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.

801 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 "avcodec.h"
  30. #include "internal.h"
  31. #include "get_bits.h"
  32. #include "qcelpdata.h"
  33. #include "celp_math.h"
  34. #include "celp_filters.h"
  35. #include "acelp_filters.h"
  36. #include "acelp_vectors.h"
  37. #include "lsp.h"
  38. #undef NDEBUG
  39. #include <assert.h>
  40. typedef enum
  41. {
  42. I_F_Q = -1, /**< insufficient frame quality */
  43. SILENCE,
  44. RATE_OCTAVE,
  45. RATE_QUARTER,
  46. RATE_HALF,
  47. RATE_FULL
  48. } qcelp_packet_rate;
  49. typedef struct
  50. {
  51. GetBitContext gb;
  52. qcelp_packet_rate bitrate;
  53. QCELPFrame frame; /**< unpacked data frame */
  54. uint8_t erasure_count;
  55. uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
  56. float prev_lspf[10];
  57. float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
  58. float pitch_synthesis_filter_mem[303];
  59. float pitch_pre_filter_mem[303];
  60. float rnd_fir_filter_mem[180];
  61. float formant_mem[170];
  62. float last_codebook_gain;
  63. int prev_g1[2];
  64. int prev_bitrate;
  65. float pitch_gain[4];
  66. uint8_t pitch_lag[4];
  67. uint16_t first16bits;
  68. uint8_t warned_buf_mismatch_bitrate;
  69. /* postfilter */
  70. float postfilter_synth_mem[10];
  71. float postfilter_agc_mem;
  72. float postfilter_tilt_mem;
  73. } QCELPContext;
  74. /**
  75. * Initialize the speech codec according to the specification.
  76. *
  77. * TIA/EIA/IS-733 2.4.9
  78. */
  79. static av_cold int qcelp_decode_init(AVCodecContext *avctx)
  80. {
  81. QCELPContext *q = avctx->priv_data;
  82. int i;
  83. avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
  84. for(i=0; i<10; i++)
  85. q->prev_lspf[i] = (i+1)/11.;
  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 ?
  107. q->prev_lspf : 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. assert(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.;
  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,
  172. float *gain) {
  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. assert(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, 20 * sizeof(float));
  321. break;
  322. case RATE_OCTAVE:
  323. cbseed = q->first16bits;
  324. for (i = 0; i < 8; i++) {
  325. tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
  326. for (j = 0; j < 20; j++) {
  327. cbseed = 521 * cbseed + 259;
  328. *cdn_vector++ = tmp_gain * (int16_t)cbseed;
  329. }
  330. }
  331. break;
  332. case I_F_Q:
  333. cbseed = -44; // random codebook index
  334. for (i = 0; i < 4; i++) {
  335. tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
  336. for(j=0; j<40; j++)
  337. *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
  338. }
  339. break;
  340. case SILENCE:
  341. memset(cdn_vector, 0, 160 * sizeof(float));
  342. break;
  343. }
  344. }
  345. /**
  346. * Apply generic gain control.
  347. *
  348. * @param v_out output vector
  349. * @param v_in gain-controlled vector
  350. * @param v_ref vector to control gain of
  351. *
  352. * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
  353. */
  354. static void apply_gain_ctrl(float *v_out, const float *v_ref,
  355. const float *v_in)
  356. {
  357. int i;
  358. for (i = 0; i < 160; i += 40)
  359. ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i,
  360. ff_dot_productf(v_ref + i,
  361. v_ref + i, 40),
  362. 40);
  363. }
  364. /**
  365. * Apply filter in pitch-subframe steps.
