You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

796 lines
26KB

  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. * @file
  23. * QCELP decoder
  24. * @author Reynaldo H. Verdejo Pinochet
  25. * @remark FFmpeg merging spearheaded by Kenan Gillet
  26. * @remark Development mentored by Benjamin Larson
  27. */
  28. #include <stddef.h>
  29. #include "libavutil/avassert.h"
  30. #include "libavutil/channel_layout.h"
  31. #include "libavutil/float_dsp.h"
  32. #include "avcodec.h"
  33. #include "internal.h"
  34. #include "get_bits.h"
  35. #include "qcelpdata.h"
  36. #include "celp_filters.h"
  37. #include "acelp_filters.h"
  38. #include "acelp_vectors.h"
  39. #include "lsp.h"
  40. typedef enum {
  41. I_F_Q = -1, /**< insufficient frame quality */
  42. SILENCE,
  43. RATE_OCTAVE,
  44. RATE_QUARTER,
  45. RATE_HALF,
  46. RATE_FULL
  47. } qcelp_packet_rate;
  48. typedef struct {
  49. GetBitContext gb;
  50. qcelp_packet_rate bitrate;
  51. QCELPFrame frame; /**< unpacked data frame */
  52. uint8_t erasure_count;
  53. uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
  54. float prev_lspf[10];
  55. float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
  56. float pitch_synthesis_filter_mem[303];
  57. float pitch_pre_filter_mem[303];
  58. float rnd_fir_filter_mem[180];
  59. float formant_mem[170];
  60. float last_codebook_gain;
  61. int prev_g1[2];
  62. int prev_bitrate;
  63. float pitch_gain[4];
  64. uint8_t pitch_lag[4];
  65. uint16_t first16bits;
  66. uint8_t warned_buf_mismatch_bitrate;
  67. /* postfilter */
  68. float postfilter_synth_mem[10];
  69. float postfilter_agc_mem;
  70. float postfilter_tilt_mem;
  71. } QCELPContext;
  72. /**
  73. * Initialize the speech codec according to the specification.
  74. *
  75. * TIA/EIA/IS-733 2.4.9
  76. */
  77. static av_cold int qcelp_decode_init(AVCodecContext *avctx)
  78. {
  79. QCELPContext *q = avctx->priv_data;
  80. int i;
  81. avctx->channels = 1;
  82. avctx->channel_layout = AV_CH_LAYOUT_MONO;
  83. avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
  84. for (i = 0; i < 10; i++)
  85. q->prev_lspf[i] = (i + 1) / 11.0;
  86. return 0;
  87. }
  88. /**
  89. * Decode the 10 quantized LSP frequencies from the LSPV/LSP
  90. * transmission codes of any bitrate and check for badly received packets.
  91. *
  92. * @param q the context
  93. * @param lspf line spectral pair frequencies
  94. *
  95. * @return 0 on success, -1 if the packet is badly received
  96. *
  97. * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
  98. */
  99. static int decode_lspf(QCELPContext *q, float *lspf)
  100. {
  101. int i;
  102. float tmp_lspf, smooth, erasure_coeff;
  103. const float *predictors;
  104. if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
  105. predictors = q->prev_bitrate != RATE_OCTAVE &&
  106. q->prev_bitrate != I_F_Q ? q->prev_lspf
  107. : q->predictor_lspf;
  108. if (q->bitrate == RATE_OCTAVE) {
  109. q->octave_count++;
  110. for (i = 0; i < 10; i++) {
  111. q->predictor_lspf[i] =
  112. lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
  113. : -QCELP_LSP_SPREAD_FACTOR) +
  114. predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR +
  115. (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11);
  116. }
  117. smooth = q->octave_count < 10 ? .875 : 0.1;
  118. } else {
  119. erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
  120. av_assert2(q->bitrate == I_F_Q);
  121. if (q->erasure_count > 1)
  122. erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7;
  123. for (i = 0; i < 10; i++) {
  124. q->predictor_lspf[i] =
  125. lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 +
  126. erasure_coeff * predictors[i];
  127. }
  128. smooth = 0.125;
  129. }
  130. // Check the stability of the LSP frequencies.
