[FFmpeg-cvslog] avfilter: add Dynamic Audio Normalizer filter

LoRd_MuldeR git at videolan.org
Fri Jul 17 13:31:24 CEST 2015


ffmpeg | branch: master | LoRd_MuldeR <mulder2 at gmx.de> | Tue Jul  7 16:19:59 2015 +0000| [21436b95dc96e9cb2ae3f583f219349976ec1b7e] | committer: Paul B Mahol

avfilter: add Dynamic Audio Normalizer filter

> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=21436b95dc96e9cb2ae3f583f219349976ec1b7e
---

 doc/filters.texi            |  158 ++++++++++
 libavfilter/Makefile        |    1 +
 libavfilter/af_dynaudnorm.c |  734 +++++++++++++++++++++++++++++++++++++++++++
 libavfilter/allfilters.c    |    1 +
 4 files changed, 894 insertions(+)

diff --git a/doc/filters.texi b/doc/filters.texi
index 49fab59..518aef8 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -1544,6 +1544,164 @@ Optional. It should have a value much less than 1 (e.g. 0.05 or 0.02) and is
 used to prevent clipping.
 @end table
 
+ at section dynaudnorm
+Dynamic Audio Normalizer.
+
+This filter applies a certain amount of gain to the input audio in order
+to bring its peak magnitude to a target level (e.g. 0 dBFS). However, in
+contrast to more "simple" normalization algorithms, the Dynamic Audio
+Normalizer *dynamically* re-adjusts the gain factor to the input audio.
+This allows for applying extra gain to the "quiet" sections of the audio
+while avoiding distortions or clipping the "loud" sections. In other words:
+The Dynamic Audio Normalizer will "even out" the volume of quiet and loud
+sections, in the sense that the volume of each section is brought to the
+same target level. Note, however, that the Dynamic Audio Normalizer achieves
+this goal *without* applying "dynamic range compressing". It will retain 100%
+of the dynamic range *within* each section of the audio file.
+
+ at table @option
+ at item f
+Set the frame length in milliseconds. In range from 10 to 8000 milliseconds.
+Default is 500 milliseconds.
+The Dynamic Audio Normalizer processes the input audio in small chunks,
+referred to as frames. This is required, because a peak magnitude has no
+meaning for just a single sample value. Instead, we need to determine the
+peak magnitude for a contiguous sequence of sample values. While a "standard"
+normalizer would simply use the peak magnitude of the complete file, the
+Dynamic Audio Normalizer determines the peak magnitude individually for each
+frame. The length of a frame is specified in milliseconds. By default, the
+Dynamic Audio Normalizer uses a frame length of 500 milliseconds, which has
+been found to give good results with most files.
+Note that the exact frame length, in number of samples, will be determined
+automatically, based on the sampling rate of the individual input audio file.
+
+ at item g
+Set the Gaussian filter window size. In range from 3 to 301, must be odd
+number. Default is 31.
+Probably the most important parameter of the Dynamic Audio Normalizer is the
+ at code{window size} of the Gaussian smoothing filter. The filter's window size
+is specified in frames, centered around the current frame. For the sake of
+simplicity, this must be an odd number. Consequently, the default value of 31
+takes into account the current frame, as well as the 15 preceding frames and
+the 15 subsequent frames. Using a larger window results in a stronger
+smoothing effect and thus in less gain variation, i.e. slower gain
+adaptation. Conversely, using a smaller window results in a weaker smoothing
+effect and thus in more gain variation, i.e. faster gain adaptation.
+In other words, the more you increase this value, the more the Dynamic Audio
+Normalizer will behave like a "traditional" normalization filter. On the
+contrary, the more you decrease this value, the more the Dynamic Audio
+Normalizer will behave like a dynamic range compressor.
+
+ at item p
+Set the target peak value. This specifies the highest permissible magnitude
+level for the normalized audio input. This filter will try to approach the
+target peak magnitude as closely as possible, but at the same time it also
+makes sure that the normalized signal will never exceed the peak magnitude.
+A frame's maximum local gain factor is imposed directly by the target peak
+magnitude. The default value is 0.95 and thus leaves a headroom of 5%*.
+It is not recommended to go above this value.
+
+ at item m
+Set the maximum gain factor. In range from 1.0 to 100.0. Default is 10.0.
+The Dynamic Audio Normalizer determines the maximum possible (local) gain
+factor for each input frame, i.e. the maximum gain factor that does not
+result in clipping or distortion. The maximum gain factor is determined by
+the frame's highest magnitude sample. However, the Dynamic Audio Normalizer
+additionally bounds the frame's maximum gain factor by a predetermined
+(global) maximum gain factor. This is done in order to avoid excessive gain
+factors in "silent" or almost silent frames. By default, the maximum gain
+factor is 10.0, For most inputs the default value should be sufficient and
+it usually is not recommended to increase this value. Though, for input
+with an extremely low overall volume level, it may be necessary to allow even
+higher gain factors. Note, however, that the Dynamic Audio Normalizer does
+not simply apply a "hard" threshold (i.e. cut off values above the threshold).
