2;
442 const double initial_value = s->alt_boundary_mode ? current_gain_factor : 1.0;
444 s->prev_amplification_factor[channel] = initial_value;
446 while (cqueue_size(s->gain_history_original[channel]) < pre_fill_size) {
447 cqueue_enqueue(s->gain_history_original[channel], initial_value);
451 cqueue_enqueue(s->gain_history_original[channel], current_gain_factor);
453 while (cqueue_size(s->gain_history_original[channel]) >= s->filter_size) {
455 av_assert0(cqueue_size(s->gain_history_original[channel]) == s->filter_size);
457 if (cqueue_empty(s->gain_history_minimum[channel])) {
458 const int pre_fill_size = s->filter_size / 2;
459 double initial_value = s->alt_boundary_mode ? cqueue_peek(s->gain_history_original[channel], 0) : 1.0;
460 int input = pre_fill_size;
462 while (cqueue_size(s->gain_history_minimum[channel]) < pre_fill_size) {
464 initial_value = FFMIN(initial_value, cqueue_peek(s->gain_history_original[channel], input));
465 cqueue_enqueue(s->gain_history_minimum[channel], initial_value);
469 minimum = minimum_filter(s->gain_history_original[channel]);
471 cqueue_enqueue(s->gain_history_minimum[channel], minimum);
473 cqueue_pop(s->gain_history_original[channel]);
476 while (cqueue_size(s->gain_history_minimum[channel]) >= s->filter_size) {
478 av_assert0(cqueue_size(s->gain_history_minimum[channel]) == s->filter_size);
479 smoothed = gaussian_filter(s, s->gain_history_minimum[channel]);
481 cqueue_enqueue(s->gain_history_smoothed[channel], smoothed);
483 cqueue_pop(s->gain_history_minimum[channel]);
487 static inline double update_value(double new, double old, double aggressiveness)
489 av_assert0((aggressiveness >= 0.0) && (aggressiveness <= 1.0));
490 return aggressiveness * new + (1.0 - aggressiveness) * old;
493 static void perform_dc_correction(DynamicAudioNormalizerContext *s, AVFrame *frame)
495 const double diff = 1.0 / frame->nb_samples;
496 int is_first_frame = cqueue_empty(s->gain_history_original[0]);
499 for (c = 0; c < s->channels; c++) {
500 double *dst_ptr = (double *)frame->extended_data[c];
501 double current_average_value = 0.0;
504 for (i = 0; i < frame->nb_samples; i++)
505 current_average_value += dst_ptr[i] * diff;
507 prev_value = is_first_frame ? current_average_value : s->dc_correction_value[c];
508 s->dc_correction_value[c] = is_first_frame ? current_average_value : update_value(current_average_value, s->dc_correction_value[c], 0.1);
510 for (i = 0; i < frame->nb_samples; i++) {
511 dst_ptr[i] -= fade(prev_value, s->dc_correction_value[c], i, s->fade_factors);
516 static double setup_compress_thresh(double threshold)
518 if ((threshold > DBL_EPSILON) && (threshold < (1.0 - DBL_EPSILON))) {
519 double current_threshold = threshold;
520 double step_size = 1.0;
522 while (step_size > DBL_EPSILON) {
523 while ((llrint((current_threshold + step_size) * (UINT64_C(1) << 63)) >
524 llrint(current_threshold * (UINT64_C(1) << 63))) &&
525 (bound(current_threshold + step_size, 1.0) <= threshold)) {
526 current_threshold += step_size;
532 return current_threshold;
538 static double compute_frame_std_dev(DynamicAudioNormalizerContext *s,
539 AVFrame *frame, int channel)
541 double variance = 0.0;
545 for (c = 0; c < s->channels; c++) {
546 const double *data_ptr = (double *)frame->extended_data[c];
548 for (i = 0; i < frame->nb_samples; i++) {
549 variance += pow_2(data_ptr[i]); // Assume that MEAN is *zero*
552 variance /= (s->channels * frame->nb_samples) - 1;
554 const double *data_ptr = (double *)frame->extended_data[channel];
556 for (i = 0; i < frame->nb_samples; i++) {
557 variance += pow_2(data_ptr[i]); // Assume that MEAN is *zero*
559 variance /= frame->nb_samples - 1;
562 return FFMAX(sqrt(variance), DBL_EPSILON);
565 static void perform_compression(DynamicAudioNormalizerContext *s, AVFrame *frame)
567 int is_first_frame = cqueue_empty(s->gain_history_original[0]);
570 if (s->channels_coupled) {
571 const double standard_deviation = compute_frame_std_dev(s, frame, -1);
572 const double current_threshold = FFMIN(1.0, s->compress_factor * standard_deviation);
574 const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[0];
575 double prev_actual_thresh, curr_actual_thresh;
576 s->compress_threshold[0] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[0], (1.0/3.0));
578 prev_actual_thresh = setup_compress_thresh(prev_value);
579 curr_actual_thresh = setup_compress_thresh(s->compress_threshold[0]);
