blob: b4afda2fb1c26161b1297fefbd9b00d0ea66c31a
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 "internal.h" |
37 | #include "get_bits.h" |
38 | #include "qcelpdata.h" |
39 | #include "celp_filters.h" |
40 | #include "acelp_filters.h" |
41 | #include "acelp_vectors.h" |
42 | #include "lsp.h" |
43 | |
44 | typedef enum { |
45 | I_F_Q = -1, /**< insufficient frame quality */ |
46 | SILENCE, |
47 | RATE_OCTAVE, |
48 | RATE_QUARTER, |
49 | RATE_HALF, |
50 | RATE_FULL |
51 | } qcelp_packet_rate; |
52 | |
53 | typedef struct QCELPContext { |
54 | GetBitContext gb; |
55 | qcelp_packet_rate bitrate; |
56 | QCELPFrame frame; /**< unpacked data frame */ |
57 | |
58 | uint8_t erasure_count; |
59 | uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */ |
60 | float prev_lspf[10]; |
61 | float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */ |
62 | float pitch_synthesis_filter_mem[303]; |
63 | float pitch_pre_filter_mem[303]; |
64 | float rnd_fir_filter_mem[180]; |
65 | float formant_mem[170]; |
66 | float last_codebook_gain; |
67 | int prev_g1[2]; |
68 | int prev_bitrate; |
69 | float pitch_gain[4]; |
70 | uint8_t pitch_lag[4]; |
71 | uint16_t first16bits; |
72 | uint8_t warned_buf_mismatch_bitrate; |
73 | |
74 | /* postfilter */ |
75 | float postfilter_synth_mem[10]; |
76 | float postfilter_agc_mem; |
77 | float postfilter_tilt_mem; |
78 | } QCELPContext; |
79 | |
80 | /** |
81 | * Initialize the speech codec according to the specification. |
82 | * |
83 | * TIA/EIA/IS-733 2.4.9 |
84 | */ |
85 | static av_cold int qcelp_decode_init(AVCodecContext *avctx) |
86 | { |
87 | QCELPContext *q = avctx->priv_data; |
88 | int i; |
89 | |
90 | avctx->channels = 1; |
91 | avctx->channel_layout = AV_CH_LAYOUT_MONO; |
92 | avctx->sample_fmt = AV_SAMPLE_FMT_FLT; |
93 | |
94 | for (i = 0; i < 10; i++) |
95 | q->prev_lspf[i] = (i + 1) / 11.0; |
96 | |
97 | return 0; |
98 | } |
99 | |
100 | /** |
101 | * Decode the 10 quantized LSP frequencies from the LSPV/LSP |
102 | * transmission codes of any bitrate and check for badly received packets. |
103 | * |
104 | * @param q the context |
105 | * @param lspf line spectral pair frequencies |
106 | * |
107 | * @return 0 on success, -1 if the packet is badly received |
108 | * |
109 | * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3 |
110 | */ |
111 | static int decode_lspf(QCELPContext *q, float *lspf) |
112 | { |
113 | int i; |
114 | float tmp_lspf, smooth, erasure_coeff; |
115 | const float *predictors; |
116 | |
117 | if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) { |
118 | predictors = q->prev_bitrate != RATE_OCTAVE && |
119 | q->prev_bitrate != I_F_Q ? q->prev_lspf |
120 | : q->predictor_lspf; |
121 | |
122 | if (q->bitrate == RATE_OCTAVE) { |
123 | q->octave_count++; |
124 | |
125 | for (i = 0; i < 10; i++) { |
126 | q->predictor_lspf[i] = |
127 | lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR |
128 | : -QCELP_LSP_SPREAD_FACTOR) + |
129 | predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR + |
130 | (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11); |
131 | } |
132 | smooth = q->octave_count < 10 ? .875 : 0.1; |
133 | } else { |
134 | erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR; |
135 | |
136 | av_assert2(q->bitrate == I_F_Q); |
137 | |
138 | if (q->erasure_count > 1) |
139 | erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7; |
140 | |
141 | for (i = 0; i < 10; i++) { |
142 | q->predictor_lspf[i] = |
143 | lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 + |
144 | erasure_coeff * predictors[i]; |
145 | } |
146 | smooth = 0.125; |
147 | } |
148 | |
149 | // Check the stability of the LSP frequencies. |
150 | lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR); |
151 | for (i = 1; i < 10; i++) |