  366. *
  367. * @param memory buffer for the previous state of the filter
  368. * - must be able to contain 303 elements
  369. * - the 143 first elements are from the previous state
  370. * - the next 160 are for output
  371. * @param v_in input filter vector
  372. * @param gain per-subframe gain array, each element is between 0.0 and 2.0
  373. * @param lag per-subframe lag array, each element is
  374. * - between 16 and 143 if its corresponding pfrac is 0,
  375. * - between 16 and 139 otherwise
  376. * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
  377. * otherwise
  378. *
  379. * @return filter output vector
  380. */
  381. static const float *do_pitchfilter(float memory[303], const float v_in[160],
  382. const float gain[4], const uint8_t *lag,
  383. const uint8_t pfrac[4])
  384. {
  385. int i, j;
  386. float *v_lag, *v_out;
  387. const float *v_len;
  388. v_out = memory + 143; // Output vector starts at memory[143].
  389. for (i = 0; i < 4; i++) {
  390. if (gain[i]) {
  391. v_lag = memory + 143 + 40 * i - lag[i];
  392. for (v_len = v_in + 40; v_in < v_len; v_in++) {
  393. if (pfrac[i]) { // If it is a fractional lag...
  394. for(j=0, *v_out=0.; j<4; j++)
  395. *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
  396. }else
  397. *v_out = *v_lag;
  398. *v_out = *v_in + gain[i] * *v_out;
  399. v_lag++;
  400. v_out++;
  401. }
  402. } else {
  403. memcpy(v_out, v_in, 40 * sizeof(float));
  404. v_in += 40;
  405. v_out += 40;
  406. }
  407. }
  408. memmove(memory, memory + 160, 143 * sizeof(float));
  409. return memory + 143;
  410. }
  411. /**
  412. * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
  413. * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
  414. *
  415. * @param q the context
  416. * @param cdn_vector the scaled codebook vector
  417. */
  418. static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
  419. {
  420. int i;
  421. const float *v_synthesis_filtered, *v_pre_filtered;
  422. if(q->bitrate >= RATE_HALF ||
  423. q->bitrate == SILENCE ||
  424. (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
  425. if(q->bitrate >= RATE_HALF) {
  426. // Compute gain & lag for the whole frame.
  427. for (i = 0; i < 4; i++) {
  428. q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
  429. q->pitch_lag[i] = q->frame.plag[i] + 16;
  430. }
  431. } else {
  432. float max_pitch_gain;
  433. if (q->bitrate == I_F_Q) {
  434. if (q->erasure_count < 3)
  435. max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
  436. else
  437. max_pitch_gain = 0.0;
  438. } else {
  439. assert(q->bitrate == SILENCE);
  440. max_pitch_gain = 1.0;
  441. }
  442. for(i=0; i<4; i++)
  443. q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
  444. memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
  445. }
  446. // pitch synthesis filter
  447. v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
  448. cdn_vector, q->pitch_gain,
  449. q->pitch_lag, q->frame.pfrac);
  450. // pitch prefilter update
  451. for(i=0; i<4; i++)
  452. q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
  453. v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
  454. v_synthesis_filtered,
  455. q->pitch_gain, q->pitch_lag,
  456. q->frame.pfrac);
  457. apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
  458. } else {
  459. memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
  460. 143 * sizeof(float));
  461. memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
  462. memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
  463. memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
  464. }
  465. }
  466. /**
  467. * Reconstruct LPC coefficients from the line spectral pair frequencies
  468. * and perform bandwidth expansion.
  469. *
  470. * @param lspf line spectral pair frequencies
  471. * @param lpc linear predictive coding coefficients
  472. *
  473. * @note: bandwidth_expansion_coeff could be precalculated into a table
  474. * but it seems to be slower on x86
  475. *
  476. * TIA/EIA/IS-733 2.4.3.3.5
  477. */
  478. static void lspf2lpc(const float *lspf, float *lpc)
  479. {
  480. double lsp[10];
  481. double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
  482. int i;
  483. for (i=0; i<10; i++)
  484. lsp[i] = cos(M_PI * lspf[i]);
  485. ff_acelp_lspd2lpc(lsp, lpc, 5);
  486. for (i = 0; i < 10; i++) {
  487. lpc[i] *= bandwidth_expansion_coeff;
  488. bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
  489. }
  490. }
  491. /**
  492. * Interpolate LSP frequencies and compute LPC coefficients
  493. * for a given bitrate & pitch subframe.