+Instead, a "sigmoid" threshold function will be applied. This way, the
+gain factors will smoothly approach the threshold value, but never exceed that
+value.
+
+ at item r
+Set the target RMS. In range from 0.0 to 1.0. Default is 0.0 - disabled.
+By default, the Dynamic Audio Normalizer performs "peak" normalization.
+This means that the maximum local gain factor for each frame is defined
+(only) by the frame's highest magnitude sample. This way, the samples can
+be amplified as much as possible without exceeding the maximum signal
+level, i.e. without clipping. Optionally, however, the Dynamic Audio
+Normalizer can also take into account the frame's root mean square,
+abbreviated RMS. In electrical engineering, the RMS is commonly used to
+determine the power of a time-varying signal. It is therefore considered
+that the RMS is a better approximation of the "perceived loudness" than
+just looking at the signal's peak magnitude. Consequently, by adjusting all
+frames to a constant RMS value, a uniform "perceived loudness" can be
+established. If a target RMS value has been specified, a frame's local gain
+factor is defined as the factor that would result in exactly that RMS value.
+Note, however, that the maximum local gain factor is still restricted by the
+frame's highest magnitude sample, in order to prevent clipping.
+
+ at item n
+Enable channels coupling. By default is enabled.
+By default, the Dynamic Audio Normalizer will amplify all channels by the same
+amount. This means the same gain factor will be applied to all channels, i.e.
+the maximum possible gain factor is determined by the "loudest" channel.
+However, in some recordings, it may happen that the volume of the different
+channels is uneven, e.g. one channel may be "quieter" than the other one(s).
+In this case, this option can be used to disable the channel coupling. This way,
+the gain factor will be determined independently for each channel, depending
+only on the individual channel's highest magnitude sample. This allows for
+harmonizing the volume of the different channels.
+
+ at item c
+Enable DC bias correction. By default is disabled.
+An audio signal (in the time domain) is a sequence of sample values.
+In the Dynamic Audio Normalizer these sample values are represented in the
+-1.0 to 1.0 range, regardless of the original input format. Normally, the
+audio signal, or "waveform", should be centered around the zero point.
+That means if we calculate the mean value of all samples in a file, or in a
+single frame, then the result should be 0.0 or at least very close to that
+value. If, however, there is a significant deviation of the mean value from
+0.0, in either positive or negative direction, this is referred to as a
+DC bias or DC offset. Since a DC bias is clearly undesirable, the Dynamic
+Audio Normalizer provides optional DC bias correction.
+With DC bias correction enabled, the Dynamic Audio Normalizer will determine
+the mean value, or "DC correction" offset, of each input frame and subtract
+that value from all of the frame's sample values which ensures those samples
+are centered around 0.0 again. Also, in order to avoid "gaps" at the frame
+boundaries, the DC correction offset values will be interpolated smoothly
+between neighbouring frames.
+
+ at item b
+Enable alternative boundary mode. By default is disabled.
+The Dynamic Audio Normalizer takes into account a certain neighbourhood
+around each frame. This includes the preceding frames as well as the
+subsequent frames. However, for the "boundary" frames, located at the very
+beginning and at the very end of the audio file, not all neighbouring
+frames are available. In particular, for the first few frames in the audio
+file, the preceding frames are not known. And, similarly, for the last few
+frames in the audio file, the subsequent frames are not known. Thus, the
+question arises which gain factors should be assumed for the missing frames
+in the "boundary" region. The Dynamic Audio Normalizer implements two modes
+to deal with this situation. The default boundary mode assumes a gain factor
+of exactly 1.0 for the missing frames, resulting in a smooth "fade in" and
+"fade out" at the beginning and at the end of the input, respectively.
+
+ at item s
+Set the compress factor. In range from 0.0 to 30.0. Default is 0.0.
+By default, the Dynamic Audio Normalizer does not apply "traditional"
+compression. This means that signal peaks will not be pruned and thus the
+full dynamic range will be retained within each local neighbourhood. However,
+in some cases it may be desirable to combine the Dynamic Audio Normalizer's
+normalization algorithm with a more "traditional" compression.
+For this purpose, the Dynamic Audio Normalizer provides an optional compression
+(thresholding) function. If (and only if) the compression feature is enabled,
+all input frames will be processed by a soft knee thresholding function prior
+to the actual normalization process. Put simply, the thresholding function is
+going to prune all samples whose magnitude exceeds a certain threshold value.