  131. lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
  132. for (i = 1; i < 10; i++)
  133. lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR);
  134. lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR);
  135. for (i = 9; i > 0; i--)
  136. lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR);
  137. // Low-pass filter the LSP frequencies.
  138. ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10);
  139. } else {
  140. q->octave_count = 0;
  141. tmp_lspf = 0.0;
  142. for (i = 0; i < 5; i++) {
  143. lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
  144. lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
  145. }
  146. // Check for badly received packets.
  147. if (q->bitrate == RATE_QUARTER) {
  148. if (lspf[9] <= .70 || lspf[9] >= .97)
  149. return -1;
  150. for (i = 3; i < 10; i++)
  151. if (fabs(lspf[i] - lspf[i - 2]) < .08)
  152. return -1;
  153. } else {
  154. if (lspf[9] <= .66 || lspf[9] >= .985)
  155. return -1;
  156. for (i = 4; i < 10; i++)
  157. if (fabs(lspf[i] - lspf[i - 4]) < .0931)
  158. return -1;
  159. }
  160. }
  161. return 0;
  162. }
  163. /**
  164. * Convert codebook transmission codes to GAIN and INDEX.
  165. *
  166. * @param q the context
  167. * @param gain array holding the decoded gain
  168. *
  169. * TIA/EIA/IS-733 2.4.6.2
  170. */
  171. static void decode_gain_and_index(QCELPContext *q, float *gain)
  172. {
  173. int i, subframes_count, g1[16];
  174. float slope;
  175. if (q->bitrate >= RATE_QUARTER) {
  176. switch (q->bitrate) {
  177. case RATE_FULL: subframes_count = 16; break;
  178. case RATE_HALF: subframes_count = 4; break;
  179. default: subframes_count = 5;
  180. }
  181. for (i = 0; i < subframes_count; i++) {
  182. g1[i] = 4 * q->frame.cbgain[i];
  183. if (q->bitrate == RATE_FULL && !((i + 1) & 3)) {
  184. g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32);
  185. }
  186. gain[i] = qcelp_g12ga[g1[i]];
  187. if (q->frame.cbsign[i]) {
  188. gain[i] = -gain[i];
  189. q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127;
  190. }
  191. }
  192. q->prev_g1[0] = g1[i - 2];
  193. q->prev_g1[1] = g1[i - 1];
  194. q->last_codebook_gain = qcelp_g12ga[g1[i - 1]];
  195. if (q->bitrate == RATE_QUARTER) {
  196. // Provide smoothing of the unvoiced excitation energy.
  197. gain[7] = gain[4];
  198. gain[6] = 0.4 * gain[3] + 0.6 * gain[4];
  199. gain[5] = gain[3];
  200. gain[4] = 0.8 * gain[2] + 0.2 * gain[3];
  201. gain[3] = 0.2 * gain[1] + 0.8 * gain[2];
  202. gain[2] = gain[1];
  203. gain[1] = 0.6 * gain[0] + 0.4 * gain[1];
  204. }
  205. } else if (q->bitrate != SILENCE) {
  206. if (q->bitrate == RATE_OCTAVE) {
  207. g1[0] = 2 * q->frame.cbgain[0] +
  208. av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
  209. subframes_count = 8;
  210. } else {
  211. av_assert2(q->bitrate == I_F_Q);
  212. g1[0] = q->prev_g1[1];
  213. switch (q->erasure_count) {
  214. case 1 : break;
  215. case 2 : g1[0] -= 1; break;
  216. case 3 : g1[0] -= 2; break;
  217. default: g1[0] -= 6;
  218. }
  219. if (g1[0] < 0)
  220. g1[0] = 0;
  221. subframes_count = 4;
  222. }
  223. // This interpolation is done to produce smoother background noise.
  224. slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
  225. for (i = 1; i <= subframes_count; i++)
  226. gain[i - 1] = q->last_codebook_gain + slope * i;
  227. q->last_codebook_gain = gain[i - 2];
  228. q->prev_g1[0] = q->prev_g1[1];
  229. q->prev_g1[1] = g1[0];
  230. }
  231. }
  232. /**
  233. * If the received packet is Rate 1/4 a further sanity check is made of the
  234. * codebook gain.
