avfilter: add speechnorm filter
authorPaul B Mahol <onemda@gmail.com>
Sat, 2 May 2020 18:27:37 +0000 (20:27 +0200)
committerPaul B Mahol <onemda@gmail.com>
Sun, 15 Nov 2020 11:59:40 +0000 (12:59 +0100)
Changelog
doc/filters.texi
libavfilter/Makefile
libavfilter/af_speechnorm.c [new file with mode: 0644]
libavfilter/allfilters.c
libavfilter/version.h

index d8f151b..fcdabce 100644 (file)
--- a/Changelog
+++ b/Changelog
@@ -45,6 +45,7 @@ version <next>:
 - AMV muxer
 - NVDEC AV1 hwaccel
 - DXVA2/D3D11VA hardware accelerated AV1 decoding
+- speechnorm filter
 
 
 version 4.3:
index 8380f6c..8c9eccb 100644 (file)
@@ -5276,6 +5276,69 @@ and also with custom gain:
 @end example
 @end itemize
 
+@section speechnorm
+Speech Normalizer.
+
+This filter expands or compresses each half-cycle of audio samples
+(local set of samples all above or all below zero and between two nearest zero crossings) depending
+on threshold value, so audio reaches target peak value under conditions controlled by below options.
+
+The filter accepts the following options:
+
+@table @option
+@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.
+
+@item expansion, e
+Set the maximum expansion factor. Allowed range is from 1.0 to 50.0. 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 reaches target peak value but never to surpass it and that
+ratio between new and previous peak value does not surpass this option value.
+
+@item compression, c
+Set the maximum compression factor. Allowed range is from 1.0 to 50.0. 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, then 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 will be compressed by current compression factor.
+
+@item threshold, t
+Set the threshold value. Default value is 0.0. Allowed range is from 0.0 to 1.0.
+This option specifies which half-cycles of samples will be compressed and which will be expanded.
+Any half-cycle samples with their local peak value below or same as this option value will be
+compressed by current compression factor, otherwise, if greater than threshold value they will be
+expanded with expansion factor so that it could reach peak target value but never surpass it.
+
+@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.
+Setting this options too high may lead to distortions.
+
+@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.
+
+@item channels, h
+Specify which channels to filter, by default all available channels are filtered.
+
+@item invert, i
+Enable inverted filtering, by default is disabled. This inverts interpretation of @option{threshold}
+option. When enabled any half-cycle of samples with their local peak value below or same as
+@option{threshold} option will be expanded otherwise it will be compressed.
+
+@item link, l
+Link channels when calculating gain applied to each filtered channel sample, by default is disabled.
+When disabled each filtered channel gain calculation is independent, otherwise when this option
+is enabled the minimum of all possible gains for each filtered channel is used.
+@end table
+
+@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
index 30e39b7..794a55a 100644 (file)
@@ -137,6 +137,7 @@ 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 (file)
index 0000000..eb46cf2
--- /dev/null
@@ -0,0 +1,579 @@
+/*
+ * Copyright (c) 2020 Paul B Mahol
+ *
+ * Speech Normalizer
+ *
+ * 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
+#define MIN_PEAK (1. / 32768.)