+However, the Dynamic Audio Normalizer does not simply apply a fixed threshold
+value. Instead, the threshold value will be adjusted for each individual
+frame.
+In general, smaller parameters result in stronger compression, and vice versa.
+Values below 3.0 are not recommended, because audible distortion may appear.
+ at end table
+
 @section earwax
 
 Make audio easier to listen to on headphones.
diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index 058b9e9..a259851 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -67,6 +67,7 @@ OBJS-$(CONFIG_CHANNELSPLIT_FILTER)           += af_channelsplit.o
 OBJS-$(CONFIG_CHORUS_FILTER)                 += af_chorus.o generate_wave_table.o
 OBJS-$(CONFIG_COMPAND_FILTER)                += af_compand.o
 OBJS-$(CONFIG_DCSHIFT_FILTER)                += af_dcshift.o
+OBJS-$(CONFIG_DYNAUDNORM_FILTER)             += af_dynaudnorm.o
 OBJS-$(CONFIG_EARWAX_FILTER)                 += af_earwax.o
 OBJS-$(CONFIG_EBUR128_FILTER)                += f_ebur128.o
 OBJS-$(CONFIG_EQUALIZER_FILTER)              += af_biquads.o
diff --git a/libavfilter/af_dynaudnorm.c b/libavfilter/af_dynaudnorm.c
new file mode 100644
index 0000000..fb83c20
--- /dev/null
+++ b/libavfilter/af_dynaudnorm.c
@@ -0,0 +1,734 @@
+/*
+ * Dynamic Audio Normalizer
+ * Copyright (c) 2015 LoRd_MuldeR <mulder2 at gmx.de>. Some rights reserved.
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * FFmpeg is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with FFmpeg; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+/**
+ * @file
+ * Dynamic Audio Normalizer
+ */
+
+#include <float.h>
+
+#include "libavutil/avassert.h"
+#include "libavutil/opt.h"
+
+#define FF_BUFQUEUE_SIZE 302
+#include "libavfilter/bufferqueue.h"
+
+#include "audio.h"
+#include "avfilter.h"
+#include "internal.h"
+
+typedef struct cqueue {
+    double *elements;
+    int size;
+    int nb_elements;
+    int first;
+} cqueue;
+
+typedef struct DynamicAudioNormalizerContext {
+    const AVClass *class;
+
+    struct FFBufQueue queue;
+
+    int frame_len;
+    int frame_len_msec;
+    int filter_size;
+    int dc_correction;
+    int channels_coupled;
+    int alt_boundary_mode;
+
+    double peak_value;
+    double max_amplification;
+    double target_rms;
+    double compress_factor;
+    double *prev_amplification_factor;
+    double *dc_correction_value;
+    double *compress_threshold;
+    double *fade_factors[2];
+    double *weights;
+
+    int channels;
+    int delay;
+
+    cqueue **gain_history_original;
+    cqueue **gain_history_minimum;
+    cqueue **gain_history_smoothed;
+} DynamicAudioNormalizerContext;
+
+#define OFFSET(x) offsetof(DynamicAudioNormalizerContext, x)
+#define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM
+
+static const AVOption dynaudnorm_options[] = {
+    { "f", "set the frame length in msec",     OFFSET(frame_len_msec),    AV_OPT_TYPE_INT,    {.i64 = 500},   10,  8000, FLAGS },
+    { "g", "set the filter size",              OFFSET(filter_size),       AV_OPT_TYPE_INT,    {.i64 = 31},     3,   301, FLAGS },
+    { "p", "set the peak value",               OFFSET(peak_value),        AV_OPT_TYPE_DOUBLE, {.dbl = 0.95}, 0.0,   1.0, FLAGS },
+    { "m", "set the max amplification",        OFFSET(max_amplification), AV_OPT_TYPE_DOUBLE, {.dbl = 10.0}, 1.0, 100.0, FLAGS },
+    { "r", "set the target RMS",               OFFSET(target_rms),        AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,   1.0, FLAGS },
+    { "n", "enable channel coupling",          OFFSET(channels_coupled),  AV_OPT_TYPE_INT,    {.i64 = 1},      0,     1, FLAGS },
+    { "c", "enable DC correction",             OFFSET(dc_correction),     AV_OPT_TYPE_INT,    {.i64 = 0},      0,     1, FLAGS },
+    { "b", "enable alternative boundary mode", OFFSET(alt_boundary_mode), AV_OPT_TYPE_INT,    {.i64 = 0},      0,     1, FLAGS },
+    { "s", "set the compress factor",          OFFSET(compress_factor),   AV_OPT_TYPE_DOUBLE, {.dbl = 0.0},  0.0,  30.0, FLAGS },
+    { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(dynaudnorm);
+
+static av_cold int init(AVFilterContext *ctx)
+{
+    DynamicAudioNormalizerContext *s = ctx->priv;
