[FFmpeg-devel] [PATCH] avfilter: add speechnorm filter

Paul B Mahol onemda at gmail.com
Sat Nov 7 01:27:15 EET 2020


Signed-off-by: Paul B Mahol <onemda at gmail.com>
---
 doc/filters.texi            |  49 +++++
 libavfilter/Makefile        |   1 +
 libavfilter/af_speechnorm.c | 422 ++++++++++++++++++++++++++++++++++++
 libavfilter/allfilters.c    |   1 +
 4 files changed, 473 insertions(+)
 create mode 100644 libavfilter/af_speechnorm.c

diff --git a/doc/filters.texi b/doc/filters.texi
index 8380f6cac2..7db43dd34c 100644
--- a/doc/filters.texi
+++ b/doc/filters.texi
@@ -5276,6 +5276,55 @@ and also with custom gain:
 @end example
 @end itemize
 
+ at section speechnorm
+Speech Normalizer.
+
+This filter expands or compress each half-cycle of audio samples
+(local set of samples all above or all bellow zero) depending on threshold value,
+so audio reaches target peak value under conditions controlled by bellow options.
+
+The filter accepts the following options:
+
+ at table @option
+ at item peak, p
+Set the expansion target peak value. This specifies the highest allowed absolute amplitude
+level for the normalized audio input. Default value is 0.95. Allowed range is from 0.0 to 1.0.
+
+ at item expansion, e
+Set the maximum expansion factor. Allowed range is from 1.0 to 10.0. Default value is 2.0.
+This option controls maximum local half-cycle of samples expansion. The maximum expansion
+would be such that local peak value reach target peak value but never to surpass it and that
+ratio between new and previous peak value does not surpass this option value.
+
+ at item compression, c
+Set the maximum compression factor. Allowed range is from 1.0 to 10.0. Default value is 2.0.
+This option controls maximum local half-cycle of samples compression. This option is used
+only if @option{threshold} option is set to value greater than 0.0, than in such cases
+when local peak is lower or same as value set by @option{threshold} all samples belonging to
+that peak's half-cycle would be compressed by current compression factor.
+
+ at item threshold, t
+Set the threshold value. Default value is 0.0. Allowed range is from 0.0 to 1.0.
+This option specify which half-cycles of samples will be compressed and which will be expanded.
+Any half-cycle samples with their local peak value bellow or same as this option value will be
+compressed by current compression factor, otherwise, if greater than threshold value it will be
+expanded with expansion factor so that it could reach peak target value but never surpass it.
+
+ at item raise, r
+Set the expansion raising amount per each half-cycle of samples. Default value is 0.001.
+Allowed range is from 0.0 to 1.0. This controls how fast expansion factor is raised per
+each new half-cycle until it reaches @option{expansion} value.
+
+ at item fall, f
+Set the compression raising amount per each half-cycle of samples. Default value is 0.001.
+Allowed range is from 0.0 to 1.0. This controls how fast compression factor is raised per
+each new half-cycle until it reaches @option{compression} value.
+ at end table
+
+ at subsection Commands
+
+This filter supports the all above options as @ref{commands}.
+
 @section stereotools
 
 This filter has some handy utilities to manage stereo signals, for converting
diff --git a/libavfilter/Makefile b/libavfilter/Makefile
index 0c2a5d1cf4..36f3d2d0e4 100644
--- a/libavfilter/Makefile
+++ b/libavfilter/Makefile
@@ -138,6 +138,7 @@ OBJS-$(CONFIG_SIDECHAINGATE_FILTER)          += af_agate.o
 OBJS-$(CONFIG_SILENCEDETECT_FILTER)          += af_silencedetect.o
 OBJS-$(CONFIG_SILENCEREMOVE_FILTER)          += af_silenceremove.o
 OBJS-$(CONFIG_SOFALIZER_FILTER)              += af_sofalizer.o
+OBJS-$(CONFIG_SPEECHNORM_FILTER)             += af_speechnorm.o
 OBJS-$(CONFIG_STEREOTOOLS_FILTER)            += af_stereotools.o
 OBJS-$(CONFIG_STEREOWIDEN_FILTER)            += af_stereowiden.o
 OBJS-$(CONFIG_SUPEREQUALIZER_FILTER)         += af_superequalizer.o
diff --git a/libavfilter/af_speechnorm.c b/libavfilter/af_speechnorm.c
new file mode 100644