152 | lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR); |
153 | |
154 | lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR); |
155 | for (i = 9; i > 0; i--) |
156 | lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR); |
157 | |
158 | // Low-pass filter the LSP frequencies. |
159 | ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10); |
160 | } else { |
161 | q->octave_count = 0; |
162 | |
163 | tmp_lspf = 0.0; |
164 | for (i = 0; i < 5; i++) { |
165 | lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001; |
166 | lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001; |
167 | } |
168 | |
169 | // Check for badly received packets. |
170 | if (q->bitrate == RATE_QUARTER) { |
171 | if (lspf[9] <= .70 || lspf[9] >= .97) |
172 | return -1; |
173 | for (i = 3; i < 10; i++) |
174 | if (fabs(lspf[i] - lspf[i - 2]) < .08) |
175 | return -1; |
176 | } else { |
177 | if (lspf[9] <= .66 || lspf[9] >= .985) |
178 | return -1; |
179 | for (i = 4; i < 10; i++) |
180 | if (fabs(lspf[i] - lspf[i - 4]) < .0931) |
181 | return -1; |
182 | } |
183 | } |
184 | return 0; |
185 | } |
186 | |
187 | /** |
188 | * Convert codebook transmission codes to GAIN and INDEX. |
189 | * |
190 | * @param q the context |
191 | * @param gain array holding the decoded gain |
192 | * |
193 | * TIA/EIA/IS-733 2.4.6.2 |
194 | */ |
195 | static void decode_gain_and_index(QCELPContext *q, float *gain) |
196 | { |
197 | int i, subframes_count, g1[16]; |
198 | float slope; |
199 | |
200 | if (q->bitrate >= RATE_QUARTER) { |
201 | switch (q->bitrate) { |
202 | case RATE_FULL: subframes_count = 16; break; |
203 | case RATE_HALF: subframes_count = 4; break; |
204 | default: subframes_count = 5; |
205 | } |
206 | for (i = 0; i < subframes_count; i++) { |
207 | g1[i] = 4 * q->frame.cbgain[i]; |
208 | if (q->bitrate == RATE_FULL && !((i + 1) & 3)) { |
209 | g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32); |
210 | } |
211 | |
212 | gain[i] = qcelp_g12ga[g1[i]]; |
213 | |
214 | if (q->frame.cbsign[i]) { |
215 | gain[i] = -gain[i]; |
216 | q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127; |
217 | } |
218 | } |
219 | |
220 | q->prev_g1[0] = g1[i - 2]; |
221 | q->prev_g1[1] = g1[i - 1]; |
222 | q->last_codebook_gain = qcelp_g12ga[g1[i - 1]]; |
223 | |
224 | if (q->bitrate == RATE_QUARTER) { |
225 | // Provide smoothing of the unvoiced excitation energy. |
226 | gain[7] = gain[4]; |
227 | gain[6] = 0.4 * gain[3] + 0.6 * gain[4]; |
228 | gain[5] = gain[3]; |
229 | gain[4] = 0.8 * gain[2] + 0.2 * gain[3]; |
230 | gain[3] = 0.2 * gain[1] + 0.8 * gain[2]; |
231 | gain[2] = gain[1]; |
232 | gain[1] = 0.6 * gain[0] + 0.4 * gain[1]; |
233 | } |
234 | } else if (q->bitrate != SILENCE) { |
235 | if (q->bitrate == RATE_OCTAVE) { |
236 | g1[0] = 2 * q->frame.cbgain[0] + |
237 | av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54); |
238 | subframes_count = 8; |
239 | } else { |
240 | av_assert2(q->bitrate == I_F_Q); |
241 | |
242 | g1[0] = q->prev_g1[1]; |
243 | switch (q->erasure_count) { |
244 | case 1 : break; |
245 | case 2 : g1[0] -= 1; break; |
246 | case 3 : g1[0] -= 2; break; |
247 | default: g1[0] -= 6; |
248 | } |
249 | if (g1[0] < 0) |
250 | g1[0] = 0; |
251 | subframes_count = 4; |
252 | } |
253 | // This interpolation is done to produce smoother background noise. |
254 | slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count; |
255 | for (i = 1; i <= subframes_count; i++) |
256 | gain[i - 1] = q->last_codebook_gain + slope * i; |
257 | |
258 | q->last_codebook_gain = gain[i - 2]; |
259 | q->prev_g1[0] = q->prev_g1[1]; |
260 | q->prev_g1[1] = g1[0]; |
261 | } |
262 | } |
263 | |
264 | /** |
265 | * If the received packet is Rate 1/4 a further sanity check is made of the |