+
+    if (!(s->filter_size & 1)) {
+        av_log(ctx, AV_LOG_ERROR, "filter size %d is invalid. Must be an odd value.\n", s->filter_size);
+        return AVERROR(EINVAL);
+    }
+
+    return 0;
+}
+
+static int query_formats(AVFilterContext *ctx)
+{
+    AVFilterFormats *formats;
+    AVFilterChannelLayouts *layouts;
+    static const enum AVSampleFormat sample_fmts[] = {
+        AV_SAMPLE_FMT_DBLP,
+        AV_SAMPLE_FMT_NONE
+    };
+    int ret;
+
+    layouts = ff_all_channel_layouts();
+    if (!layouts)
+        return AVERROR(ENOMEM);
+    ret = ff_set_common_channel_layouts(ctx, layouts);
+    if (ret < 0)
+        return ret;
+
+    formats = ff_make_format_list(sample_fmts);
+    if (!formats)
+        return AVERROR(ENOMEM);
+    ret = ff_set_common_formats(ctx, formats);
+    if (ret < 0)
+        return ret;
+
+    formats = ff_all_samplerates();
+    if (!formats)
+        return AVERROR(ENOMEM);
+    return ff_set_common_samplerates(ctx, formats);
+}
+
+static inline int frame_size(int sample_rate, int frame_len_msec)
+{
+    const int frame_size = round((double)sample_rate * (frame_len_msec / 1000.0));
+    return frame_size + (frame_size % 2);
+}
+
+static void precalculate_fade_factors(double *fade_factors[2], int frame_len)
+{
+    const double step_size = 1.0 / frame_len;
+    int pos;
+
+    for (pos = 0; pos < frame_len; pos++) {
+        fade_factors[0][pos] = 1.0 - (step_size * (pos + 1.0));
+        fade_factors[1][pos] = 1.0 - fade_factors[0][pos];
+    }
+}
+
+static cqueue *cqueue_create(int size)
+{
+    cqueue *q;
+
+    q = av_malloc(sizeof(cqueue));
+    if (!q)
+        return NULL;
+
+    q->size = size;
+    q->nb_elements = 0;
+    q->first = 0;
+
+    q->elements = av_malloc(sizeof(double) * size);
+    if (!q->elements) {
+        av_free(q);
+        return NULL;
+    }
+
+    return q;
+}
+
+static void cqueue_free(cqueue *q)
+{
+    av_free(q->elements);
+    av_free(q);
+}
+
+static int cqueue_size(cqueue *q)
+{
+    return q->nb_elements;
+}
+
+static int cqueue_empty(cqueue *q)
+{
+    return !q->nb_elements;
+}
+
+static int cqueue_enqueue(cqueue *q, double element)
+{
+    int i;
+
+    av_assert2(q->nb_elements |= q->size);
+
+    i = (q->first + q->nb_elements) % q->size;
+    q->elements[i] = element;
+    q->nb_elements++;
+
+    return 0;
+}
+
+static double cqueue_peek(cqueue *q, int index)
+{
+    av_assert2(index < q->nb_elements);
+    return q->elements[(q->first + index) % q->size];
+}
+
+static int cqueue_dequeue(cqueue *q, double *element)
+{
+    av_assert2(!cqueue_empty(q));
+
+    *element = q->elements[q->first];
+    q->first = (q->first + 1) % q->size;
+    q->nb_elements--;
+
+    return 0;
+}
+
+static int cqueue_pop(cqueue *q)
+{
+    av_assert2(!cqueue_empty(q));
+
+    q->first = (q->first + 1) % q->size;
+    q->nb_elements--;
+
+    return 0;
+}
+
+static const double s_pi = 3.1415926535897932384626433832795028841971693993751058209749445923078164062862089986280348253421170679;
+
+static void init_gaussian_filter(DynamicAudioNormalizerContext *s)
+{
+    double total_weight = 0.0;
+    const double sigma = (((s->filter_size / 2.0) - 1.0) / 3.0) + (1.0 / 3.0);
+    double adjust;
+    int i;
+
+    // Pre-compute constants
+    const int offset = s->filter_size / 2;
+    const double c1 = 1.0 / (sigma * sqrt(2.0 * s_pi));
+    const double c2 = 2.0 * pow(sigma, 2.0);
+
+    // Compute weights
+    for (i = 0; i < s->filter_size; i++) {
+        const int x = i - offset;
+
+        s->weights[i] = c1 * exp(-(pow(x, 2.0) / c2));
+        total_weight += s->weights[i];
+    }
+
+    // Adjust weights
+    adjust = 1.0 / total_weight;
+    for (i = 0; i < s->filter_size; i++) {
+        s->weights[i] *= adjust;
+    }
+}
+
+static int config_input(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx = inlink->dst;
+    DynamicAudioNormalizerContext *s = ctx->priv;
+    int c;
+
+    s->frame_len =
+    inlink->min_samples =
+    inlink->max_samples =
+    inlink->partial_buf_size = frame_size(inlink->sample_rate, s->frame_len_msec);
+    av_log(ctx, AV_LOG_DEBUG, "frame len %d\n", s->frame_len);
+
