1 /*
2 * QCELP decoder
3 * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
4 *
5 * This file is part of FFmpeg.
6 *
7 * FFmpeg is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
11 *
12 * FFmpeg is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with FFmpeg; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 /**
23 * @file
24 * QCELP decoder
25 * @author Reynaldo H. Verdejo Pinochet
26 * @remark FFmpeg merging spearheaded by Kenan Gillet
27 * @remark Development mentored by Benjamin Larson
28 */
29
30 #include <stddef.h>
31
32 #include "libavutil/avassert.h"
33 #include "libavutil/channel_layout.h"
34 #include "libavutil/float_dsp.h"
35 #include "avcodec.h"
36 #include "codec_internal.h"
37 #include "internal.h"
38 #include "get_bits.h"
39 #include "qcelpdata.h"
40 #include "celp_filters.h"
41 #include "acelp_filters.h"
42 #include "acelp_vectors.h"
43 #include "lsp.h"
44
45 typedef enum {
46 I_F_Q = -1, /**< insufficient frame quality */
47 SILENCE,
48 RATE_OCTAVE,
49 RATE_QUARTER,
50 RATE_HALF,
51 RATE_FULL
52 } qcelp_packet_rate;
53
54 typedef struct QCELPContext {
55 GetBitContext gb;
56 qcelp_packet_rate bitrate;
57 QCELPFrame frame; /**< unpacked data frame */
58
59 uint8_t erasure_count;
60 uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
61 float prev_lspf[10];
62 float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
63 float pitch_synthesis_filter_mem[303];
64 float pitch_pre_filter_mem[303];
65 float rnd_fir_filter_mem[180];
66 float formant_mem[170];
67 float last_codebook_gain;
68 int prev_g1[2];
69 int prev_bitrate;
70 float pitch_gain[4];
71 uint8_t pitch_lag[4];
72 uint16_t first16bits;
73 uint8_t warned_buf_mismatch_bitrate;
74
75 /* postfilter */
76 float postfilter_synth_mem[10];
77 float postfilter_agc_mem;
78 float postfilter_tilt_mem;
79 } QCELPContext;
80
81 /**
82 * Initialize the speech codec according to the specification.
83 *
84 * TIA/EIA/IS-733 2.4.9
85 */
qcelp_decode_init(AVCodecContext *avctx)86 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
87 {
88 QCELPContext *q = avctx->priv_data;
89 int i;
90
91 av_channel_layout_uninit(&avctx->ch_layout);
92 avctx->ch_layout = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO;
93 avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
94
95 for (i = 0; i < 10; i++)
96 q->prev_lspf[i] = (i + 1) / 11.0;
97
98 return 0;
99 }
100
101 /**
102 * Decode the 10 quantized LSP frequencies from the LSPV/LSP
103 * transmission codes of any bitrate and check for badly received packets.
104 *
105 * @param q the context
106 * @param lspf line spectral pair frequencies
107 *
108 * @return 0 on success, -1 if the packet is badly received
109 *
110 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
111 */
decode_lspf(QCELPContext *q, float *lspf)112 static int decode_lspf(QCELPContext *q, float *lspf)
113 {
114 int i;
115 float tmp_lspf, smooth, erasure_coeff;
116 const float *predictors;
117
118 if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
119 predictors = q->prev_bitrate != RATE_OCTAVE &&
120 q->prev_bitrate != I_F_Q ? q->prev_lspf
121 : q->predictor_lspf;
122
123 if (q->bitrate == RATE_OCTAVE) {
124 q->octave_count++;
125
126 for (i = 0; i < 10; i++) {
127 q->predictor_lspf[i] =
128 lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
129 : -QCELP_LSP_SPREAD_FACTOR) +
130 predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR +
131 (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11);
132 }
133 smooth = q->octave_count < 10 ? .875 : 0.1;
134 } else {
135 erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
136
137 av_assert2(q->bitrate == I_F_Q);
138
139 if (q->erasure_count > 1)
140 erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7;
141
142 for (i = 0; i < 10; i++) {
143 q->predictor_lspf[i] =
144 lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 +
145 erasure_coeff * predictors[i];
146 }
147 smooth = 0.125;
148 }
149
150 // Check the stability of the LSP frequencies.