  494. *
  495. * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
  496. *
  497. * @param q the context
  498. * @param curr_lspf LSP frequencies vector of the current frame
  499. * @param lpc float vector for the resulting LPC
  500. * @param subframe_num frame number in decoded stream
  501. */
  502. static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
  503. float *lpc, const int subframe_num)
  504. {
  505. float interpolated_lspf[10];
  506. float weight;
  507. if(q->bitrate >= RATE_QUARTER)
  508. weight = 0.25 * (subframe_num + 1);
  509. else if(q->bitrate == RATE_OCTAVE && !subframe_num)
  510. weight = 0.625;
  511. else
  512. weight = 1.0;
  513. if (weight != 1.0) {
  514. ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
  515. weight, 1.0 - weight, 10);
  516. lspf2lpc(interpolated_lspf, lpc);
  517. }else if(q->bitrate >= RATE_QUARTER ||
  518. (q->bitrate == I_F_Q && !subframe_num))
  519. lspf2lpc(curr_lspf, lpc);
  520. else if(q->bitrate == SILENCE && !subframe_num)
  521. lspf2lpc(q->prev_lspf, lpc);
  522. }
  523. static qcelp_packet_rate buf_size2bitrate(const int buf_size)
  524. {
  525. switch (buf_size) {
  526. case 35: return RATE_FULL;
  527. case 17: return RATE_HALF;
  528. case 8: return RATE_QUARTER;
  529. case 4: return RATE_OCTAVE;
  530. case 1: return SILENCE;
  531. }
  532. return I_F_Q;
  533. }
  534. /**
  535. * Determine the bitrate from the frame size and/or the first byte of the frame.
  536. *
  537. * @param avctx the AV codec context
  538. * @param buf_size length of the buffer
  539. * @param buf the bufffer
  540. *
  541. * @return the bitrate on success,
  542. * I_F_Q if the bitrate cannot be satisfactorily determined
  543. *
  544. * TIA/EIA/IS-733 2.4.8.7.1
  545. */
  546. static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, const int buf_size,
  547. const uint8_t **buf)
  548. {
  549. qcelp_packet_rate bitrate;
  550. if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
  551. if (bitrate > **buf) {
  552. QCELPContext *q = avctx->priv_data;
  553. if (!q->warned_buf_mismatch_bitrate) {
  554. av_log(avctx, AV_LOG_WARNING,
  555. "Claimed bitrate and buffer size mismatch.\n");
  556. q->warned_buf_mismatch_bitrate = 1;
  557. }
  558. bitrate = **buf;
  559. } else if (bitrate < **buf) {
  560. av_log(avctx, AV_LOG_ERROR,
  561. "Buffer is too small for the claimed bitrate.\n");
  562. return I_F_Q;
  563. }
  564. (*buf)++;
  565. } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
  566. av_log(avctx, AV_LOG_WARNING,
  567. "Bitrate byte is missing, guessing the bitrate from packet size.\n");
  568. }else
  569. return I_F_Q;
  570. if (bitrate == SILENCE) {
  571. //FIXME: Remove experimental warning when tested with samples.