+    s->fade_factors[0] = av_malloc(s->frame_len * sizeof(*s->fade_factors[0]));
+    s->fade_factors[1] = av_malloc(s->frame_len * sizeof(*s->fade_factors[1]));
+
+    s->prev_amplification_factor = av_malloc(inlink->channels * sizeof(*s->prev_amplification_factor));
+    s->dc_correction_value = av_calloc(inlink->channels, sizeof(*s->dc_correction_value));
+    s->compress_threshold = av_calloc(inlink->channels, sizeof(*s->compress_threshold));
+    s->gain_history_original = av_calloc(inlink->channels, sizeof(*s->gain_history_original));
+    s->gain_history_minimum = av_calloc(inlink->channels, sizeof(*s->gain_history_minimum));
+    s->gain_history_smoothed = av_calloc(inlink->channels, sizeof(*s->gain_history_smoothed));
+    s->weights = av_malloc(s->filter_size * sizeof(*s->weights));
+    if (!s->prev_amplification_factor || !s->dc_correction_value ||
+        !s->compress_threshold || !s->fade_factors[0] || !s->fade_factors[1] ||
+        !s->gain_history_original || !s->gain_history_minimum ||
+        !s->gain_history_smoothed || !s->weights)
+        return AVERROR(ENOMEM);
+
+    for (c = 0; c < inlink->channels; c++) {
+        s->prev_amplification_factor[c] = 1.0;
+
+        s->gain_history_original[c] = cqueue_create(s->filter_size);
+        s->gain_history_minimum[c]  = cqueue_create(s->filter_size);
+        s->gain_history_smoothed[c] = cqueue_create(s->filter_size);
+
+        if (!s->gain_history_original[c] || !s->gain_history_minimum[c] ||
+            !s->gain_history_smoothed[c])
+            return AVERROR(ENOMEM);
+    }
+
+    precalculate_fade_factors(s->fade_factors, s->frame_len);
+    init_gaussian_filter(s);
+
+    s->channels = inlink->channels;
+    s->delay = s->filter_size;
+
+    return 0;
+}
+
+static int config_output(AVFilterLink *outlink)
+{
+    outlink->flags |= FF_LINK_FLAG_REQUEST_LOOP;
+    return 0;
+}
+
+static inline double fade(double prev, double next, int pos,
+                          double *fade_factors[2])
+{
+    return fade_factors[0][pos] * prev + fade_factors[1][pos] * next;
+}
+
+static inline double pow2(const double value)
+{
+    return value * value;
+}
+
+static inline double bound(const double threshold, const double val)
+{
+    const double CONST = 0.8862269254527580136490837416705725913987747280611935; //sqrt(PI) / 2.0
+    return erf(CONST * (val / threshold)) * threshold;
+}
+
+static double find_peak_magnitude(AVFrame *frame, int channel)
+{
+    double max = DBL_EPSILON;
+    int c, i;
+
+    if (channel == -1) {
+        for (c = 0; c < frame->channels; c++) {
+            double *data_ptr = (double *)frame->extended_data[c];
+
+            for (i = 0; i < frame->nb_samples; i++)
+                max = FFMAX(max, fabs(data_ptr[i]));
+        }
+    } else {
+        double *data_ptr = (double *)frame->extended_data[channel];
+
+        for (i = 0; i < frame->nb_samples; i++)
+            max = FFMAX(max, fabs(data_ptr[i]));
+    }
+
+    return max;
+}
+
+static double compute_frame_rms(AVFrame *frame, int channel)
+{
+    double rms_value = 0.0;
+    int c, i;
+
+    if (channel == -1) {
+        for (c = 0; c < frame->channels; c++) {
+            const double *data_ptr = (double *)frame->extended_data[c];
+
+            for (i = 0; i < frame->nb_samples; i++) {
+                rms_value += pow2(data_ptr[i]);
+            }
+        }
+
+        rms_value /= frame->nb_samples * frame->channels;
+    } else {
+        const double *data_ptr = (double *)frame->extended_data[channel];
+        for (i = 0; i < frame->nb_samples; i++) {
+            rms_value += pow2(data_ptr[i]);
+        }
+
+        rms_value /= frame->nb_samples;
+    }
+
+    return FFMAX(sqrt(rms_value), DBL_EPSILON);
+}
+
+static double get_max_local_gain(DynamicAudioNormalizerContext *s, AVFrame *frame,
+                                 int channel)
+{
+    const double maximum_gain = s->peak_value / find_peak_magnitude(frame, channel);
+    const double rms_gain = s->target_rms > DBL_EPSILON ? (s->target_rms / compute_frame_rms(frame, channel)) : DBL_MAX;
+    return bound(s->max_amplification, FFMIN(maximum_gain, rms_gain));
+}
+
