1 /*
2 * Copyright (c) 2022 Paul B Mahol
3 *
4 * This file is part of FFmpeg.
5 *
6 * FFmpeg is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public License
8 * as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
10 *
11 * FFmpeg is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public License
17 * along with FFmpeg; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19 */
20
21 #include "libavutil/channel_layout.h"
22 #include "libavutil/opt.h"
23 #include "libavutil/tx.h"
24 #include "audio.h"
25 #include "avfilter.h"
26 #include "filters.h"
27 #include "internal.h"
28 #include "window_func.h"
29
30 #include <float.h>
31
32 typedef struct AudioDialogueEnhancementContext {
33 const AVClass *class;
34
35 double original, enhance, voice;
36
37 int fft_size;
38 int overlap;
39
40 float *window;
41 float prev_vad;
42
43 AVFrame *in;
44 AVFrame *in_frame;
45 AVFrame *out_dist_frame;
46 AVFrame *windowed_frame;
47 AVFrame *windowed_out;
48 AVFrame *windowed_prev;
49 AVFrame *center_frame;
50
51 AVTXContext *tx_ctx[2], *itx_ctx;
52 av_tx_fn tx_fn, itx_fn;
53 } AudioDialogueEnhanceContext;
54
55 #define OFFSET(x) offsetof(AudioDialogueEnhanceContext, x)
56 #define FLAGS AV_OPT_FLAG_AUDIO_PARAM | AV_OPT_FLAG_FILTERING_PARAM | AV_OPT_FLAG_RUNTIME_PARAM
57
58 static const AVOption dialoguenhance_options[] = {
59 { "original", "set original center factor", OFFSET(original), AV_OPT_TYPE_DOUBLE, {.dbl=1}, 0, 1, FLAGS },
60 { "enhance", "set dialogue enhance factor",OFFSET(enhance), AV_OPT_TYPE_DOUBLE, {.dbl=1}, 0, 3, FLAGS },
61 { "voice", "set voice detection factor", OFFSET(voice), AV_OPT_TYPE_DOUBLE, {.dbl=2}, 2,32, FLAGS },
62 {NULL}
63 };
64
65 AVFILTER_DEFINE_CLASS(dialoguenhance);
66
query_formats(AVFilterContext * ctx)67 static int query_formats(AVFilterContext *ctx)
68 {
69 AVFilterFormats *formats = NULL;
70 AVFilterChannelLayouts *in_layout = NULL, *out_layout = NULL;
71 int ret;
72
73 if ((ret = ff_add_format (&formats, AV_SAMPLE_FMT_FLTP )) < 0 ||
74 (ret = ff_set_common_formats (ctx , formats )) < 0 ||
75 (ret = ff_add_channel_layout (&in_layout , &(AVChannelLayout)AV_CHANNEL_LAYOUT_STEREO)) < 0 ||
76 (ret = ff_channel_layouts_ref(in_layout, &ctx->inputs[0]->outcfg.channel_layouts)) < 0 ||
77 (ret = ff_add_channel_layout (&out_layout , &(AVChannelLayout)AV_CHANNEL_LAYOUT_SURROUND)) < 0 ||
78 (ret = ff_channel_layouts_ref(out_layout, &ctx->outputs[0]->incfg.channel_layouts)) < 0)
79 return ret;
80
81 return ff_set_common_all_samplerates(ctx);
82 }
83
config_input(AVFilterLink * inlink)84 static int config_input(AVFilterLink *inlink)
85 {
86 AVFilterContext *ctx = inlink->dst;
87 AudioDialogueEnhanceContext *s = ctx->priv;
88 float scale = 1.f, iscale, overlap;
89 int ret;
90
91 s->fft_size = inlink->sample_rate > 100000 ? 8192 : inlink->sample_rate > 50000 ? 4096 : 2048;
92 s->overlap = s->fft_size / 4;
93
94 s->window = av_calloc(s->fft_size, sizeof(*s->window));
95 if (!s->window)
96 return AVERROR(ENOMEM);
97
98 s->in_frame = ff_get_audio_buffer(inlink, (s->fft_size + 2) * 2);
99 s->center_frame = ff_get_audio_buffer(inlink, (s->fft_size + 2) * 2);
100 s->out_dist_frame = ff_get_audio_buffer(inlink, (s->fft_size + 2) * 2);
101 s->windowed_frame = ff_get_audio_buffer(inlink, (s->fft_size + 2) * 2);
102 s->windowed_out = ff_get_audio_buffer(inlink, (s->fft_size + 2) * 2);
103 s->windowed_prev = ff_get_audio_buffer(inlink, (s->fft_size + 2) * 2);