index 0000000000..ef06c72fee
--- /dev/null
+++ b/libavfilter/af_speechnorm.c
@@ -0,0 +1,422 @@
+/*
+ * Speech Normalizer
+ * Copyright (c) 2020 Paul B Mahol
+ *
+ * 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
+ * Speech Normalizer
+ */
+
+#include <float.h>
+
+#include "libavutil/avassert.h"
+#include "libavutil/opt.h"
+
+#define FF_BUFQUEUE_SIZE (1024)
+#include "bufferqueue.h"
+
+#include "audio.h"
+#include "avfilter.h"
+#include "filters.h"
+#include "internal.h"
+
+#define MAX_ITEMS  882000
+
+typedef struct PeriodItem {
+    int size;
+    int type;
+    double max_peak;
+} PeriodItem;
+
+typedef struct ChannelContext {
+    int state;
+    PeriodItem pi[MAX_ITEMS];
+    double gain_state;
+    int pi_start;
+    int pi_end;
+} ChannelContext;
+
+typedef struct SpeechNormalizerContext {
+    const AVClass *class;
+
+    double peak_value;
+    double max_expansion;
+    double max_compression;
+    double threshold_value;
+    double raise_amount;
+    double fall_amount;
+    int channels;
+
+    ChannelContext *cc;
+
+    int max_period;
+    int eof;
+    int64_t pts;
+
+    struct FFBufQueue queue;
+} SpeechNormalizerContext;
+
+#define OFFSET(x) offsetof(SpeechNormalizerContext, x)
+#define FLAGS AV_OPT_FLAG_AUDIO_PARAM|AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_RUNTIME_PARAM
+
+static const AVOption speechnorm_options[] = {
+    { "peak", "set the peak value", OFFSET(peak_value), AV_OPT_TYPE_DOUBLE, {.dbl=0.95}, 0.0, 1.0, FLAGS },
+    { "p",    "set the peak value", OFFSET(peak_value), AV_OPT_TYPE_DOUBLE, {.dbl=0.95}, 0.0, 1.0, FLAGS },
+    { "expansion", "set the max expansion factor", OFFSET(max_expansion), AV_OPT_TYPE_DOUBLE, {.dbl=2.0}, 1.0, 10.0, FLAGS },
+    { "e",         "set the max expansion factor", OFFSET(max_expansion), AV_OPT_TYPE_DOUBLE, {.dbl=2.0}, 1.0, 10.0, FLAGS },
+    { "compresion", "set the max compression factor", OFFSET(max_compression), AV_OPT_TYPE_DOUBLE, {.dbl=2.0}, 1.0, 10.0, FLAGS },
+    { "c",          "set the max compression factor", OFFSET(max_compression), AV_OPT_TYPE_DOUBLE, {.dbl=2.0}, 1.0, 10.0, FLAGS },
+    { "threshold", "set the threshold value", OFFSET(threshold_value), AV_OPT_TYPE_DOUBLE, {.dbl=0}, 0.0, 1.0, FLAGS },
+    { "t",         "set the threshold value", OFFSET(threshold_value), AV_OPT_TYPE_DOUBLE, {.dbl=0}, 0.0, 1.0, FLAGS },
+    { "raise", "set the expansion raising amount", OFFSET(raise_amount), AV_OPT_TYPE_DOUBLE, {.dbl=0.001}, 0.0, 1.0, FLAGS },
+    { "r",     "set the expansion raising amount", OFFSET(raise_amount), AV_OPT_TYPE_DOUBLE, {.dbl=0.001}, 0.0, 1.0, FLAGS },
+    { "fall", "set the compression raising amount", OFFSET(fall_amount), AV_OPT_TYPE_DOUBLE, {.dbl=0.001}, 0.0, 1.0, FLAGS },
+    { "f",    "set the compression raising amount", OFFSET(fall_amount), AV_OPT_TYPE_DOUBLE, {.dbl=0.001}, 0.0, 1.0, FLAGS },
+    { NULL }
+};
+
+AVFILTER_DEFINE_CLASS(speechnorm);
+
+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_counts();
+    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 int config_input(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx = inlink->dst;
+    SpeechNormalizerContext *s = ctx->priv;
+
+    s->max_period = inlink->sample_rate / (2 * 20);
+    s->channels = inlink->channels;
+
+    s->cc = av_calloc(inlink->channels, sizeof(*s->cc));
+    if (!s->cc)
+        return AVERROR(ENOMEM);
+
+    for (int ch = 0; ch < s->channels; ch++) {
+        ChannelContext *cc = &s->cc[ch];
+
+        cc->state = -1;
+        cc->gain_state = 1.;
+    }
+
+    return 0;
+}
+
+static int get_pi_samples(PeriodItem *pi, int start, int end, int mode)
+{
+    int sum;
+
+    if (mode && pi[start].type == 0)
+        return 0;
+
+    sum = pi[start].size;
+    av_assert0(sum >= 0);
+    while (start != end) {
+        start++;
+        if (start >= MAX_ITEMS)
+            start = 0;
+        if (mode && pi[start].type == 0)
+            break;
+        av_assert0(pi[start].size > 0);
+        sum += pi[start].size;
+        if (pi[start].type == 0)
+            break;
+    }
+
+    return sum;
+}
+