+
+typedef struct PeriodItem {
+    int size;
+    int type;
+    double max_peak;
+} PeriodItem;
+
+typedef struct ChannelContext {
+    int state;
+    int bypass;
+    PeriodItem pi[MAX_ITEMS];
+    double gain_state;
+    double pi_max_peak;
+    int pi_start;
+    int pi_end;
+    int pi_size;
+} 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;
+    uint64_t channels;
+    int invert;
+    int link;
+
+    ChannelContext *cc;
+    double prev_gain;
+
+    int max_period;
+    int eof;
+    int64_t pts;
+
+    struct FFBufQueue queue;
+
+    void (*analyze_channel)(AVFilterContext *ctx, ChannelContext *cc,
+                            const uint8_t *srcp, int nb_samples);
+    void (*filter_channels[2])(AVFilterContext *ctx,
+                               AVFrame *in, int nb_samples);
+} 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, 50.0, FLAGS },
+    { "e",         "set the max expansion factor", OFFSET(max_expansion), AV_OPT_TYPE_DOUBLE, {.dbl=2.0}, 1.0, 50.0, FLAGS },
+    { "compression", "set the max compression factor", OFFSET(max_compression), AV_OPT_TYPE_DOUBLE, {.dbl=2.0}, 1.0, 50.0, FLAGS },
+    { "c",           "set the max compression factor", OFFSET(max_compression), AV_OPT_TYPE_DOUBLE, {.dbl=2.0}, 1.0, 50.0, 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 },
+    { "channels", "set channels to filter", OFFSET(channels), AV_OPT_TYPE_CHANNEL_LAYOUT, {.i64=-1}, INT64_MIN, INT64_MAX, FLAGS },
+    { "h",        "set channels to filter", OFFSET(channels), AV_OPT_TYPE_CHANNEL_LAYOUT, {.i64=-1}, INT64_MIN, INT64_MAX, FLAGS },
+    { "invert", "set inverted filtering", OFFSET(invert), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
+    { "i",      "set inverted filtering", OFFSET(invert), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
+    { "link", "set linked channels filtering", OFFSET(link), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
+    { "l",    "set linked channels filtering", OFFSET(link), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, 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_FLTP, 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 get_pi_samples(PeriodItem *pi, int start, int end, int remain)
+{
+    int sum;
+
+    if (pi[start].type == 0)
+        return remain;
+
+    sum = remain;
+    while (start != end) {
+        start++;
+        if (start >= MAX_ITEMS)
+            start = 0;
+        if (pi[start].type == 0)
+            break;
+        av_assert0(pi[start].size > 0);
+        sum += pi[start].size;
+    }
+
+    return sum;
+}
+
+static int available_samples(AVFilterContext *ctx)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+    AVFilterLink *inlink = ctx->inputs[0];
+    int min_pi_nb_samples;
+
+    min_pi_nb_samples = get_pi_samples(s->cc[0].pi, s->cc[0].pi_start, s->cc[0].pi_end, s->cc[0].pi_size);
+    for (int ch = 1; ch < inlink->channels && min_pi_nb_samples > 0; 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, cc->pi_size));
+    }
+
+    return min_pi_nb_samples;
+}
+
+static void consume_pi(ChannelContext *cc, int nb_samples)
+{
+    if (cc->pi_size >= nb_samples) {
+        cc->pi_size -= nb_samples;
+    } else {
+        av_assert0(0);
+    }
+}
+
+static double next_gain(AVFilterContext *ctx, double pi_max_peak, int bypass, double state)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+    const double expansion = FFMIN(s->max_expansion, s->peak_value / pi_max_peak);
+    const double compression = 1. / s->max_compression;
+    const int type = s->invert ? pi_max_peak <= s->threshold_value : pi_max_peak >= s->threshold_value;
+
+    if (bypass) {
+        return 1.;
+    } else if (type) {
+        return FFMIN(expansion, state + s->raise_amount);
+    } else {
+        return FFMIN(expansion, FFMAX(compression, state - s->fall_amount));
+    }
+}
+
+static void next_pi(AVFilterContext *ctx, ChannelContext *cc, int bypass)
+{
+    av_assert0(cc->pi_size >= 0);
+    if (cc->pi_size == 0) {
+        SpeechNormalizerContext *s = ctx->priv;
+        int start = cc->pi_start;
+
+        av_assert0(cc->pi[start].size > 0);
+        av_assert0(cc->pi[start].type > 0 || s->eof);
+        cc->pi_size = cc->pi[start].size;
+        cc->pi_max_peak = cc->pi[start].max_peak;
+        av_assert0(cc->pi_start != cc->pi_end || s->eof);
+        start++;
+        if (start >= MAX_ITEMS)
+            start = 0;
+        cc->pi_start = start;
+        cc->gain_state = next_gain(ctx, cc->pi_max_peak, bypass, cc->gain_state);
+    }
+}
+
+static double min_gain(AVFilterContext *ctx, ChannelContext *cc, int max_size)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+    double min_gain = s->max_expansion;
+    double gain_state = cc->gain_state;
+    int size = cc->pi_size;
+    int idx = cc->pi_start;
+
+    min_gain = FFMIN(min_gain, gain_state);
+    while (size <= max_size) {
+        if (idx == cc->pi_end)
+            break;
+        gain_state = next_gain(ctx, cc->pi[idx].max_peak, 0, gain_state);
+        min_gain = FFMIN(min_gain, gain_state);
+        size += cc->pi[idx].size;
+        idx++;
+        if (idx >= MAX_ITEMS)
+            idx = 0;
+    }
+
+    return min_gain;