104 if (!s->in_frame || !s->windowed_out || !s->windowed_prev ||
105 !s->out_dist_frame || !s->windowed_frame || !s->center_frame)
106 return AVERROR(ENOMEM);
107
108 generate_window_func(s->window, s->fft_size, WFUNC_SINE, &overlap);
109
110 iscale = 1.f / s->fft_size;
111
112 ret = av_tx_init(&s->tx_ctx[0], &s->tx_fn, AV_TX_FLOAT_RDFT, 0, s->fft_size, &scale, 0);
113 if (ret < 0)
114 return ret;
115
116 ret = av_tx_init(&s->tx_ctx[1], &s->tx_fn, AV_TX_FLOAT_RDFT, 0, s->fft_size, &scale, 0);
117 if (ret < 0)
118 return ret;
119
120 ret = av_tx_init(&s->itx_ctx, &s->itx_fn, AV_TX_FLOAT_RDFT, 1, s->fft_size, &iscale, 0);
121 if (ret < 0)
122 return ret;
123
124 return 0;
125 }
126
apply_window(AudioDialogueEnhanceContext * s,const float * in_frame,float * out_frame,const int add_to_out_frame)127 static void apply_window(AudioDialogueEnhanceContext *s,
128 const float *in_frame, float *out_frame, const int add_to_out_frame)
129 {
130 const float *window = s->window;
131
132 if (add_to_out_frame) {
133 for (int i = 0; i < s->fft_size; i++)
134 out_frame[i] += in_frame[i] * window[i];
135 } else {
136 for (int i = 0; i < s->fft_size; i++)
137 out_frame[i] = in_frame[i] * window[i];
138 }
139 }
140
sqrf(float x)141 static float sqrf(float x)
142 {
143 return x * x;
144 }
145
get_centere(AVComplexFloat * left,AVComplexFloat * right,AVComplexFloat * center,int N)146 static void get_centere(AVComplexFloat *left, AVComplexFloat *right,
147 AVComplexFloat *center, int N)
148 {
149 for (int i = 0; i < N; i++) {
150 const float l_re = left[i].re;
151 const float l_im = left[i].im;
152 const float r_re = right[i].re;
153 const float r_im = right[i].im;
154 const float a = 0.5f * (1.f - sqrtf((sqrf(l_re - r_re) + sqrf(l_im - r_im))/
155 (sqrf(l_re + r_re) + sqrf(l_im + r_im) + FLT_EPSILON)));
156
157 center[i].re = a * (l_re + r_re);
158 center[i].im = a * (l_im + r_im);
159 }
160 }
161
flux(float * curf,float * prevf,int N)162 static float flux(float *curf, float *prevf, int N)
163 {
164 AVComplexFloat *cur = (AVComplexFloat *)curf;
165 AVComplexFloat *prev = (AVComplexFloat *)prevf;
166 float sum = 0.f;
167
168 for (int i = 0; i < N; i++) {
169 float c_re = cur[i].re;
170 float c_im = cur[i].im;
171 float p_re = prev[i].re;
172 float p_im = prev[i].im;
173
174 sum += sqrf(hypotf(c_re, c_im) - hypotf(p_re, p_im));
175 }
176
177 return sum;
178 }
179
fluxlr(float * lf,float * lpf,float * rf,float * rpf,int N)180 static float fluxlr(float *lf, float *lpf,
181 float *rf, float *rpf,
182 int N)
183 {
184 AVComplexFloat *l = (AVComplexFloat *)lf;
185 AVComplexFloat *lp = (AVComplexFloat *)lpf;
186 AVComplexFloat *r = (AVComplexFloat *)rf;
187 AVComplexFloat *rp = (AVComplexFloat *)rpf;
188 float sum = 0.f;
189
190 for (int i = 0; i < N; i++) {
191 float c_re = l[i].re - r[i].re;
192 float c_im = l[i].im - r[i].im;
193 float p_re = lp[i].re - rp[i].re;
194 float p_im = lp[i].im - rp[i].im;
195
196 sum += sqrf(hypotf(c_re, c_im) - hypotf(p_re, p_im));
197 }
198
199 return sum;
200 }
201
calc_vad(float fc,float flr,float a)202 static float calc_vad(float fc, float flr, float a)
203 {
204 const float vad = a * (fc / (fc + flr) - 0.5f);
205
206 return av_clipf(vad, 0.f, 1.f);
207 }
208
get_final(float * c,float * l,float * r,float vad,int N,float original,float enhance)209 static void get_final(float *c, float *l,
210 float *r, float vad, int N,
211 float original, float enhance)