  235. *
  236. * @param cbgain the unpacked cbgain array
  237. * @return -1 if the sanity check fails, 0 otherwise
  238. *
  239. * TIA/EIA/IS-733 2.4.8.7.3
  240. */
  241. static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
  242. {
  243. int i, diff, prev_diff = 0;
  244. for (i = 1; i < 5; i++) {
  245. diff = cbgain[i] - cbgain[i-1];
  246. if (FFABS(diff) > 10)
  247. return -1;
  248. else if (FFABS(diff - prev_diff) > 12)
  249. return -1;
  250. prev_diff = diff;
  251. }
  252. return 0;
  253. }
  254. /**
  255. * Compute the scaled codebook vector Cdn From INDEX and GAIN
  256. * for all rates.
  257. *
  258. * The specification lacks some information here.
  259. *
  260. * TIA/EIA/IS-733 has an omission on the codebook index determination
  261. * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
  262. * you have to subtract the decoded index parameter from the given scaled
  263. * codebook vector index 'n' to get the desired circular codebook index, but
  264. * it does not mention that you have to clamp 'n' to [0-9] in order to get
  265. * RI-compliant results.
  266. *
  267. * The reason for this mistake seems to be the fact they forgot to mention you
  268. * have to do these calculations per codebook subframe and adjust given
  269. * equation values accordingly.
  270. *
  271. * @param q the context
  272. * @param gain array holding the 4 pitch subframe gain values
  273. * @param cdn_vector array for the generated scaled codebook vector
  274. */
  275. static void compute_svector(QCELPContext *q, const float *gain,
  276. float *cdn_vector)
  277. {
  278. int i, j, k;
  279. uint16_t cbseed, cindex;
  280. float *rnd, tmp_gain, fir_filter_value;
  281. switch (q->bitrate) {
  282. case RATE_FULL:
  283. for (i = 0; i < 16; i++) {
  284. tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
  285. cindex = -q->frame.cindex[i];
  286. for (j = 0; j < 10; j++)
  287. *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
  288. }
  289. break;
  290. case RATE_HALF:
  291. for (i = 0; i < 4; i++) {
  292. tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
  293. cindex = -q->frame.cindex[i];
  294. for (j = 0; j < 40; j++)
  295. *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
  296. }
  297. break;
  298. case RATE_QUARTER:
  299. cbseed = (0x0003 & q->frame.lspv[4]) << 14 |
  300. (0x003F & q->frame.lspv[3]) << 8 |
  301. (0x0060 & q->frame.lspv[2]) << 1 |
  302. (0x0007 & q->frame.lspv[1]) << 3 |
  303. (0x0038 & q->frame.lspv[0]) >> 3;
  304. rnd = q->rnd_fir_filter_mem + 20;
  305. for (i = 0; i < 8; i++) {
  306. tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
  307. for (k = 0; k < 20; k++) {
  308. cbseed = 521 * cbseed + 259;
  309. *rnd = (int16_t) cbseed;
  310. // FIR filter
  311. fir_filter_value = 0.0;
  312. for (j = 0; j < 10; j++)
  313. fir_filter_value += qcelp_rnd_fir_coefs[j] *
  314. (rnd[-j] + rnd[-20+j]);
  315. fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
  316. *cdn_vector++ = tmp_gain * fir_filter_value;
  317. rnd++;
  318. }
  319. }
  320. memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160,
  321. 20 * sizeof(float));
  322. break;
  323. case RATE_OCTAVE:
  324. cbseed = q->first16bits;
  325. for (i = 0; i < 8; i++) {
  326. tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
  327. for (j = 0; j < 20; j++) {
  328. cbseed = 521 * cbseed + 259;
  329. *cdn_vector++ = tmp_gain * (int16_t) cbseed;
  330. }
  331. }
  332. break;
  333. case I_F_Q:
  334. cbseed = -44; // random codebook index
  335. for (i = 0; i < 4; i++) {
  336. tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
  337. for (j = 0; j < 40; j++)
  338. *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
  339. }
  340. break;
  341. case SILENCE:
  342. memset(cdn_vector, 0, 160 * sizeof(float));
  343. break;
  344. }
  345. }
  346. /**
  347. * Apply generic gain control.