  572. av_log_ask_for_sample(avctx, "'Blank frame handling is experimental.");
  573. }
  574. return bitrate;
  575. }
  576. static void warn_insufficient_frame_quality(AVCodecContext *avctx,
  577. const char *message)
  578. {
  579. av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
  580. message);
  581. }
  582. static void postfilter(QCELPContext *q, float *samples, float *lpc)
  583. {
  584. static const float pow_0_775[10] = {
  585. 0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
  586. 0.216676, 0.167924, 0.130141, 0.100859, 0.078166
  587. }, pow_0_625[10] = {
  588. 0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
  589. 0.059605, 0.037253, 0.023283, 0.014552, 0.009095
  590. };
  591. float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
  592. int n;
  593. for (n = 0; n < 10; n++) {
  594. lpc_s[n] = lpc[n] * pow_0_625[n];
  595. lpc_p[n] = lpc[n] * pow_0_775[n];
  596. }
  597. ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
  598. q->formant_mem + 10, 160, 10);
  599. memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
  600. ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
  601. memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
  602. ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
  603. ff_adaptive_gain_control(samples, pole_out + 10,
  604. ff_dot_productf(q->formant_mem + 10, q->formant_mem + 10, 160),
  605. 160, 0.9375, &q->postfilter_agc_mem);
  606. }
  607. static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
  608. AVPacket *avpkt)
  609. {
  610. const uint8_t *buf = avpkt->data;
  611. int buf_size = avpkt->size;
  612. QCELPContext *q = avctx->priv_data;
  613. float *outbuffer = data;
  614. int i, out_size;
  615. float quantized_lspf[10], lpc[10];
  616. float gain[16];
  617. float *formant_mem;
  618. out_size = 160 * av_get_bytes_per_sample(avctx->sample_fmt);
  619. if (*data_size < out_size) {
  620. av_log(avctx, AV_LOG_ERROR, "Output buffer is too small\n");
  621. return AVERROR(EINVAL);
  622. }
  623. if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
  624. warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
  625. goto erasure;
  626. }
  627. if(q->bitrate == RATE_OCTAVE &&
  628. (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
  629. warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
  630. goto erasure;
  631. }
  632. if (q->bitrate > SILENCE) {
  633. const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
  634. const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
  635. + qcelp_unpacking_bitmaps_lengths[q->bitrate];
  636. uint8_t *unpacked_data = (uint8_t *)&q->frame;
  637. init_get_bits(&q->gb, buf, 8*buf_size);
  638. memset(&q->frame, 0, sizeof(QCELPFrame));
  639. for(; bitmaps < bitmaps_end; bitmaps++)
  640. unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
  641. // Check for erasures/blanks on rates 1, 1/4 and 1/8.
  642. if (q->frame.reserved) {
  643. warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
  644. goto erasure;
  645. }
  646. if(q->bitrate == RATE_QUARTER &&
  647. codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
  648. warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
  649. goto erasure;
  650. }
  651. if (q->bitrate >= RATE_HALF) {
  652. for (i = 0; i < 4; i++) {
  653. if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
  654. warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
  655. goto erasure;
  656. }
  657. }
  658. }
  659. }
  660. decode_gain_and_index(q, gain);
  661. compute_svector(q, gain, outbuffer);
  662. if (decode_lspf(q, quantized_lspf) < 0) {
  663. warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
  664. goto erasure;
  665. }
  666. apply_pitch_filters(q, outbuffer);
  667. if (q->bitrate == I_F_Q) {
  668. erasure:
  669. q->bitrate = I_F_Q;
  670. q->erasure_count++;
  671. decode_gain_and_index(q, gain);
  672. compute_svector(q, gain, outbuffer);
  673. decode_lspf(q, quantized_lspf);
  674. apply_pitch_filters(q, outbuffer);
  675. }else
  676. q->erasure_count = 0;
  677. formant_mem = q->formant_mem + 10;
  678. for (i = 0; i < 4; i++) {
  679. interpolate_lpc(q, quantized_lspf, lpc, i);
  680. ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40,
  681. 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. *data_size = out_size;
  690. return buf_size;
  691. }
  692. AVCodec ff_qcelp_decoder =
  693. {
  694. .name = "qcelp",
  695. .type = AVMEDIA_TYPE_AUDIO,
  696. .id = CODEC_ID_QCELP,
  697. .init = qcelp_decode_init,
  698. .decode = qcelp_decode_frame,
  699. .priv_data_size = sizeof(QCELPContext),
  700. .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
  701. };