266 | * codebook gain. |
267 | * |
268 | * @param cbgain the unpacked cbgain array |
269 | * @return -1 if the sanity check fails, 0 otherwise |
270 | * |
271 | * TIA/EIA/IS-733 2.4.8.7.3 |
272 | */ |
273 | static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain) |
274 | { |
275 | int i, diff, prev_diff = 0; |
276 | |
277 | for (i = 1; i < 5; i++) { |
278 | diff = cbgain[i] - cbgain[i-1]; |
279 | if (FFABS(diff) > 10) |
280 | return -1; |
281 | else if (FFABS(diff - prev_diff) > 12) |
282 | return -1; |
283 | prev_diff = diff; |
284 | } |
285 | return 0; |
286 | } |
287 | |
288 | /** |
289 | * Compute the scaled codebook vector Cdn From INDEX and GAIN |
290 | * for all rates. |
291 | * |
292 | * The specification lacks some information here. |
293 | * |
294 | * TIA/EIA/IS-733 has an omission on the codebook index determination |
295 | * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says |
296 | * you have to subtract the decoded index parameter from the given scaled |
297 | * codebook vector index 'n' to get the desired circular codebook index, but |
298 | * it does not mention that you have to clamp 'n' to [0-9] in order to get |
299 | * RI-compliant results. |
300 | * |
301 | * The reason for this mistake seems to be the fact they forgot to mention you |
302 | * have to do these calculations per codebook subframe and adjust given |
303 | * equation values accordingly. |
304 | * |
305 | * @param q the context |
306 | * @param gain array holding the 4 pitch subframe gain values |
307 | * @param cdn_vector array for the generated scaled codebook vector |
308 | */ |
309 | static void compute_svector(QCELPContext *q, const float *gain, |
310 | float *cdn_vector) |
311 | { |
312 | int i, j, k; |
313 | uint16_t cbseed, cindex; |
314 | float *rnd, tmp_gain, fir_filter_value; |
315 | |
316 | switch (q->bitrate) { |
317 | case RATE_FULL: |
318 | for (i = 0; i < 16; i++) { |
319 | tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; |
320 | cindex = -q->frame.cindex[i]; |
321 | for (j = 0; j < 10; j++) |
322 | *cdn_vector++ = tmp_gain * |
323 | qcelp_rate_full_codebook[cindex++ & 127]; |
324 | } |
325 | break; |
326 | case RATE_HALF: |
327 | for (i = 0; i < 4; i++) { |
328 | tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO; |
329 | cindex = -q->frame.cindex[i]; |
330 | for (j = 0; j < 40; j++) |
331 | *cdn_vector++ = tmp_gain * |
332 | qcelp_rate_half_codebook[cindex++ & 127]; |
333 | } |
334 | break; |
335 | case RATE_QUARTER: |
336 | cbseed = (0x0003 & q->frame.lspv[4]) << 14 | |
337 | (0x003F & q->frame.lspv[3]) << 8 | |
338 | (0x0060 & q->frame.lspv[2]) << 1 | |
339 | (0x0007 & q->frame.lspv[1]) << 3 | |
340 | (0x0038 & q->frame.lspv[0]) >> 3; |
341 | rnd = q->rnd_fir_filter_mem + 20; |
342 | for (i = 0; i < 8; i++) { |
343 | tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); |
344 | for (k = 0; k < 20; k++) { |
345 | cbseed = 521 * cbseed + 259; |
346 | *rnd = (int16_t) cbseed; |
347 | |
348 | // FIR filter |
349 | fir_filter_value = 0.0; |
350 | for (j = 0; j < 10; j++) |
351 | fir_filter_value += qcelp_rnd_fir_coefs[j] * |
352 | (rnd[-j] + rnd[-20+j]); |
353 | |
354 | fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10]; |
355 | *cdn_vector++ = tmp_gain * fir_filter_value; |
356 | rnd++; |
357 | } |
358 | } |
359 | memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, |
360 | 20 * sizeof(float)); |
361 | break; |
362 | case RATE_OCTAVE: |
363 | cbseed = q->first16bits; |
364 | for (i = 0; i < 8; i++) { |
365 | tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); |
366 | for (j = 0; j < 20; j++) { |
367 | cbseed = 521 * cbseed + 259; |
368 | *cdn_vector++ = tmp_gain * (int16_t) cbseed; |
369 | } |
370 | } |
371 | break; |
372 | case I_F_Q: |