  348. *
  349. * @param v_out output vector
  350. * @param v_in gain-controlled vector
  351. * @param v_ref vector to control gain of
  352. *
  353. * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
  354. */
  355. static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)
  356. {
  357. int i;
  358. for (i = 0; i < 160; i += 40) {
  359. float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40);
  360. ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40);
  361. }
  362. }
  363. /**
  364. * Apply filter in pitch-subframe steps.
  365. *
  366. * @param memory buffer for the previous state of the filter
  367. * - must be able to contain 303 elements
  368. * - the 143 first elements are from the previous state
  369. * - the next 160 are for output
  370. * @param v_in input filter vector
  371. * @param gain per-subframe gain array, each element is between 0.0 and 2.0
  372. * @param lag per-subframe lag array, each element is
  373. * - between 16 and 143 if its corresponding pfrac is 0,
  374. * - between 16 and 139 otherwise
  375. * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
  376. * otherwise
  377. *
  378. * @return filter output vector
  379. */
  380. static const float *do_pitchfilter(float memory[303], const float v_in[160],
  381. const float gain[4], const uint8_t *lag,
  382. const uint8_t pfrac[4])
  383. {
  384. int i, j;
  385. float *v_lag, *v_out;
  386. const float *v_len;
  387. v_out = memory + 143; // Output vector starts at memory[143].
  388. for (i = 0; i < 4; i++) {
  389. if (gain[i]) {
  390. v_lag = memory + 143 + 40 * i - lag[i];
  391. for (v_len = v_in + 40; v_in < v_len; v_in++) {
  392. if (pfrac[i]) { // If it is a fractional lag...
  393. for (j = 0, *v_out = 0.0; j < 4; j++)
  394. *v_out += qcelp_hammsinc_table[j] * (v_lag[j - 4] + v_lag[3 - j]);
  395. } else
  396. *v_out = *v_lag;
  397. *v_out = *v_in + gain[i] * *v_out;
  398. v_lag++;
  399. v_out++;
  400. }
  401. } else {
  402. memcpy(v_out, v_in, 40 * sizeof(float));
  403. v_in += 40;
  404. v_out += 40;
  405. }
  406. }
  407. memmove(memory, memory + 160, 143 * sizeof(float));
  408. return memory + 143;
  409. }
  410. /**
  411. * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
  412. * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
  413. *
  414. * @param q the context
  415. * @param cdn_vector the scaled codebook vector
  416. */
  417. static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
  418. {
  419. int i;
  420. const float *v_synthesis_filtered, *v_pre_filtered;
  421. if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE ||
  422. (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
  423. if (q->bitrate >= RATE_HALF) {
  424. // Compute gain & lag for the whole frame.
  425. for (i = 0; i < 4; i++) {
  426. q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
  427. q->pitch_lag[i] = q->frame.plag[i] + 16;
  428. }
  429. } else {
  430. float max_pitch_gain;
  431. if (q->bitrate == I_F_Q) {
  432. if (q->erasure_count < 3)
  433. max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
  434. else
  435. max_pitch_gain = 0.0;
  436. } else {
  437. av_assert2(q->bitrate == SILENCE);
  438. max_pitch_gain = 1.0;
  439. }
  440. for (i = 0; i < 4; i++)
  441. q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
  442. memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
  443. }
  444. // pitch synthesis filter
  445. v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
  446. cdn_vector, q->pitch_gain,
  447. q->pitch_lag, q->frame.pfrac);
  448. // pitch prefilter update
  449. for (i = 0; i < 4; i++)
  450. q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
  451. v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
  452. v_synthesis_filtered,
  453. q->pitch_gain, q->pitch_lag,
  454. q->frame.pfrac);
  455. apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
  456. } else {
  457. memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17, 143 * sizeof(float));
  458. memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
  459. memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
  460. memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
  461. }
  462. }
  463. /**
  464. * Reconstruct LPC coefficients from the line spectral pair frequencies
  465. * and perform bandwidth expansion.