581 for (c = 0; c < s->channels; c++) {
582 double *const dst_ptr = (double *)frame->extended_data[c];
583 for (i = 0; i < frame->nb_samples; i++) {
584 const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
585 dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
589 for (c = 0; c < s->channels; c++) {
590 const double standard_deviation = compute_frame_std_dev(s, frame, c);
591 const double current_threshold = setup_compress_thresh(FFMIN(1.0, s->compress_factor * standard_deviation));
593 const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[c];
594 double prev_actual_thresh, curr_actual_thresh;
596 s->compress_threshold[c] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[c], 1.0/3.0);
598 prev_actual_thresh = setup_compress_thresh(prev_value);
599 curr_actual_thresh = setup_compress_thresh(s->compress_threshold[c]);
601 dst_ptr = (double *)frame->extended_data[c];
602 for (i = 0; i < frame->nb_samples; i++) {
603 const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
604 dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
610 static void analyze_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
612 if (s->dc_correction) {
613 perform_dc_correction(s, frame);
616 if (s->compress_factor > DBL_EPSILON) {
617 perform_compression(s, frame);
620 if (s->channels_coupled) {
621 const double current_gain_factor = get_max_local_gain(s, frame, -1);
624 for (c = 0; c < s->channels; c++)
625 update_gain_history(s, c, current_gain_factor);
629 for (c = 0; c < s->channels; c++)
630 update_gain_history(s, c, get_max_local_gain(s, frame, c));
634 static void amplify_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
638 for (c = 0; c < s->channels; c++) {
639 double *dst_ptr = (double *)frame->extended_data[c];
640 double current_amplification_factor;
642 cqueue_dequeue(s->gain_history_smoothed[c], ¤t_amplification_factor);
644 for (i = 0; i < frame->nb_samples; i++) {
645 const double amplification_factor = fade(s->prev_amplification_factor[c],
646 current_amplification_factor, i,
649 dst_ptr[i] *= amplification_factor;
651 if (fabs(dst_ptr[i]) > s->peak_value)
652 dst_ptr[i] = copysign(s->peak_value, dst_ptr[i]);
655 s->prev_amplification_factor[c] = current_amplification_factor;
659 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
661 AVFilterContext *ctx = inlink->dst;
662 DynamicAudioNormalizerContext *s = ctx->priv;
663 AVFilterLink *outlink = inlink->dst->outputs[0];
666 if (!cqueue_empty(s->gain_history_smoothed[0])) {
667 AVFrame *out = ff_bufqueue_get(&s->queue);
669 amplify_frame(s, out);
670 ret = ff_filter_frame(outlink, out);
673 analyze_frame(s, in);
674 ff_bufqueue_add(ctx, &s->queue, in);
679 static int flush_buffer(DynamicAudioNormalizerContext *s, AVFilterLink *inlink,
680 AVFilterLink *outlink)
682 AVFrame *out = ff_get_audio_buffer(outlink, s->frame_len);
686 return AVERROR(ENOMEM);
688 for (c = 0; c < s->channels; c++) {
689 double *dst_ptr = (double *)out->extended_data[c];
691 for (i = 0; i < out->nb_samples; i++) {
692 dst_ptr[i] = s->alt_boundary_mode ? DBL_EPSILON : ((s->target_rms > DBL_EPSILON) ? FFMIN(s->peak_value, s->target_rms) : s->peak_value);
693 if (s->dc_correction) {
694 dst_ptr[i] *= ((i % 2) == 1) ? -1 : 1;
695 dst_ptr[i] += s->dc_correction_value[c];
701 return filter_frame(inlink, out);
704 static int request_frame(AVFilterLink *outlink)
706 AVFilterContext *ctx = outlink->src;
707 DynamicAudioNormalizerContext *s = ctx->priv;
710 ret = ff_request_frame(ctx->inputs[0]);
712 if (ret == AVERROR_EOF && !ctx->is_disabled && s->delay) {
713 if (!cqueue_empty(s->gain_history_smoothed[0])) {
714 ret = flush_buffer(s, ctx->inputs[0], outlink);
715 } else if (s->queue.available) {
716 AVFrame *out = ff_bufqueue_get(&s->queue);
718 ret = ff_filter_frame(outlink, out);
725 static const AVFilterPad avfilter_af_dynaudnorm_inputs[] = {
728 .type = AVMEDIA_TYPE_AUDIO,
729 .filter_frame = filter_frame,
730 .config_props = config_input,
736 static const AVFilterPad avfilter_af_dynaudnorm_outputs[] = {
739 .type = AVMEDIA_TYPE_AUDIO,
740 .request_frame = request_frame,
745 AVFilter ff_af_dynaudnorm = {
746 .name = "dynaudnorm",
747 .description = NULL_IF_CONFIG_SMALL("Dynamic Audio Normalizer."),
748 .query_formats = query_formats,
749 .priv_size = sizeof(DynamicAudioNormalizerContext),
752 .inputs = avfilter_af_dynaudnorm_inputs,
753 .outputs = avfilter_af_dynaudnorm_outputs,
754 .priv_class = &dynaudnorm_class,