373 | cbseed = -44; // random codebook index |
374 | for (i = 0; i < 4; i++) { |
375 | tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; |
376 | for (j = 0; j < 40; j++) |
377 | *cdn_vector++ = tmp_gain * |
378 | qcelp_rate_full_codebook[cbseed++ & 127]; |
379 | } |
380 | break; |
381 | case SILENCE: |
382 | memset(cdn_vector, 0, 160 * sizeof(float)); |
383 | break; |
384 | } |
385 | } |
386 | |
387 | /** |
388 | * Apply generic gain control. |
389 | * |
390 | * @param v_out output vector |
391 | * @param v_in gain-controlled vector |
392 | * @param v_ref vector to control gain of |
393 | * |
394 | * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6 |
395 | */ |
396 | static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in) |
397 | { |
398 | int i; |
399 | |
400 | for (i = 0; i < 160; i += 40) { |
401 | float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40); |
402 | ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40); |
403 | } |
404 | } |
405 | |
406 | /** |
407 | * Apply filter in pitch-subframe steps. |
408 | * |
409 | * @param memory buffer for the previous state of the filter |
410 | * - must be able to contain 303 elements |
411 | * - the 143 first elements are from the previous state |
412 | * - the next 160 are for output |
413 | * @param v_in input filter vector |
414 | * @param gain per-subframe gain array, each element is between 0.0 and 2.0 |
415 | * @param lag per-subframe lag array, each element is |
416 | * - between 16 and 143 if its corresponding pfrac is 0, |
417 | * - between 16 and 139 otherwise |
418 | * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0 |
419 | * otherwise |
420 | * |
421 | * @return filter output vector |
422 | */ |
423 | static const float *do_pitchfilter(float memory[303], const float v_in[160], |
424 | const float gain[4], const uint8_t *lag, |
425 | const uint8_t pfrac[4]) |
426 | { |
427 | int i, j; |
428 | float *v_lag, *v_out; |
429 | const float *v_len; |
430 | |
431 | v_out = memory + 143; // Output vector starts at memory[143]. |
432 | |
433 | for (i = 0; i < 4; i++) { |
434 | if (gain[i]) { |
435 | v_lag = memory + 143 + 40 * i - lag[i]; |
436 | for (v_len = v_in + 40; v_in < v_len; v_in++) { |
437 | if (pfrac[i]) { // If it is a fractional lag... |
438 | for (j = 0, *v_out = 0.0; j < 4; j++) |
439 | *v_out += qcelp_hammsinc_table[j] * |
440 | (v_lag[j - 4] + v_lag[3 - j]); |
441 | } else |
442 | *v_out = *v_lag; |
443 | |
444 | *v_out = *v_in + gain[i] * *v_out; |
445 | |
446 | v_lag++; |
447 | v_out++; |
448 | } |
449 | } else { |
450 | memcpy(v_out, v_in, 40 * sizeof(float)); |
451 | v_in += 40; |
452 | v_out += 40; |
453 | } |
454 | } |
455 | |
456 | memmove(memory, memory + 160, 143 * sizeof(float)); |
457 | return memory + 143; |
458 | } |
459 | |
460 | /** |
461 | * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector. |
462 | * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2 |
463 | * |
464 | * @param q the context |
465 | * @param cdn_vector the scaled codebook vector |
466 | */ |
467 | static void apply_pitch_filters(QCELPContext *q, float *cdn_vector) |
468 | { |
469 | int i; |
470 | const float *v_synthesis_filtered, *v_pre_filtered; |
471 | |
472 | if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE || |
473 | (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) { |
474 | |
475 | if (q->bitrate >= RATE_HALF) { |
476 | // Compute gain & lag for the whole frame. |
477 | for (i = 0; i < 4; i++) { |
478 | q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0; |
479 | |
480 | q->pitch_lag[i] = q->frame.plag[i] + 16; |
481 | } |
482 | } else { |
483 | float max_pitch_gain; |
484 | |
485 | if (q->bitrate == I_F_Q) { |
486 | if (q->erasure_count < 3) |
487 | max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1); |