  466. *
  467. * @param lspf line spectral pair frequencies
  468. * @param lpc linear predictive coding coefficients
  469. *
  470. * @note: bandwidth_expansion_coeff could be precalculated into a table
  471. * but it seems to be slower on x86
  472. *
  473. * TIA/EIA/IS-733 2.4.3.3.5
  474. */
  475. static void lspf2lpc(const float *lspf, float *lpc)
  476. {
  477. double lsp[10];
  478. double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
  479. int i;
  480. for (i = 0; i < 10; i++)
  481. lsp[i] = cos(M_PI * lspf[i]);
  482. ff_acelp_lspd2lpc(lsp, lpc, 5);
  483. for (i = 0; i < 10; i++) {
  484. lpc[i] *= bandwidth_expansion_coeff;
  485. bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
  486. }
  487. }
  488. /**
  489. * Interpolate LSP frequencies and compute LPC coefficients
  490. * for a given bitrate & pitch subframe.
  491. *
  492. * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
  493. *
  494. * @param q the context
  495. * @param curr_lspf LSP frequencies vector of the current frame
  496. * @param lpc float vector for the resulting LPC
  497. * @param subframe_num frame number in decoded stream
  498. */
  499. static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
  500. float *lpc, const int subframe_num)
  501. {
  502. float interpolated_lspf[10];
  503. float weight;
  504. if (q->bitrate >= RATE_QUARTER)
  505. weight = 0.25 * (subframe_num + 1);
  506. else if (q->bitrate == RATE_OCTAVE && !subframe_num)
  507. weight = 0.625;
  508. else
  509. weight = 1.0;
  510. if (weight != 1.0) {
  511. ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
  512. weight, 1.0 - weight, 10);
  513. lspf2lpc(interpolated_lspf, lpc);
  514. } else if (q->bitrate >= RATE_QUARTER ||
  515. (q->bitrate == I_F_Q && !subframe_num))
  516. lspf2lpc(curr_lspf, lpc);
  517. else if (q->bitrate == SILENCE && !subframe_num)
  518. lspf2lpc(q->prev_lspf, lpc);
  519. }
  520. static qcelp_packet_rate buf_size2bitrate(const int buf_size)
  521. {
  522. switch (buf_size) {
  523. case 35: return RATE_FULL;
  524. case 17: return RATE_HALF;
  525. case 8: return RATE_QUARTER;
  526. case 4: return RATE_OCTAVE;
  527. case 1: return SILENCE;
  528. }
  529. return I_F_Q;
  530. }
  531. /**
  532. * Determine the bitrate from the frame size and/or the first byte of the frame.
  533. *
  534. * @param avctx the AV codec context
  535. * @param buf_size length of the buffer
  536. * @param buf the bufffer
  537. *
  538. * @return the bitrate on success,
  539. * I_F_Q if the bitrate cannot be satisfactorily determined
  540. *
  541. * TIA/EIA/IS-733 2.4.8.7.1
  542. */
  543. static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx,
  544. const int buf_size,
  545. const uint8_t **buf)
  546. {
  547. qcelp_packet_rate bitrate;
  548. if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
  549. if (bitrate > **buf) {
  550. QCELPContext *q = avctx->priv_data;
  551. if (!q->warned_buf_mismatch_bitrate) {
  552. av_log(avctx, AV_LOG_WARNING,
  553. "Claimed bitrate and buffer size mismatch.\n");
  554. q->warned_buf_mismatch_bitrate = 1;
  555. }
  556. bitrate = **buf;
  557. } else if (bitrate < **buf) {
  558. av_log(avctx, AV_LOG_ERROR,
  559. "Buffer is too small for the claimed bitrate.\n");
  560. return I_F_Q;
  561. }
  562. (*buf)++;
  563. } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
  564. av_log(avctx, AV_LOG_WARNING,
  565. "Bitrate byte is missing, guessing the bitrate from packet size.\n");
  566. } else
  567. return I_F_Q;
  568. if (bitrate == SILENCE) {
  569. // FIXME: Remove this warning when tested with samples.