+static double minimum_filter(cqueue *q)
+{
+    double min = DBL_MAX;
+    int i;
+
+    for (i = 0; i < cqueue_size(q); i++) {
+        min = FFMIN(min, cqueue_peek(q, i));
+    }
+
+    return min;
+}
+
+static double gaussian_filter(DynamicAudioNormalizerContext *s, cqueue *q)
+{
+    double result = 0.0;
+    int i;
+
+    for (i = 0; i < cqueue_size(q); i++) {
+        result += cqueue_peek(q, i) * s->weights[i];
+    }
+
+    return result;
+}
+
+static void update_gain_history(DynamicAudioNormalizerContext *s, int channel,
+                                double current_gain_factor)
+{
+    if (cqueue_empty(s->gain_history_original[channel]) ||
+        cqueue_empty(s->gain_history_minimum[channel])) {
+        const int pre_fill_size = s->filter_size / 2;
+
+        s->prev_amplification_factor[channel] = s->alt_boundary_mode ? current_gain_factor : 1.0;
+
+        while (cqueue_size(s->gain_history_original[channel]) < pre_fill_size) {
+            cqueue_enqueue(s->gain_history_original[channel], s->alt_boundary_mode ? current_gain_factor : 1.0);
+        }
+
+        while (cqueue_size(s->gain_history_minimum[channel]) < pre_fill_size) {
+            cqueue_enqueue(s->gain_history_minimum[channel], s->alt_boundary_mode ? current_gain_factor : 1.0);
+        }
+    }
+
+    cqueue_enqueue(s->gain_history_original[channel], current_gain_factor);
+
+    while (cqueue_size(s->gain_history_original[channel]) >= s->filter_size) {
+        av_assert0(cqueue_size(s->gain_history_original[channel]) == s->filter_size);
+        const double minimum = minimum_filter(s->gain_history_original[channel]);
+
+        cqueue_enqueue(s->gain_history_minimum[channel], minimum);
+
+        cqueue_pop(s->gain_history_original[channel]);
+    }
+
+    while (cqueue_size(s->gain_history_minimum[channel]) >= s->filter_size) {
+        av_assert0(cqueue_size(s->gain_history_minimum[channel]) == s->filter_size);
+        const double smoothed = gaussian_filter(s, s->gain_history_minimum[channel]);
+
+        cqueue_enqueue(s->gain_history_smoothed[channel], smoothed);
+
+        cqueue_pop(s->gain_history_minimum[channel]);
+    }
+}
+
+static inline double update_value(double new, double old, double aggressiveness)
+{
+    av_assert0((aggressiveness >= 0.0) && (aggressiveness <= 1.0));
+    return aggressiveness * new + (1.0 - aggressiveness) * old;
+}
+
+static void perform_dc_correction(DynamicAudioNormalizerContext *s, AVFrame *frame)
+{
+    const double diff = 1.0 / frame->nb_samples;
+    int is_first_frame = cqueue_empty(s->gain_history_original[0]);
+    int c, i;
+
+    for (c = 0; c < s->channels; c++) {
+        double *dst_ptr = (double *)frame->extended_data[c];
+        double current_average_value = 0.0;
+
+        for (i = 0; i < frame->nb_samples; i++)
+            current_average_value += dst_ptr[i] * diff;
+
+        const double prev_value = is_first_frame ? current_average_value : s->dc_correction_value[c];
+        s->dc_correction_value[c] = is_first_frame ? current_average_value : update_value(current_average_value, s->dc_correction_value[c], 0.1);
+
+        for (i = 0; i < frame->nb_samples; i++) {
+            dst_ptr[i] -= fade(prev_value, s->dc_correction_value[c], i, s->fade_factors);
+        }
+    }
+}
+
+static double setup_compress_thresh(double threshold)
+{
+    if ((threshold > DBL_EPSILON) && (threshold < (1.0 - DBL_EPSILON))) {
+        double current_threshold = threshold;
+        double step_size = 1.0;
+
+        while (step_size > DBL_EPSILON) {
+            while ((current_threshold + step_size > current_threshold) &&
+                   (bound(current_threshold + step_size, 1.0) <= threshold)) {
+                current_threshold += step_size;
+            }
+
+            step_size /= 2.0;
+        }
+
+        return current_threshold;
+    } else {
+        return threshold;
+    }
+}
+
+static double compute_frame_std_dev(DynamicAudioNormalizerContext *s,
+                                    AVFrame *frame, int channel)
+{
+    double variance = 0.0;
+    int i, c;
+
+    if (channel == -1) {
+        for (c = 0; c < s->channels; c++) {