+}
+
+#define ANALYZE_CHANNEL(name, ptype, zero)                                                 \
+static void analyze_channel_## name (AVFilterContext *ctx, ChannelContext *cc,             \
+                                     const uint8_t *srcp, int nb_samples)                  \
+{                                                                                          \
+    SpeechNormalizerContext *s = ctx->priv;                                                \
+    const ptype *src = (const ptype *)srcp;                                                \
+    int n = 0;                                                                             \
+                                                                                           \
+    if (cc->state < 0)                                                                     \
+        cc->state = src[0] >= zero;                                                        \
+                                                                                           \
+    while (n < nb_samples) {                                                               \
+        if ((cc->state != (src[n] >= zero)) ||                                             \
+            (cc->pi[cc->pi_end].size > s->max_period)) {                                   \
+            double max_peak = cc->pi[cc->pi_end].max_peak;                                 \
+            int state = cc->state;                                                         \
+            cc->state = src[n] >= zero;                                                    \
+            av_assert0(cc->pi[cc->pi_end].size > 0);                                       \
+            if (cc->pi[cc->pi_end].max_peak >= MIN_PEAK ||                                 \
+                cc->pi[cc->pi_end].size > s->max_period) {                                 \
+                cc->pi[cc->pi_end].type = 1;                                               \
+                cc->pi_end++;                                                              \
+                if (cc->pi_end >= MAX_ITEMS)                                               \
+                    cc->pi_end = 0;                                                        \
+                if (cc->state != state)                                                    \
+                    cc->pi[cc->pi_end].max_peak = DBL_MIN;                                 \
+                else                                                                       \
+                    cc->pi[cc->pi_end].max_peak = max_peak;                                \
+                cc->pi[cc->pi_end].type = 0;                                               \
+                cc->pi[cc->pi_end].size = 0;                                               \
+                av_assert0(cc->pi_end != cc->pi_start);                                    \
+            }                                                                              \
+        }                                                                                  \
+                                                                                           \
+        if (cc->state) {                                                                   \
+            while (src[n] >= zero) {                                                       \
+                cc->pi[cc->pi_end].max_peak = FFMAX(cc->pi[cc->pi_end].max_peak,  src[n]); \
+                cc->pi[cc->pi_end].size++;                                                 \
+                n++;                                                                       \
+                if (n >= nb_samples)                                                       \
+                    break;                                                                 \
+            }                                                                              \
+        } else {                                                                           \
+            while (src[n] < zero) {                                                        \
+                cc->pi[cc->pi_end].max_peak = FFMAX(cc->pi[cc->pi_end].max_peak, -src[n]); \
+                cc->pi[cc->pi_end].size++;                                                 \
+                n++;                                                                       \
+                if (n >= nb_samples)                                                       \
+                    break;                                                                 \
+            }                                                                              \
+        }                                                                                  \
+    }                                                                                      \
+}
+
+ANALYZE_CHANNEL(dbl, double, 0.0)
+ANALYZE_CHANNEL(flt, float,  0.f)
+
+#define FILTER_CHANNELS(name, ptype)                                            \
+static void filter_channels_## name (AVFilterContext *ctx,                      \
+                                     AVFrame *in, int nb_samples)               \