  570. avpriv_request_sample(avctx, "Blank frame handling");
  571. }
  572. return bitrate;
  573. }
  574. static void warn_insufficient_frame_quality(AVCodecContext *avctx,
  575. const char *message)
  576. {
  577. av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n",
  578. avctx->frame_number, message);
  579. }
  580. static void postfilter(QCELPContext *q, float *samples, float *lpc)
  581. {
  582. static const float pow_0_775[10] = {
  583. 0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
  584. 0.216676, 0.167924, 0.130141, 0.100859, 0.078166
  585. }, pow_0_625[10] = {
  586. 0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
  587. 0.059605, 0.037253, 0.023283, 0.014552, 0.009095
  588. };
  589. float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
  590. int n;
  591. for (n = 0; n < 10; n++) {
  592. lpc_s[n] = lpc[n] * pow_0_625[n];
  593. lpc_p[n] = lpc[n] * pow_0_775[n];
  594. }
  595. ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
  596. q->formant_mem + 10, 160, 10);
  597. memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
  598. ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
  599. memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
  600. ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
  601. ff_adaptive_gain_control(samples, pole_out + 10,
  602. avpriv_scalarproduct_float_c(q->formant_mem + 10,
  603. q->formant_mem + 10,
  604. 160),
  605. 160, 0.9375, &q->postfilter_agc_mem);
  606. }
  607. static int qcelp_decode_frame(AVCodecContext *avctx, void *data,
  608. int *got_frame_ptr, AVPacket *avpkt)
  609. {
  610. const uint8_t *buf = avpkt->data;
  611. int buf_size = avpkt->size;
  612. QCELPContext *q = avctx->priv_data;
  613. AVFrame *frame = data;
  614. float *outbuffer;
  615. int i, ret;
  616. float quantized_lspf[10], lpc[10];
  617. float gain[16];
  618. float *formant_mem;
  619. /* get output buffer */
  620. frame->nb_samples = 160;
  621. if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
  622. return ret;
  623. outbuffer = (float *)frame->data[0];
  624. if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
  625. warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
  626. goto erasure;
  627. }
  628. if (q->bitrate == RATE_OCTAVE &&
  629. (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
  630. warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
  631. goto erasure;
  632. }
  633. if (q->bitrate > SILENCE) {
  634. const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
  635. const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] +
  636. qcelp_unpacking_bitmaps_lengths[q->bitrate];
  637. uint8_t *unpacked_data = (uint8_t *)&q->frame;
  638. init_get_bits(&q->gb, buf, 8 * buf_size);
  639. memset(&q->frame, 0, sizeof(QCELPFrame));
  640. for (; bitmaps < bitmaps_end; bitmaps++)
  641. unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
  642. // Check for erasures/blanks on rates 1, 1/4 and 1/8.
  643. if (q->frame.reserved) {
  644. warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
  645. goto erasure;
  646. }
  647. if (q->bitrate == RATE_QUARTER &&
  648. codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
  649. warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
  650. goto erasure;
  651. }
  652. if (q->bitrate >= RATE_HALF) {
  653. for (i = 0; i < 4; i++) {
  654. if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
  655. warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
  656. goto erasure;
  657. }
  658. }
  659. }
  660. }
  661. decode_gain_and_index(q, gain);
  662. compute_svector(q, gain, outbuffer);
  663. if (decode_lspf(q, quantized_lspf) < 0) {
  664. warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
  665. goto erasure;
  666. }
  667. apply_pitch_filters(q, outbuffer);
  668. if (q->bitrate == I_F_Q) {
  669. erasure:
  670. q->bitrate = I_F_Q;
  671. q->erasure_count++;
  672. decode_gain_and_index(q, gain);
  673. compute_svector(q, gain, outbuffer);
  674. decode_lspf(q, quantized_lspf);
  675. apply_pitch_filters(q, outbuffer);
  676. } else
  677. q->erasure_count = 0;
  678. formant_mem = q->formant_mem + 10;
  679. for (i = 0; i < 4; i++) {
  680. interpolate_lpc(q, quantized_lspf, lpc, i);
  681. ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40, 10);
  682. formant_mem += 40;
  683. }
  684. // postfilter, as per TIA/EIA/IS-733 2.4.8.6
  685. postfilter(q, outbuffer, lpc);
  686. memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
  687. memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
  688. q->prev_bitrate = q->bitrate;
  689. *got_frame_ptr = 1;
  690. return buf_size;
  691. }
  692. AVCodec ff_qcelp_decoder = {
  693. .name = "qcelp",
  694. .type = AVMEDIA_TYPE_AUDIO,
  695. .id = AV_CODEC_ID_QCELP,
  696. .init = qcelp_decode_init,
  697. .decode = qcelp_decode_frame,
  698. .capabilities = CODEC_CAP_DR1,
  699. .priv_data_size = sizeof(QCELPContext),
  700. .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
  701. };