+            const double *data_ptr = (double *)frame->extended_data[c];
+
+            for (i = 0; i < frame->nb_samples; i++) {
+                variance += pow2(data_ptr[i]);  // Assume that MEAN is *zero*
+            }
+        }
+        variance /= (s->channels * frame->nb_samples) - 1;
+    } else {
+        const double *data_ptr = (double *)frame->extended_data[channel];
+
+        for (i = 0; i < frame->nb_samples; i++) {
+            variance += pow2(data_ptr[i]);      // Assume that MEAN is *zero*
+        }
+        variance /= frame->nb_samples - 1;
+    }
+
+    return FFMAX(sqrt(variance), DBL_EPSILON);
+}
+
+static void perform_compression(DynamicAudioNormalizerContext *s, AVFrame *frame)
+{
+    int is_first_frame = cqueue_empty(s->gain_history_original[0]);
+    int c, i;
+
+    if (s->channels_coupled) {
+        const double standard_deviation = compute_frame_std_dev(s, frame, -1);
+        const double current_threshold  = FFMIN(1.0, s->compress_factor * standard_deviation);
+
+        const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[0];
+        s->compress_threshold[0] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[0], (1.0/3.0));
+
+        const double prev_actual_thresh = setup_compress_thresh(prev_value);
+        const double curr_actual_thresh = setup_compress_thresh(s->compress_threshold[0]);
+
+        for (c = 0; c < s->channels; c++) {
+            double *const dst_ptr = (double *)frame->extended_data[c];
+            for (i = 0; i < frame->nb_samples; i++) {
+                const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
+                dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
+            }
+        }
+    } else {
+        for (c = 0; c < s->channels; c++) {
+            const double standard_deviation = compute_frame_std_dev(s, frame, c);
+            const double current_threshold  = setup_compress_thresh(FFMIN(1.0, s->compress_factor * standard_deviation));
+
+            const double prev_value = is_first_frame ? current_threshold : s->compress_threshold[c];
+            s->compress_threshold[c] = is_first_frame ? current_threshold : update_value(current_threshold, s->compress_threshold[c], 1.0/3.0);
+
+            const double prev_actual_thresh = setup_compress_thresh(prev_value);
+            const double curr_actual_thresh = setup_compress_thresh(s->compress_threshold[c]);
+
+            double *const dst_ptr = (double *)frame->extended_data[c];
+            for (i = 0; i < frame->nb_samples; i++) {
+                const double localThresh = fade(prev_actual_thresh, curr_actual_thresh, i, s->fade_factors);
+                dst_ptr[i] = copysign(bound(localThresh, fabs(dst_ptr[i])), dst_ptr[i]);
+            }
+        }
+    }
+}
+
+static void analyze_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
+{
+    if (s->dc_correction) {
+        perform_dc_correction(s, frame);
+    }
+
+    if (s->compress_factor > DBL_EPSILON) {
+        perform_compression(s, frame);
+    }
+
+    if (s->channels_coupled) {
+        const double current_gain_factor = get_max_local_gain(s, frame, -1);
+        int c;
+
+        for (c = 0; c < s->channels; c++)
+            update_gain_history(s, c, current_gain_factor);
+    } else {
+        int c;
+
+        for (c = 0; c < s->channels; c++)
+            update_gain_history(s, c, get_max_local_gain(s, frame, c));
+    }
+}
+
+static void amplify_frame(DynamicAudioNormalizerContext *s, AVFrame *frame)
+{
+    int c, i;
+
+    for (c = 0; c < s->channels; c++) {
+        double *dst_ptr = (double *)frame->extended_data[c];
+        double current_amplification_factor;
+
+        cqueue_dequeue(s->gain_history_smoothed[c], &current_amplification_factor);
+
+        for (i = 0; i < frame->nb_samples; i++) {
+            const double amplification_factor = fade(s->prev_amplification_factor[c],
+                                                     current_amplification_factor, i,
+                                                     s->fade_factors);
+
+            dst_ptr[i] *= amplification_factor;
+
+            if (fabs(dst_ptr[i]) > s->peak_value)
+                dst_ptr[i] = copysign(s->peak_value, dst_ptr[i]);