151 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
152 for (i = 1; i < 10; i++)
153 lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR);
154
155 lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR);
156 for (i = 9; i > 0; i--)
157 lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR);
158
159 // Low-pass filter the LSP frequencies.
160 ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10);
161 } else {
162 q->octave_count = 0;
163
164 tmp_lspf = 0.0;
165 for (i = 0; i < 5; i++) {
166 lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
167 lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
168 }
169
170 // Check for badly received packets.
171 if (q->bitrate == RATE_QUARTER) {
172 if (lspf[9] <= .70 || lspf[9] >= .97)
173 return -1;
174 for (i = 3; i < 10; i++)
175 if (fabs(lspf[i] - lspf[i - 2]) < .08)
176 return -1;
177 } else {
178 if (lspf[9] <= .66 || lspf[9] >= .985)
179 return -1;
180 for (i = 4; i < 10; i++)
181 if (fabs(lspf[i] - lspf[i - 4]) < .0931)
182 return -1;
183 }
184 }
185 return 0;
186 }
187
188 /**
189 * Convert codebook transmission codes to GAIN and INDEX.
190 *
191 * @param q the context
192 * @param gain array holding the decoded gain
193 *
194 * TIA/EIA/IS-733 2.4.6.2
195 */
decode_gain_and_index(QCELPContext *q, float *gain)196 static void decode_gain_and_index(QCELPContext *q, float *gain)
197 {
198 int i, subframes_count, g1[16];
199 float slope;
200
201 if (q->bitrate >= RATE_QUARTER) {
202 switch (q->bitrate) {
203 case RATE_FULL: subframes_count = 16; break;
204 case RATE_HALF: subframes_count = 4; break;
205 default: subframes_count = 5;
206 }
207 for (i = 0; i < subframes_count; i++) {
208 g1[i] = 4 * q->frame.cbgain[i];
209 if (q->bitrate == RATE_FULL && !((i + 1) & 3)) {
210 g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32);
211 }
212
213 gain[i] = qcelp_g12ga[g1[i]];
214
215 if (q->frame.cbsign[i]) {
216 gain[i] = -gain[i];
217 q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127;
218 }
219 }
220
221 q->prev_g1[0] = g1[i - 2];
222 q->prev_g1[1] = g1[i - 1];
223 q->last_codebook_gain = qcelp_g12ga[g1[i - 1]];
224
225 if (q->bitrate == RATE_QUARTER) {
226 // Provide smoothing of the unvoiced excitation energy.
227 gain[7] = gain[4];
228 gain[6] = 0.4 * gain[3] + 0.6 * gain[4];
229 gain[5] = gain[3];
230 gain[4] = 0.8 * gain[2] + 0.2 * gain[3];
231 gain[3] = 0.2 * gain[1] + 0.8 * gain[2];
232 gain[2] = gain[1];
233 gain[1] = 0.6 * gain[0] + 0.4 * gain[1];
234 }
235 } else if (q->bitrate != SILENCE) {
236 if (q->bitrate == RATE_OCTAVE) {
237 g1[0] = 2 * q->frame.cbgain[0] +
238 av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
239 subframes_count = 8;
240 } else {
241 av_assert2(q->bitrate == I_F_Q);
242
243 g1[0] = q->prev_g1[1];
244 switch (q->erasure_count) {
245 case 1 : break;
246 case 2 : g1[0] -= 1; break;
247 case 3 : g1[0] -= 2; break;
248 default: g1[0] -= 6;
249 }
250 if (g1[0] < 0)
251 g1[0] = 0;
252 subframes_count = 4;
253 }
254 // This interpolation is done to produce smoother background noise.
255 slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
256 for (i = 1; i <= subframes_count; i++)
257 gain[i - 1] = q->last_codebook_gain + slope * i;
258
259 q->last_codebook_gain = gain[i - 2];
260 q->prev_g1[0] = q->prev_g1[1];
261 q->prev_g1[1] = g1[0];
262 }
263 }
264
265 /**
266 * If the received packet is Rate 1/4 a further sanity check is made of the
267 * codebook gain.