212 {
213 AVComplexFloat *center = (AVComplexFloat *)c;
214 AVComplexFloat *left = (AVComplexFloat *)l;
215 AVComplexFloat *right = (AVComplexFloat *)r;
216
217 for (int i = 0; i < N; i++) {
218 float cP = sqrf(center[i].re) + sqrf(center[i].im);
219 float lrP = sqrf(left[i].re - right[i].re) + sqrf(left[i].im - right[i].im);
220 float G = cP / (cP + lrP + FLT_EPSILON);
221 float re, im;
222
223 re = center[i].re * (original + vad * G * enhance);
224 im = center[i].im * (original + vad * G * enhance);
225
226 center[i].re = re;
227 center[i].im = im;
228 }
229 }
230
de_stereo(AVFilterContext * ctx,AVFrame * out)231 static int de_stereo(AVFilterContext *ctx, AVFrame *out)
232 {
233 AudioDialogueEnhanceContext *s = ctx->priv;
234 float *center = (float *)s->center_frame->extended_data[0];
235 float *center_prev = (float *)s->center_frame->extended_data[1];
236 float *left_in = (float *)s->in_frame->extended_data[0];
237 float *right_in = (float *)s->in_frame->extended_data[1];
238 float *left_out = (float *)s->out_dist_frame->extended_data[0];
239 float *right_out = (float *)s->out_dist_frame->extended_data[1];
240 float *left_samples = (float *)s->in->extended_data[0];
241 float *right_samples = (float *)s->in->extended_data[1];
242 float *windowed_left = (float *)s->windowed_frame->extended_data[0];
243 float *windowed_right = (float *)s->windowed_frame->extended_data[1];
244 float *windowed_oleft = (float *)s->windowed_out->extended_data[0];
245 float *windowed_oright = (float *)s->windowed_out->extended_data[1];
246 float *windowed_pleft = (float *)s->windowed_prev->extended_data[0];
247 float *windowed_pright = (float *)s->windowed_prev->extended_data[1];
248 float *left_osamples = (float *)out->extended_data[0];
249 float *right_osamples = (float *)out->extended_data[1];
250 float *center_osamples = (float *)out->extended_data[2];
251 const int offset = s->fft_size - s->overlap;
252 const int nb_samples = FFMIN(s->overlap, s->in->nb_samples);
253 float vad;
254
255 // shift in/out buffers
256 memmove(left_in, &left_in[s->overlap], offset * sizeof(float));
257 memmove(right_in, &right_in[s->overlap], offset * sizeof(float));
258 memmove(left_out, &left_out[s->overlap], offset * sizeof(float));
259 memmove(right_out, &right_out[s->overlap], offset * sizeof(float));
260
261 memcpy(&left_in[offset], left_samples, nb_samples * sizeof(float));
262 memcpy(&right_in[offset], right_samples, nb_samples * sizeof(float));
263 memset(&left_out[offset], 0, s->overlap * sizeof(float));
264 memset(&right_out[offset], 0, s->overlap * sizeof(float));
265
266 apply_window(s, left_in, windowed_left, 0);
267 apply_window(s, right_in, windowed_right, 0);
268
269 s->tx_fn(s->tx_ctx[0], windowed_oleft, windowed_left, sizeof(float));
270 s->tx_fn(s->tx_ctx[1], windowed_oright, windowed_right, sizeof(float));
271
272 get_centere((AVComplexFloat *)windowed_oleft,
273 (AVComplexFloat *)windowed_oright,
274 (AVComplexFloat *)center,
275 s->fft_size / 2 + 1);
276
277 vad = calc_vad(flux(center, center_prev, s->fft_size / 2 + 1),
278 fluxlr(windowed_oleft, windowed_pleft,
279 windowed_oright, windowed_pright, s->fft_size / 2 + 1), s->voice);
280 vad = vad * 0.1 + 0.9 * s->prev_vad;
281 s->prev_vad = vad;
282
283 memcpy(center_prev, center, s->fft_size * sizeof(float));
284 memcpy(windowed_pleft, windowed_oleft, s->fft_size * sizeof(float));
285 memcpy(windowed_pright, windowed_oright, s->fft_size * sizeof(float));