+static int consume_pi(PeriodItem *pi, int start, int end, int nb_samples)
+{
+    int sum;
+
+    sum = pi[start].size;
+    av_assert0(pi[start].size > 0);
+    while (sum < nb_samples) {
+        av_assert0(pi[start].type == 1);
+        av_assert0(start != end);
+        start++;
+        if (start >= MAX_ITEMS)
+            start = 0;
+        av_assert0(pi[start].size > 0);
+        sum += pi[start].size;
+    }
+
+    av_assert0(pi[start].size >= sum - nb_samples);
+    pi[start].size = sum - nb_samples;
+    av_assert0(pi[start].size >= 0);
+    if (pi[start].size == 0 && start != end) {
+        start++;
+        if (start >= MAX_ITEMS)
+            start = 0;
+    }
+
+    return start;
+}
+
+static int get_queued_samples(SpeechNormalizerContext *s)
+{
+    int sum = 0;
+
+    for (int i = 0; i < s->queue.available; i++) {
+        AVFrame *frame = ff_bufqueue_peek(&s->queue, i);
+        sum += frame->nb_samples;
+    }
+
+    return sum;
+}
+
+static int filter_frame(AVFilterContext *ctx)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+    AVFilterLink *outlink = ctx->outputs[0];
+    AVFilterLink *inlink = ctx->inputs[0];
+    int min_pi_nb_samples;
+    AVFrame *in = NULL;
+    int ret;
+
+    for (int f = 0; f < ff_inlink_queued_frames(inlink); f++) {
+        ret = ff_inlink_consume_frame(inlink, &in);
+        if (ret < 0)
+            return ret;
+        if (ret == 0)
+            break;
+
+        ff_bufqueue_add(ctx, &s->queue, in);
+
+        for (int ch = 0; ch < inlink->channels; ch++) {
+            ChannelContext *cc = &s->cc[ch];
+            const double *src = (const double *)in->extended_data[ch];
+            int n = 0;
+
+            if (cc->state < 0)
+                cc->state = src[0] >= 0.;
+
+            while (n < in->nb_samples) {
+                if (cc->state != (src[n] >= 0.) || cc->pi[cc->pi_end].size > s->max_period) {
+                    cc->state = src[n] >= 0.;
+                    av_assert0(cc->pi[cc->pi_end].size > 0);
+                    cc->pi[cc->pi_end].type = 1;
+                    cc->pi_end++;
+                    if (cc->pi_end >= MAX_ITEMS)
+                        cc->pi_end = 0;
+                    cc->pi[cc->pi_end].max_peak = DBL_MIN;
+                    cc->pi[cc->pi_end].type = 0;
+                    cc->pi[cc->pi_end].size = 0;
+                    av_assert0(cc->pi_end != cc->pi_start);
+                }
+
+                if (src[n] >= 0.) {
+                    while (src[n] >= 0.) {
+                        cc->pi[cc->pi_end].max_peak = FFMAX(cc->pi[cc->pi_end].max_peak, FFABS(src[n]));
+                        cc->pi[cc->pi_end].size++;
+                        n++;
+                        if (n >= in->nb_samples)
+                            break;
+                    }
+                } else {
+                    while (src[n] < 0.) {
+                        cc->pi[cc->pi_end].max_peak = FFMAX(cc->pi[cc->pi_end].max_peak, FFABS(src[n]));
+                        cc->pi[cc->pi_end].size++;
+                        n++;
+                        if (n >= in->nb_samples)
+                            break;
+                    }
+                }
+            }
+        }
+    }
+
+    if (s->queue.available > 0) {
+        in = ff_bufqueue_peek(&s->queue, 0);
+        if (!in)
+            return 1;
+    } else {
+        return 1;
+    }
+
+    min_pi_nb_samples = get_pi_samples(s->cc[0].pi, s->cc[0].pi_start, s->cc[0].pi_end, 1);
+    for (int ch = 1; ch < inlink->channels; ch++) {
+        ChannelContext *cc = &s->cc[ch];
+        min_pi_nb_samples = FFMIN(min_pi_nb_samples, get_pi_samples(cc->pi, cc->pi_start, cc->pi_end, 1));
+    }
+
+    if (min_pi_nb_samples >= in->nb_samples || s->eof) {
+        int nb_samples = get_queued_samples(s);
+
+        in = ff_bufqueue_get(&s->queue);
+
+        av_frame_make_writable(in);
+
+        nb_samples -= in->nb_samples;
+
+        for (int ch = 0; ch < inlink->channels; ch++) {
+            ChannelContext *cc = &s->cc[ch];
+            double *src = (double *)in->extended_data[ch];