+{                                                                               \
+    SpeechNormalizerContext *s = ctx->priv;                                     \
+    AVFilterLink *inlink = ctx->inputs[0];                                      \
+                                                                                \
+    for (int ch = 0; ch < inlink->channels; ch++) {                             \
+        ChannelContext *cc = &s->cc[ch];                                        \
+        ptype *dst = (ptype *)in->extended_data[ch];                            \
+        const int bypass = !(av_channel_layout_extract_channel(inlink->channel_layout, ch) & s->channels); \
+        int n = 0;                                                              \
+                                                                                \
+        while (n < nb_samples) {                                                \
+            ptype gain;                                                         \
+            int size;                                                           \
+                                                                                \
+            next_pi(ctx, cc, bypass);                                           \
+            size = FFMIN(nb_samples - n, cc->pi_size);                          \
+            av_assert0(size > 0);                                               \
+            gain = cc->gain_state;                                              \
+            consume_pi(cc, size);                                               \
+            for (int i = n; i < n + size; i++)                                  \
+                dst[i] *= gain;                                                 \
+            n += size;                                                          \
+        }                                                                       \
+    }                                                                           \
+}
+
+FILTER_CHANNELS(dbl, double)
+FILTER_CHANNELS(flt, float)
+
+static double lerp(double min, double max, double mix)
+{
+    return min + (max - min) * mix;
+}
+
+#define FILTER_LINK_CHANNELS(name, ptype)                                       \
+static void filter_link_channels_## name (AVFilterContext *ctx,                 \
+                                          AVFrame *in, int nb_samples)          \
+{                                                                               \
+    SpeechNormalizerContext *s = ctx->priv;                                     \
+    AVFilterLink *inlink = ctx->inputs[0];                                      \
+    int n = 0;                                                                  \
+                                                                                \
+    while (n < nb_samples) {                                                    \
+        int min_size = nb_samples - n;                                          \
+        int max_size = 1;                                                       \
+        ptype gain = s->max_expansion;                                          \
+                                                                                \
+        for (int ch = 0; ch < inlink->channels; ch++) {                         \
+            ChannelContext *cc = &s->cc[ch];                                    \
+                                                                                \
+            cc->bypass = !(av_channel_layout_extract_channel(inlink->channel_layout, ch) & s->channels); \
+                                                                                \
+            next_pi(ctx, cc, cc->bypass);                                       \
+            min_size = FFMIN(min_size, cc->pi_size);                            \
+            max_size = FFMAX(max_size, cc->pi_size);                            \
+        }                                                                       \
+                                                                                \
+        av_assert0(min_size > 0);                                               \
+        for (int ch = 0; ch < inlink->channels; ch++) {                         \
+            ChannelContext *cc = &s->cc[ch];                                    \
+                                                                                \
+            if (cc->bypass)                                                     \
+                continue;                                                       \
+            gain = FFMIN(gain, min_gain(ctx, cc, max_size));                    \
+        }                                                                       \
+                                                                                \
+        for (int ch = 0; ch < inlink->channels; ch++) {                         \
+            ChannelContext *cc = &s->cc[ch];                                    \
+            ptype *dst = (ptype *)in->extended_data[ch];                        \