488 | else |
489 | max_pitch_gain = 0.0; |
490 | } else { |
491 | av_assert2(q->bitrate == SILENCE); |
492 | max_pitch_gain = 1.0; |
493 | } |
494 | for (i = 0; i < 4; i++) |
495 | q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain); |
496 | |
497 | memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac)); |
498 | } |
499 | |
500 | // pitch synthesis filter |
501 | v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem, |
502 | cdn_vector, q->pitch_gain, |
503 | q->pitch_lag, q->frame.pfrac); |
504 | |
505 | // pitch prefilter update |
506 | for (i = 0; i < 4; i++) |
507 | q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0); |
508 | |
509 | v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem, |
510 | v_synthesis_filtered, |
511 | q->pitch_gain, q->pitch_lag, |
512 | q->frame.pfrac); |
513 | |
514 | apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered); |
515 | } else { |
516 | memcpy(q->pitch_synthesis_filter_mem, |
517 | cdn_vector + 17, 143 * sizeof(float)); |
518 | memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float)); |
519 | memset(q->pitch_gain, 0, sizeof(q->pitch_gain)); |
520 | memset(q->pitch_lag, 0, sizeof(q->pitch_lag)); |
521 | } |
522 | } |
523 | |
524 | /** |
525 | * Reconstruct LPC coefficients from the line spectral pair frequencies |
526 | * and perform bandwidth expansion. |
527 | * |
528 | * @param lspf line spectral pair frequencies |
529 | * @param lpc linear predictive coding coefficients |
530 | * |
531 | * @note: bandwidth_expansion_coeff could be precalculated into a table |
532 | * but it seems to be slower on x86 |
533 | * |
534 | * TIA/EIA/IS-733 2.4.3.3.5 |
535 | */ |
536 | static void lspf2lpc(const float *lspf, float *lpc) |
537 | { |
538 | double lsp[10]; |
539 | double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF; |
540 | int i; |
541 | |
542 | for (i = 0; i < 10; i++) |
543 | lsp[i] = cos(M_PI * lspf[i]); |
544 | |
545 | ff_acelp_lspd2lpc(lsp, lpc, 5); |
546 | |
547 | for (i = 0; i < 10; i++) { |
548 | lpc[i] *= bandwidth_expansion_coeff; |
549 | bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF; |
550 | } |
551 | } |
552 | |
553 | /** |
554 | * Interpolate LSP frequencies and compute LPC coefficients |
555 | * for a given bitrate & pitch subframe. |
556 | * |
557 | * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2 |
558 | * |
559 | * @param q the context |
560 | * @param curr_lspf LSP frequencies vector of the current frame |
561 | * @param lpc float vector for the resulting LPC |
562 | * @param subframe_num frame number in decoded stream |
563 | */ |
564 | static void interpolate_lpc(QCELPContext *q, const float *curr_lspf, |
565 | float *lpc, const int subframe_num) |
566 | { |
567 | float interpolated_lspf[10]; |
568 | float weight; |
569 | |
570 | if (q->bitrate >= RATE_QUARTER) |
571 | weight = 0.25 * (subframe_num + 1); |
572 | else if (q->bitrate == RATE_OCTAVE && !subframe_num) |
573 | weight = 0.625; |
574 | else |
575 | weight = 1.0; |
576 | |
577 | if (weight != 1.0) { |
578 | ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf, |
579 | weight, 1.0 - weight, 10); |
580 | lspf2lpc(interpolated_lspf, lpc); |
581 | } else if (q->bitrate >= RATE_QUARTER || |
582 | (q->bitrate == I_F_Q && !subframe_num)) |
583 | lspf2lpc(curr_lspf, lpc); |
584 | else if (q->bitrate == SILENCE && !subframe_num) |
585 | lspf2lpc(q->prev_lspf, lpc); |
586 | } |
587 | |
588 | static qcelp_packet_rate buf_size2bitrate(const int buf_size) |
589 | { |
590 | switch (buf_size) { |
591 | case 35: return RATE_FULL; |
592 | case 17: return RATE_HALF; |
593 | case 8: return RATE_QUARTER; |
594 | case 4: return RATE_OCTAVE; |
595 | case 1: return SILENCE; |
596 | } |
597 | |
598 | return I_F_Q; |