+            int start = cc->pi_start;
+            int offset = 0;
+            double gain = 1.;
+
+            for (int n = 0; n < in->nb_samples; n++) {
+                if (n >= offset) {
+                    int type = cc->pi[start].max_peak > s->threshold_value;
+
+                    if (type)
+                        gain = FFMIN(s->max_expansion, s->peak_value / cc->pi[start].max_peak);
+                    else
+                        gain = 1. / s->max_compression;
+
+                    av_assert0(cc->pi[start].size > 0);
+                    offset += cc->pi[start++].size;
+                    if (start >= MAX_ITEMS)
+                        start = 0;
+
+                    if (type)
+                        cc->gain_state = FFMIN(gain, cc->gain_state + s->raise_amount);
+                    else
+                        cc->gain_state = FFMAX(gain, cc->gain_state - s->fall_amount);
+                }
+                src[n] *= cc->gain_state;
+            }
+        }
+
+        for (int ch = 0; ch < inlink->channels; ch++) {
+            ChannelContext *cc = &s->cc[ch];
+
+            cc->pi_start = consume_pi(cc->pi, cc->pi_start, cc->pi_end, in->nb_samples);
+        }
+
+        for (int ch = 0; ch < inlink->channels; ch++) {
+            ChannelContext *cc = &s->cc[ch];
+            int pi_nb_samples = get_pi_samples(cc->pi, cc->pi_start, cc->pi_end, 0);
+
+            if (nb_samples != pi_nb_samples) {
+                av_assert0(0);
+            }
+        }
+
+        s->pts = in->pts + in->nb_samples;
+
+        return ff_filter_frame(outlink, in);
+    }
+
+    return 1;
+}
+
+static int activate(AVFilterContext *ctx)
+{
+    AVFilterLink *inlink = ctx->inputs[0];
+    AVFilterLink *outlink = ctx->outputs[0];
+    SpeechNormalizerContext *s = ctx->priv;
+    int ret = 0, status;
+    int64_t pts;
+
+    FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
+
+    ret = filter_frame(ctx);
+    if (ret <= 0)
+        return ret;
+
+    if (!s->eof && ff_inlink_acknowledge_status(inlink, &status, &pts)) {
+        if (status == AVERROR_EOF)
+            s->eof = 1;
+    }
+
+    if (s->eof && ff_inlink_queued_samples(inlink) == 0 &&
+        s->queue.available == 0) {
+        ff_outlink_set_status(outlink, AVERROR_EOF, s->pts);
+        return 0;
+    }
+
+    if (!s->eof)
+        FF_FILTER_FORWARD_WANTED(outlink, inlink);
+
+    if (s->eof && s->queue.available > 0) {
+        ff_filter_set_ready(ctx, 10);
+        return 0;
+    }
+
+    return FFERROR_NOT_READY;
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+
+    ff_bufqueue_discard_all(&s->queue);
+    av_freep(&s->cc);
+}
+
+static const AVFilterPad avfilter_af_speechnorm_inputs[] = {
+    {
+        .name           = "default",
+        .type           = AVMEDIA_TYPE_AUDIO,
+        .config_props   = config_input,
+    },
+    { NULL }
+};
+
+static const AVFilterPad avfilter_af_speechnorm_outputs[] = {
+    {
+        .name          = "default",
+        .type          = AVMEDIA_TYPE_AUDIO,
+    },
+    { NULL }
+};
+
+AVFilter ff_af_speechnorm = {
+    .name          = "speechnorm",
+    .description   = NULL_IF_CONFIG_SMALL("Speech Normalizer."),
+    .query_formats = query_formats,
+    .priv_size     = sizeof(SpeechNormalizerContext),
+    .priv_class    = &speechnorm_class,
+    .activate      = activate,
+    .uninit        = uninit,
+    .inputs        = avfilter_af_speechnorm_inputs,
+    .outputs       = avfilter_af_speechnorm_outputs,
+    .process_command = ff_filter_process_command,
+};
diff --git a/libavfilter/allfilters.c b/libavfilter/allfilters.c
index 7796959dc7..fde535d50c 100644
--- a/libavfilter/allfilters.c
+++ b/libavfilter/allfilters.c
@@ -132,6 +132,7 @@ extern AVFilter ff_af_sidechaingate;
 extern AVFilter ff_af_silencedetect;
 extern AVFilter ff_af_silenceremove;
 extern AVFilter ff_af_sofalizer;
+extern AVFilter ff_af_speechnorm;
 extern AVFilter ff_af_stereotools;
 extern AVFilter ff_af_stereowiden;
 extern AVFilter ff_af_superequalizer;
-- 
2.17.1



More information about the ffmpeg-devel mailing list