268 *
269 * @param cbgain the unpacked cbgain array
270 * @return -1 if the sanity check fails, 0 otherwise
271 *
272 * TIA/EIA/IS-733 2.4.8.7.3
273 */
codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)274 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
275 {
276 int i, diff, prev_diff = 0;
277
278 for (i = 1; i < 5; i++) {
279 diff = cbgain[i] - cbgain[i-1];
280 if (FFABS(diff) > 10)
281 return -1;
282 else if (FFABS(diff - prev_diff) > 12)
283 return -1;
284 prev_diff = diff;
285 }
286 return 0;
287 }
288
289 /**
290 * Compute the scaled codebook vector Cdn From INDEX and GAIN
291 * for all rates.
292 *
293 * The specification lacks some information here.
294 *
295 * TIA/EIA/IS-733 has an omission on the codebook index determination
296 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
297 * you have to subtract the decoded index parameter from the given scaled
298 * codebook vector index 'n' to get the desired circular codebook index, but
299 * it does not mention that you have to clamp 'n' to [0-9] in order to get
300 * RI-compliant results.
301 *
302 * The reason for this mistake seems to be the fact they forgot to mention you
303 * have to do these calculations per codebook subframe and adjust given
304 * equation values accordingly.
305 *
306 * @param q the context
307 * @param gain array holding the 4 pitch subframe gain values
308 * @param cdn_vector array for the generated scaled codebook vector
309 */
compute_svector(QCELPContext *q, const float *gain, float *cdn_vector)310 static void compute_svector(QCELPContext *q, const float *gain,
311 float *cdn_vector)
312 {
313 int i, j, k;
314 uint16_t cbseed, cindex;
315 float *rnd, tmp_gain, fir_filter_value;
316
317 switch (q->bitrate) {
318 case RATE_FULL:
319 for (i = 0; i < 16; i++) {
320 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
321 cindex = -q->frame.cindex[i];
322 for (j = 0; j < 10; j++)
323 *cdn_vector++ = tmp_gain *
324 qcelp_rate_full_codebook[cindex++ & 127];
325 }
326 break;
327 case RATE_HALF:
328 for (i = 0; i < 4; i++) {
329 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
330 cindex = -q->frame.cindex[i];
331 for (j = 0; j < 40; j++)
332 *cdn_vector++ = tmp_gain *
333 qcelp_rate_half_codebook[cindex++ & 127];
334 }
335 break;
336 case RATE_QUARTER:
337 cbseed = (0x0003 & q->frame.lspv[4]) << 14 |
338 (0x003F & q->frame.lspv[3]) << 8 |
339 (0x0060 & q->frame.lspv[2]) << 1 |
340 (0x0007 & q->frame.lspv[1]) << 3 |
341 (0x0038 & q->frame.lspv[0]) >> 3;
342 rnd = q->rnd_fir_filter_mem + 20;
343 for (i = 0; i < 8; i++) {
344 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
345 for (k = 0; k < 20; k++) {
346 cbseed = 521 * cbseed + 259;
347 *rnd = (int16_t) cbseed;
348
349 // FIR filter
350 fir_filter_value = 0.0;
351 for (j = 0; j < 10; j++)
352 fir_filter_value += qcelp_rnd_fir_coefs[j] *
353 (rnd[-j] + rnd[-20+j]);
354
355 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
356 *cdn_vector++ = tmp_gain * fir_filter_value;
357 rnd++;
358 }
359 }
360 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160,
361 20 * sizeof(float));
362 break;
363 case RATE_OCTAVE:
364 cbseed = q->first16bits;
365 for (i = 0; i < 8; i++) {
366 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
367 for (j = 0; j < 20; j++) {
368 cbseed = 521 * cbseed + 259;
369 *cdn_vector++ = tmp_gain * (int16_t) cbseed;
370 }
371 }
372 break;
373 case I_F_Q:
374 cbseed = -44; // random codebook index
375 for (i = 0; i < 4; i++) {
376 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
377 for (j = 0; j < 40; j++)
378 *cdn_vector++ = tmp_gain *
379 qcelp_rate_full_codebook[cbseed++ & 127];
380 }
381 break;
382 case SILENCE:
383 memset(cdn_vector, 0, 160 * sizeof(float));
384 break;
385 }
386 }
387
388 /**
389 * Apply generic gain control.
390 *
391 * @param v_out output vector
392 * @param v_in gain-controlled vector
393 * @param v_ref vector to control gain of
394 *
395 * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
396 */
apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)397 static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)
398 {
399 int i;
400
401 for (i = 0; i < 160; i += 40) {
402 float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40);
403 ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40);
404 }
405 }
406
407 /**
408 * Apply filter in pitch-subframe steps.