286
287 get_final(center, windowed_oleft, windowed_oright, vad, s->fft_size / 2 + 1,
288 s->original, s->enhance);
289
290 s->itx_fn(s->itx_ctx, windowed_oleft, center, sizeof(float));
291
292 apply_window(s, windowed_oleft, left_out, 1);
293
294 for (int i = 0; i < s->overlap; i++) {
295 // 4 times overlap with squared hanning window results in 1.5 time increase in amplitude
296 if (!ctx->is_disabled)
297 center_osamples[i] = left_out[i] / 1.5f;
298 else
299 center_osamples[i] = 0.f;
300 left_osamples[i] = left_in[i];
301 right_osamples[i] = right_in[i];
302 }
303
304 return 0;
305 }
306
filter_frame(AVFilterLink * inlink,AVFrame * in)307 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
308 {
309 AVFilterContext *ctx = inlink->dst;
310 AVFilterLink *outlink = ctx->outputs[0];
311 AudioDialogueEnhanceContext *s = ctx->priv;
312 AVFrame *out;
313 int ret;
314
315 out = ff_get_audio_buffer(outlink, s->overlap);
316 if (!out) {
317 ret = AVERROR(ENOMEM);
318 goto fail;
319 }
320
321 s->in = in;
322 de_stereo(ctx, out);
323
324 out->pts = in->pts;
325 out->nb_samples = in->nb_samples;
326 ret = ff_filter_frame(outlink, out);
327 fail:
328 av_frame_free(&in);
329 s->in = NULL;
330 return ret < 0 ? ret : 0;
331 }
332
activate(AVFilterContext * ctx)333 static int activate(AVFilterContext *ctx)
334 {
335 AVFilterLink *inlink = ctx->inputs[0];
336 AVFilterLink *outlink = ctx->outputs[0];
337 AudioDialogueEnhanceContext *s = ctx->priv;
338 AVFrame *in = NULL;
339 int ret = 0, status;
340 int64_t pts;
341
342 FF_FILTER_FORWARD_STATUS_BACK(outlink, inlink);
343
344 ret = ff_inlink_consume_samples(inlink, s->overlap, s->overlap, &in);
345 if (ret < 0)
346 return ret;
347
348 if (ret > 0) {
349 return filter_frame(inlink, in);
350 } else if (ff_inlink_acknowledge_status(inlink, &status, &pts)) {
351 ff_outlink_set_status(outlink, status, pts);
352 return 0;
353 } else {
354 if (ff_inlink_queued_samples(inlink) >= s->overlap) {
355 ff_filter_set_ready(ctx, 10);
356 } else if (ff_outlink_frame_wanted(outlink)) {
357 ff_inlink_request_frame(inlink);
358 }
359 return 0;
360 }
361 }
362
uninit(AVFilterContext * ctx)363 static av_cold void uninit(AVFilterContext *ctx)
364 {
365 AudioDialogueEnhanceContext *s = ctx->priv;
366
367 av_freep(&s->window);
368
369 av_frame_free(&s->in_frame);
370 av_frame_free(&s->center_frame);
371 av_frame_free(&s->out_dist_frame);
372 av_frame_free(&s->windowed_frame);
373 av_frame_free(&s->windowed_out);
374 av_frame_free(&s->windowed_prev);
375
376 av_tx_uninit(&s->tx_ctx[0]);
377 av_tx_uninit(&s->tx_ctx[1]);
378 av_tx_uninit(&s->itx_ctx);
379 }
380
381 static const AVFilterPad inputs[] = {
382 {
383 .name = "default",
384 .type = AVMEDIA_TYPE_AUDIO,
385 .config_props = config_input,
386 },
387 };
388
389 static const AVFilterPad outputs[] = {
390 {
391 .name = "default",
392 .type = AVMEDIA_TYPE_AUDIO,
393 },
394 };
395
396 const AVFilter ff_af_dialoguenhance = {
397 .name = "dialoguenhance",
398 .description = NULL_IF_CONFIG_SMALL("Audio Dialogue Enhancement."),
399 .priv_size = sizeof(AudioDialogueEnhanceContext),
400 .priv_class = &dialoguenhance_class,
401 .uninit = uninit,
402 FILTER_INPUTS(inputs),
403 FILTER_OUTPUTS(outputs),
404 FILTER_QUERY_FUNC(query_formats),
405 .flags = AVFILTER_FLAG_SUPPORT_TIMELINE_INTERNAL,
406 .activate = activate,
407 .process_command = ff_filter_process_command,
408 };
409