+        }
+
+        s->prev_amplification_factor[c] = current_amplification_factor;
+    }
+}
+
+static int filter_frame(AVFilterLink *inlink, AVFrame *in)
+{
+    AVFilterContext *ctx = inlink->dst;
+    DynamicAudioNormalizerContext *s = ctx->priv;
+    AVFilterLink *outlink = inlink->dst->outputs[0];
+    int ret = 0;
+
+    if (!cqueue_empty(s->gain_history_smoothed[0])) {
+        AVFrame *out = ff_bufqueue_get(&s->queue);
+
+        amplify_frame(s, out);
+        ret = ff_filter_frame(outlink, out);
+    }
+
+    analyze_frame(s, in);
+    ff_bufqueue_add(ctx, &s->queue, in);
+
+    return ret;
+}
+
+static int flush_buffer(DynamicAudioNormalizerContext *s, AVFilterLink *inlink,
+                        AVFilterLink *outlink)
+{
+    AVFrame *out = ff_get_audio_buffer(outlink, s->frame_len);
+    int c, i;
+
+    if (!out)
+        return AVERROR(ENOMEM);
+
+    for (c = 0; c < s->channels; c++) {
+        double *dst_ptr = (double *)out->extended_data[c];
+
+        for (i = 0; i < out->nb_samples; i++) {
+            dst_ptr[i] = s->alt_boundary_mode ? DBL_EPSILON : ((s->target_rms > DBL_EPSILON) ? FFMIN(s->peak_value, s->target_rms) : s->peak_value);
+            if (s->dc_correction) {
+                dst_ptr[i] *= ((i % 2) == 1) ? -1 : 1;
+                dst_ptr[i] += s->dc_correction_value[c];
+            }
+        }
+    }
+
+    s->delay--;
+    return filter_frame(inlink, out);
+}
+
+static int request_frame(AVFilterLink *outlink)
+{
+    AVFilterContext *ctx = outlink->src;
+    DynamicAudioNormalizerContext *s = ctx->priv;
+    int ret = 0;
+
+    ret = ff_request_frame(ctx->inputs[0]);
+
+    if (ret == AVERROR_EOF && !ctx->is_disabled && s->delay)
+        ret = flush_buffer(s, ctx->inputs[0], outlink);
+
+    return ret;
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    DynamicAudioNormalizerContext *s = ctx->priv;
+    int c;
+
+    av_freep(&s->prev_amplification_factor);
+    av_freep(&s->dc_correction_value);
+    av_freep(&s->compress_threshold);
+    av_freep(&s->fade_factors[0]);
+    av_freep(&s->fade_factors[1]);
+
+    for (c = 0; c < s->channels; c++) {
+        cqueue_free(s->gain_history_original[c]);
+        cqueue_free(s->gain_history_minimum[c]);
+        cqueue_free(s->gain_history_smoothed[c]);
+    }
+
+    av_freep(&s->gain_history_original);
+    av_freep(&s->gain_history_minimum);
+    av_freep(&s->gain_history_smoothed);
+
+    av_freep(&s->weights);
+
+    ff_bufqueue_discard_all(&s->queue);
+}
+
+static const AVFilterPad avfilter_af_dynaudnorm_inputs[] = {
+    {
+        .name           = "default",
+        .type           = AVMEDIA_TYPE_AUDIO,
+        .filter_frame   = filter_frame,
+        .config_props   = config_input,
+        .needs_writable = 1,
+    },
+    { NULL }
+};
+
+static const AVFilterPad avfilter_af_dynaudnorm_outputs[] = {
+    {
+        .name          = "default",
+        .type          = AVMEDIA_TYPE_AUDIO,
+        .config_props  = config_output,
+        .request_frame = request_frame,
+    },
+    { NULL }
+};
+
+AVFilter ff_af_dynaudnorm = {
+    .name          = "dynaudnorm",
+    .description   = NULL_IF_CONFIG_SMALL("Dynamic Audio Normalizer."),
+    .query_formats = query_formats,
+    .priv_size     = sizeof(DynamicAudioNormalizerContext),
+    .init          = init,
+    .uninit        = uninit,
+    .inputs        = avfilter_af_dynaudnorm_inputs,
+    .outputs       = avfilter_af_dynaudnorm_outputs,
+    .priv_class    = &dynaudnorm_class,
+};
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index b0d8410..01c9e38 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -83,6 +83,7 @@ void avfilter_register_all(void)
     REGISTER_FILTER(CHORUS,         chorus,         af);
     REGISTER_FILTER(COMPAND,        compand,        af);
     REGISTER_FILTER(DCSHIFT,        dcshift,        af);
+    REGISTER_FILTER(DYNAUDNORM,     dynaudnorm,     af);
     REGISTER_FILTER(EARWAX,         earwax,         af);
     REGISTER_FILTER(EBUR128,        ebur128,        af);
     REGISTER_FILTER(EQUALIZER,      equalizer,      af);



More information about the ffmpeg-cvslog mailing list