409 *
410 * @param memory buffer for the previous state of the filter
411 * - must be able to contain 303 elements
412 * - the 143 first elements are from the previous state
413 * - the next 160 are for output
414 * @param v_in input filter vector
415 * @param gain per-subframe gain array, each element is between 0.0 and 2.0
416 * @param lag per-subframe lag array, each element is
417 * - between 16 and 143 if its corresponding pfrac is 0,
418 * - between 16 and 139 otherwise
419 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
420 * otherwise
421 *
422 * @return filter output vector
423 */
do_pitchfilter(float memory[303], const float v_in[160], const float gain[4], const uint8_t *lag, const uint8_t pfrac[4])424 static const float *do_pitchfilter(float memory[303], const float v_in[160],
425 const float gain[4], const uint8_t *lag,
426 const uint8_t pfrac[4])
427 {
428 int i, j;
429 float *v_lag, *v_out;
430 const float *v_len;
431
432 v_out = memory + 143; // Output vector starts at memory[143].
433
434 for (i = 0; i < 4; i++) {
435 if (gain[i]) {
436 v_lag = memory + 143 + 40 * i - lag[i];
437 for (v_len = v_in + 40; v_in < v_len; v_in++) {
438 if (pfrac[i]) { // If it is a fractional lag...
439 for (j = 0, *v_out = 0.0; j < 4; j++)
440 *v_out += qcelp_hammsinc_table[j] *
441 (v_lag[j - 4] + v_lag[3 - j]);
442 } else
443 *v_out = *v_lag;
444
445 *v_out = *v_in + gain[i] * *v_out;
446
447 v_lag++;
448 v_out++;
449 }
450 } else {
451 memcpy(v_out, v_in, 40 * sizeof(float));
452 v_in += 40;
453 v_out += 40;
454 }
455 }
456
457 memmove(memory, memory + 160, 143 * sizeof(float));
458 return memory + 143;
459 }
460
461 /**
462 * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
463 * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
464 *
465 * @param q the context
466 * @param cdn_vector the scaled codebook vector
467 */
apply_pitch_filters(QCELPContext *q, float *cdn_vector)468 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
469 {
470 int i;
471 const float *v_synthesis_filtered, *v_pre_filtered;
472
473 if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE ||
474 (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
475
476 if (q->bitrate >= RATE_HALF) {
477 // Compute gain & lag for the whole frame.
478 for (i = 0; i < 4; i++) {
479 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
480
481 q->pitch_lag[i] = q->frame.plag[i] + 16;
482 }
483 } else {
484 float max_pitch_gain;
485
486 if (q->bitrate == I_F_Q) {
487 if (q->erasure_count < 3)
488 max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
489 else
490 max_pitch_gain = 0.0;
491 } else {
492 av_assert2(q->bitrate == SILENCE);
493 max_pitch_gain = 1.0;
494 }
495 for (i = 0; i < 4; i++)
496 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
497
498 memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
499 }
500
501 // pitch synthesis filter
502 v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
503 cdn_vector, q->pitch_gain,
504 q->pitch_lag, q->frame.pfrac);
505
506 // pitch prefilter update
507 for (i = 0; i < 4; i++)
508 q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
509
510 v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
511 v_synthesis_filtered,
512 q->pitch_gain, q->pitch_lag,
513 q->frame.pfrac);
514
515 apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
516 } else {
517 memcpy(q->pitch_synthesis_filter_mem,
518 cdn_vector + 17, 143 * sizeof(float));
519 memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
520 memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
521 memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
522 }
523 }
524
525 /**
526 * Reconstruct LPC coefficients from the line spectral pair frequencies
527 * and perform bandwidth expansion.
528 *
529 * @param lspf line spectral pair frequencies
530 * @param lpc linear predictive coding coefficients
531 *
532 * @note: bandwidth_expansion_coeff could be precalculated into a table
533 * but it seems to be slower on x86
534 *
535 * TIA/EIA/IS-733 2.4.3.3.5
536 */
lspf2lpc(const float *lspf, float *lpc)537 static void lspf2lpc(const float *lspf, float *lpc)
538 {
539 double lsp[10];
540 double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
541 int i;
542
543 for (i = 0; i < 10; i++)
544 lsp[i] = cos(M_PI * lspf[i]);
545
546 ff_acelp_lspd2lpc(lsp, lpc, 5);
547
548 for (i = 0; i < 10; i++) {
549 lpc[i] *= bandwidth_expansion_coeff;
550 bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
551 }
552 }
553
554 /**
555 * Interpolate LSP frequencies and compute LPC coefficients
556 * for a given bitrate & pitch subframe.