+                                                                                \
+            consume_pi(cc, min_size);                                           \
+            if (cc->bypass)                                                     \
+                continue;                                                       \
+                                                                                \
+            for (int i = n; i < n + min_size; i++) {                            \
+                ptype g = lerp(s->prev_gain, gain, (i - n) / (double)min_size); \
+                dst[i] *= g;                                                    \
+            }                                                                   \
+        }                                                                       \
+                                                                                \
+        s->prev_gain = gain;                                                    \
+        n += min_size;                                                          \
+    }                                                                           \
+}
+
+FILTER_LINK_CHANNELS(dbl, double)
+FILTER_LINK_CHANNELS(flt, float)
+
+static int filter_frame(AVFilterContext *ctx)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+    AVFilterLink *outlink = ctx->outputs[0];
+    AVFilterLink *inlink = ctx->inputs[0];
+    int ret;
+
+    while (s->queue.available > 0) {
+        int min_pi_nb_samples;
+        AVFrame *in;
+
+        in = ff_bufqueue_peek(&s->queue, 0);
+        if (!in)
+            break;
+
+        min_pi_nb_samples = available_samples(ctx);
+        if (min_pi_nb_samples < in->nb_samples && !s->eof)
+            break;
+
+        in = ff_bufqueue_get(&s->queue);
+
+        av_frame_make_writable(in);
+
+        s->filter_channels[s->link](ctx, in, in->nb_samples);
+
+        s->pts = in->pts + in->nb_samples;
+
+        return ff_filter_frame(outlink, in);
+    }
+
+    for (int f = 0; f < ff_inlink_queued_frames(inlink); f++) {
+        AVFrame *in;
+
+        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];
+
+            s->analyze_channel(ctx, cc, in->extended_data[ch], in->nb_samples);
+        }
+    }
+
+    return 1;
+}
+
+static int activate(AVFilterContext *ctx)
+{
+    AVFilterLink *inlink = ctx->inputs[0];
+    AVFilterLink *outlink = ctx->outputs[0];
+    SpeechNormalizerContext *s = ctx->priv;
+    int ret, 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->queue.available > 0) {
+        AVFrame *in = ff_bufqueue_peek(&s->queue, 0);
+        const int nb_samples = available_samples(ctx);
+
+        if (nb_samples >= in->nb_samples || s->eof) {
+            ff_filter_set_ready(ctx, 10);
+            return 0;
+        }
+    }
+
+    FF_FILTER_FORWARD_WANTED(outlink, inlink);
+
+    return FFERROR_NOT_READY;
+}
+
+static int config_input(AVFilterLink *inlink)
+{
+    AVFilterContext *ctx = inlink->dst;
+    SpeechNormalizerContext *s = ctx->priv;
+
+    s->max_period = inlink->sample_rate / 10;
+
+    s->prev_gain = 1.;
+    s->cc = av_calloc(inlink->channels, sizeof(*s->cc));
+    if (!s->cc)
+        return AVERROR(ENOMEM);
+
+    for (int ch = 0; ch < inlink->channels; ch++) {
+        ChannelContext *cc = &s->cc[ch];
+
+        cc->state = -1;
+        cc->gain_state = 1.;
+    }
+
+    switch (inlink->format) {
+    case AV_SAMPLE_FMT_FLTP:
+        s->analyze_channel = analyze_channel_flt;
+        s->filter_channels[0] = filter_channels_flt;
+        s->filter_channels[1] = filter_link_channels_flt;
+        break;
+    case AV_SAMPLE_FMT_DBLP:
+        s->analyze_channel = analyze_channel_dbl;
+        s->filter_channels[0] = filter_channels_dbl;
+        s->filter_channels[1] = filter_link_channels_dbl;
+        break;
+    default:
+        av_assert0(0);
+    }
+
+    return 0;
+}
+
+static int process_command(AVFilterContext *ctx, const char *cmd, const char *args,
+                           char *res, int res_len, int flags)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+    int link = s->link;
+    int ret;
+
+    ret = ff_filter_process_command(ctx, cmd, args, res, res_len, flags);
+    if (ret < 0)
+        return ret;
+    if (link != s->link)
+        s->prev_gain = 1.;
+
+    return 0;
+}
+
+static av_cold void uninit(AVFilterContext *ctx)
+{
+    SpeechNormalizerContext *s = ctx->priv;
+
+    ff_bufqueue_discard_all(&s->queue);
+    av_freep(&s->cc);
+}
+
+static const AVFilterPad inputs[] = {
+    {
+        .name         = "default",
+        .type         = AVMEDIA_TYPE_AUDIO,
+        .config_props = config_input,
+    },
+    { NULL }
+};
+
+static const AVFilterPad 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          = inputs,
+    .outputs         = outputs,
+    .process_command = process_command,
+};
index 4c671be..fbfd898 100644 (file)
@@ -131,6 +131,7 @@ 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;
index 7112ec8..23c9d37 100644 (file)
@@ -30,7 +30,7 @@
 #include "libavutil/version.h"
 
 #define LIBAVFILTER_VERSION_MAJOR   7
-#define LIBAVFILTER_VERSION_MINOR  89
+#define LIBAVFILTER_VERSION_MINOR  90
 #define LIBAVFILTER_VERSION_MICRO 100