599 | } |
600 | |
601 | /** |
602 | * Determine the bitrate from the frame size and/or the first byte of the frame. |
603 | * |
604 | * @param avctx the AV codec context |
605 | * @param buf_size length of the buffer |
606 | * @param buf the buffer |
607 | * |
608 | * @return the bitrate on success, |
609 | * I_F_Q if the bitrate cannot be satisfactorily determined |
610 | * |
611 | * TIA/EIA/IS-733 2.4.8.7.1 |
612 | */ |
613 | static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, |
614 | const int buf_size, |
615 | const uint8_t **buf) |
616 | { |
617 | qcelp_packet_rate bitrate; |
618 | |
619 | if ((bitrate = buf_size2bitrate(buf_size)) >= 0) { |
620 | if (bitrate > **buf) { |
621 | QCELPContext *q = avctx->priv_data; |
622 | if (!q->warned_buf_mismatch_bitrate) { |
623 | av_log(avctx, AV_LOG_WARNING, |
624 | "Claimed bitrate and buffer size mismatch.\n"); |
625 | q->warned_buf_mismatch_bitrate = 1; |
626 | } |
627 | bitrate = **buf; |
628 | } else if (bitrate < **buf) { |
629 | av_log(avctx, AV_LOG_ERROR, |
630 | "Buffer is too small for the claimed bitrate.\n"); |
631 | return I_F_Q; |
632 | } |
633 | (*buf)++; |
634 | } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) { |
635 | av_log(avctx, AV_LOG_WARNING, |
636 | "Bitrate byte missing, guessing bitrate from packet size.\n"); |
637 | } else |
638 | return I_F_Q; |
639 | |
640 | if (bitrate == SILENCE) { |
641 | // FIXME: Remove this warning when tested with samples. |
642 | avpriv_request_sample(avctx, "Blank frame handling"); |
643 | } |
644 | return bitrate; |
645 | } |
646 | |
647 | static void warn_insufficient_frame_quality(AVCodecContext *avctx, |
648 | const char *message) |
649 | { |
650 | av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", |
651 | avctx->frame_number, message); |
652 | } |
653 | |
654 | static void postfilter(QCELPContext *q, float *samples, float *lpc) |
655 | { |
656 | static const float pow_0_775[10] = { |
657 | 0.775000, 0.600625, 0.465484, 0.360750, 0.279582, |
658 | 0.216676, 0.167924, 0.130141, 0.100859, 0.078166 |
659 | }, pow_0_625[10] = { |
660 | 0.625000, 0.390625, 0.244141, 0.152588, 0.095367, |
661 | 0.059605, 0.037253, 0.023283, 0.014552, 0.009095 |
662 | }; |
663 | float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160]; |
664 | int n; |
665 | |
666 | for (n = 0; n < 10; n++) { |
667 | lpc_s[n] = lpc[n] * pow_0_625[n]; |
668 | lpc_p[n] = lpc[n] * pow_0_775[n]; |
669 | } |
670 | |
671 | ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s, |
672 | q->formant_mem + 10, 160, 10); |
673 | memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10); |
674 | ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10); |
675 | memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10); |
676 | |
677 | ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160); |
678 | |
679 | ff_adaptive_gain_control(samples, pole_out + 10, |
680 | avpriv_scalarproduct_float_c(q->formant_mem + 10, |
681 | q->formant_mem + 10, |
682 | 160), |
683 | 160, 0.9375, &q->postfilter_agc_mem); |
684 | } |
685 | |
686 | static int qcelp_decode_frame(AVCodecContext *avctx, void *data, |
687 | int *got_frame_ptr, AVPacket *avpkt) |
688 | { |
689 | const uint8_t *buf = avpkt->data; |
690 | int buf_size = avpkt->size; |
691 | QCELPContext *q = avctx->priv_data; |
692 | AVFrame *frame = 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 | AVCodec ff_qcelp_decoder = { |
794 | .name = "qcelp", |
795 | .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"), |
796 | .type = AVMEDIA_TYPE_AUDIO, |
797 | .id = AV_CODEC_ID_QCELP, |
798 | .init = qcelp_decode_init, |
799 | .decode = qcelp_decode_frame, |
800 | .capabilities = AV_CODEC_CAP_DR1, |
801 | .priv_data_size = sizeof(QCELPContext), |
802 | }; |
803 |