557 *
558 * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
559 *
560 * @param q the context
561 * @param curr_lspf LSP frequencies vector of the current frame
562 * @param lpc float vector for the resulting LPC
563 * @param subframe_num frame number in decoded stream
564 */
interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc, const int subframe_num)565 static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
566 float *lpc, const int subframe_num)
567 {
568 float interpolated_lspf[10];
569 float weight;
570
571 if (q->bitrate >= RATE_QUARTER)
572 weight = 0.25 * (subframe_num + 1);
573 else if (q->bitrate == RATE_OCTAVE && !subframe_num)
574 weight = 0.625;
575 else
576 weight = 1.0;
577
578 if (weight != 1.0) {
579 ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
580 weight, 1.0 - weight, 10);
581 lspf2lpc(interpolated_lspf, lpc);
582 } else if (q->bitrate >= RATE_QUARTER ||
583 (q->bitrate == I_F_Q && !subframe_num))
584 lspf2lpc(curr_lspf, lpc);
585 else if (q->bitrate == SILENCE && !subframe_num)
586 lspf2lpc(q->prev_lspf, lpc);
587 }
588
buf_size2bitrate(const int buf_size)589 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
590 {
591 switch (buf_size) {
592 case 35: return RATE_FULL;
593 case 17: return RATE_HALF;
594 case 8: return RATE_QUARTER;
595 case 4: return RATE_OCTAVE;
596 case 1: return SILENCE;
597 }
598
599 return I_F_Q;
600 }
601
602 /**
603 * Determine the bitrate from the frame size and/or the first byte of the frame.
604 *
605 * @param avctx the AV codec context
606 * @param buf_size length of the buffer
607 * @param buf the buffer
608 *
609 * @return the bitrate on success,
610 * I_F_Q if the bitrate cannot be satisfactorily determined
611 *
612 * TIA/EIA/IS-733 2.4.8.7.1
613 */
determine_bitrate(AVCodecContext *avctx, const int buf_size, const uint8_t **buf)614 static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx,
615 const int buf_size,
616 const uint8_t **buf)
617 {
618 qcelp_packet_rate bitrate;
619
620 if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
621 if (bitrate > **buf) {
622 QCELPContext *q = avctx->priv_data;
623 if (!q->warned_buf_mismatch_bitrate) {
624 av_log(avctx, AV_LOG_WARNING,
625 "Claimed bitrate and buffer size mismatch.\n");
626 q->warned_buf_mismatch_bitrate = 1;
627 }
628 bitrate = **buf;
629 } else if (bitrate < **buf) {
630 av_log(avctx, AV_LOG_ERROR,
631 "Buffer is too small for the claimed bitrate.\n");
632 return I_F_Q;
633 }
634 (*buf)++;
635 } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
636 av_log(avctx, AV_LOG_WARNING,
637 "Bitrate byte missing, guessing bitrate from packet size.\n");
638 } else
639 return I_F_Q;
640
641 if (bitrate == SILENCE) {
642 // FIXME: Remove this warning when tested with samples.
643 avpriv_request_sample(avctx, "Blank frame handling");
644 }
645 return bitrate;
646 }
647
warn_insufficient_frame_quality(AVCodecContext *avctx, const char *message)648 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
649 const char *message)
650 {
651 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n",
652 avctx->frame_number, message);
653 }
654
postfilter(QCELPContext *q, float *samples, float *lpc)655 static void postfilter(QCELPContext *q, float *samples, float *lpc)
656 {
657 static const float pow_0_775[10] = {
658 0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
659 0.216676, 0.167924, 0.130141, 0.100859, 0.078166
660 }, pow_0_625[10] = {
661 0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
662 0.059605, 0.037253, 0.023283, 0.014552, 0.009095
663 };
664 float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
665 int n;
666
667 for (n = 0; n < 10; n++) {
668 lpc_s[n] = lpc[n] * pow_0_625[n];
669 lpc_p[n] = lpc[n] * pow_0_775[n];
670 }
671
672 ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
673 q->formant_mem + 10, 160, 10);
674 memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
675 ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
676 memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
677
678 ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
679
680 ff_adaptive_gain_control(samples, pole_out + 10,
681 avpriv_scalarproduct_float_c(q->formant_mem + 10,
682 q->formant_mem + 10,
683 160),
684 160, 0.9375, &q->postfilter_agc_mem);
685 }
686
qcelp_decode_frame(AVCodecContext *avctx, AVFrame *frame, int *got_frame_ptr, AVPacket *avpkt)687 static int qcelp_decode_frame(AVCodecContext *avctx, AVFrame *frame,
688 int *got_frame_ptr, AVPacket *avpkt)
689 {
690 const uint8_t *buf = avpkt->data;
691 int buf_size = avpkt->size;
692 QCELPContext *q = avctx->priv_data;
693 float *outbuffer;
694 int i, ret;
695 float quantized_lspf[10], lpc[10];
696 float gain[16];
697 float *formant_mem;
698
699 /* get output buffer */
700 frame->nb_samples = 160;
701 if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
702 return ret;
703 outbuffer = (float *)frame->data[0];
704
705 if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
706 warn_insufficient_frame_quality(avctx, "Bitrate cannot be determined.");
707 goto erasure;
708 }
709
710 if (q->bitrate == RATE_OCTAVE &&
711 (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
712 warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
713 goto erasure;
714 }
715
716 if (q->bitrate > SILENCE) {
717 const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
718 const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] +
719 qcelp_unpacking_bitmaps_lengths[q->bitrate];
720 uint8_t *unpacked_data = (uint8_t *)&q->frame;
721
722 if ((ret = init_get_bits8(&q->gb, buf, buf_size)) < 0)
723 return ret;
724
725 memset(&q->frame, 0, sizeof(QCELPFrame));
726
727 for (; bitmaps < bitmaps_end; bitmaps++)
728 unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
729
730 // Check for erasures/blanks on rates 1, 1/4 and 1/8.
731 if (q->frame.reserved) {
732 warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
733 goto erasure;
734 }
735 if (q->bitrate == RATE_QUARTER &&
736 codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
737 warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
738 goto erasure;
739 }
740
741 if (q->bitrate >= RATE_HALF) {
742 for (i = 0; i < 4; i++) {
743 if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
744 warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
745 goto erasure;
746 }
747 }
748 }
749 }
750
751 decode_gain_and_index(q, gain);
752 compute_svector(q, gain, outbuffer);
753
754 if (decode_lspf(q, quantized_lspf) < 0) {
755 warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
756 goto erasure;
757 }
758
759 apply_pitch_filters(q, outbuffer);
760
761 if (q->bitrate == I_F_Q) {
762 erasure:
763 q->bitrate = I_F_Q;
764 q->erasure_count++;
765 decode_gain_and_index(q, gain);
766 compute_svector(q, gain, outbuffer);
767 decode_lspf(q, quantized_lspf);
768 apply_pitch_filters(q, outbuffer);
769 } else
770 q->erasure_count = 0;
771
772 formant_mem = q->formant_mem + 10;
773 for (i = 0; i < 4; i++) {
774 interpolate_lpc(q, quantized_lspf, lpc, i);
775 ff_celp_lp_synthesis_filterf(formant_mem, lpc,
776 outbuffer + i * 40, 40, 10);
777 formant_mem += 40;
778 }
779
780 // postfilter, as per TIA/EIA/IS-733 2.4.8.6
781 postfilter(q, outbuffer, lpc);
782
783 memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
784
785 memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
786 q->prev_bitrate = q->bitrate;
787
788 *got_frame_ptr = 1;
789
790 return buf_size;
791 }
792
793 const FFCodec ff_qcelp_decoder = {
794 .p.name = "qcelp",
795 .p.long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
796 .p.type = AVMEDIA_TYPE_AUDIO,
797 .p.id = AV_CODEC_ID_QCELP,
798 .init = qcelp_decode_init,
799 FF_CODEC_DECODE_CB(qcelp_decode_frame),
800 .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
801 .priv_data_size = sizeof(QCELPContext),
802 .caps_internal = FF_CODEC_CAP_INIT_THREADSAFE,
803 };
804