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 Libav.
  6. *
  7. * Libav 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. * Libav 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 Libav; 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 Libav merging spearheaded by Kenan Gillet
  26. * @remark Development mentored by Benjamin Larson
  27. */
  28. #include <stddef.h>
  29. #include "libavutil/channel_layout.h"
  30. #include "libavutil/float_dsp.h"
  31. #include "avcodec.h"
  32. #include "bitstream.h"
  33. #include "internal.h"
  34. #include "qcelpdata.h"
  35. #include "celp_filters.h"
  36. #include "acelp_filters.h"
  37. #include "acelp_vectors.h"
  38. #include "lsp.h"
  39. #undef NDEBUG
  40. #include <assert.h>
  41. typedef enum {
  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 QCELPContext {
  50. BitstreamContext bc;
  51. qcelp_packet_rate bitrate;
  52. QCELPFrame frame; /**< unpacked data frame */
  53. uint8_t erasure_count;
  54. uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */
  55. float prev_lspf[10];
  56. float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */
  57. float pitch_synthesis_filter_mem[303];
  58. float pitch_pre_filter_mem[303];
  59. float rnd_fir_filter_mem[180];
  60. float formant_mem[170];
  61. float last_codebook_gain;
  62. int prev_g1[2];
  63. int prev_bitrate;
  64. float pitch_gain[4];
  65. uint8_t pitch_lag[4];
  66. uint16_t first16bits;
  67. uint8_t warned_buf_mismatch_bitrate;
  68. /* postfilter */
  69. float postfilter_synth_mem[10];
  70. float postfilter_agc_mem;
  71. float postfilter_tilt_mem;
  72. } QCELPContext;
  73. /**
  74. * Initialize the speech codec according to the specification.
  75. *
  76. * TIA/EIA/IS-733 2.4.9
  77. */
  78. static av_cold int qcelp_decode_init(AVCodecContext *avctx)
  79. {
  80. QCELPContext *q = avctx->priv_data;
  81. int i;
  82. avctx->channels = 1;
  83. avctx->channel_layout = AV_CH_LAYOUT_MONO;
  84. avctx->sample_fmt = AV_SAMPLE_FMT_FLT;
  85. for (i = 0; i < 10; i++)
  86. q->prev_lspf[i] = (i + 1) / 11.0;
  87. return 0;
  88. }
  89. /**
  90. * Decode the 10 quantized LSP frequencies from the LSPV/LSP
  91. * transmission codes of any bitrate and check for badly received packets.
  92. *
  93. * @param q the context
  94. * @param lspf line spectral pair frequencies
  95. *
  96. * @return 0 on success, -1 if the packet is badly received
  97. *
  98. * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
  99. */
  100. static int decode_lspf(QCELPContext *q, float *lspf)
  101. {
  102. int i;
  103. float tmp_lspf, smooth, erasure_coeff;
  104. const float *predictors;
  105. if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) {
  106. predictors = q->prev_bitrate != RATE_OCTAVE &&
  107. q->prev_bitrate != I_F_Q ? q->prev_lspf
  108. : q->predictor_lspf;
  109. if (q->bitrate == RATE_OCTAVE) {
  110. q->octave_count++;
  111. for (i = 0; i < 10; i++) {
  112. q->predictor_lspf[i] =
  113. lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
  114. : -QCELP_LSP_SPREAD_FACTOR) +
  115. predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR +
  116. (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11);
  117. }
  118. smooth = q->octave_count < 10 ? .875 : 0.1;
  119. } else {
  120. erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
  121. assert(q->bitrate == I_F_Q);
  122. if (q->erasure_count > 1)
  123. erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7;
  124. for (i = 0; i < 10; i++) {
  125. q->predictor_lspf[i] =
  126. lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 +
  127. erasure_coeff * predictors[i];
  128. }
  129. smooth = 0.125;
  130. }
  131. // Check the stability of the LSP frequencies.
  132. lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
  133. for (i = 1; i < 10; i++)
  134. lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR);
  135. lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR);
  136. for (i = 9; i > 0; i--)
  137. lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR);
  138. // Low-pass filter the LSP frequencies.
  139. ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10);
  140. } else {
  141. q->octave_count = 0;
  142. tmp_lspf = 0.0;
  143. for (i = 0; i < 5; i++) {
  144. lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
  145. lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
  146. }
  147. // Check for badly received packets.
  148. if (q->bitrate == RATE_QUARTER) {
  149. if (lspf[9] <= .70 || lspf[9] >= .97)
  150. return -1;
  151. for (i = 3; i < 10; i++)
  152. if (fabs(lspf[i] - lspf[i - 2]) < .08)
  153. return -1;
  154. } else {
  155. if (lspf[9] <= .66 || lspf[9] >= .985)
  156. return -1;
  157. for (i = 4; i < 10; i++)
  158. if (fabs(lspf[i] - lspf[i - 4]) < .0931)
  159. return -1;
  160. }
  161. }
  162. return 0;
  163. }
  164. /**
  165. * Convert codebook transmission codes to GAIN and INDEX.
  166. *
  167. * @param q the context
  168. * @param gain array holding the decoded gain
  169. *
  170. * TIA/EIA/IS-733 2.4.6.2
  171. */
  172. static void decode_gain_and_index(QCELPContext *q, float *gain)
  173. {
  174. int i, subframes_count, g1[16];
  175. float slope;
  176. if (q->bitrate >= RATE_QUARTER) {
  177. switch (q->bitrate) {
  178. case RATE_FULL: subframes_count = 16; break;
  179. case RATE_HALF: subframes_count = 4; break;
  180. default: subframes_count = 5;
  181. }
  182. for (i = 0; i < subframes_count; i++) {
  183. g1[i] = 4 * q->frame.cbgain[i];
  184. if (q->bitrate == RATE_FULL && !((i + 1) & 3)) {
  185. g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32);
  186. }
  187. gain[i] = qcelp_g12ga[g1[i]];
  188. if (q->frame.cbsign[i]) {
  189. gain[i] = -gain[i];
  190. q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127;
  191. }
  192. }
  193. q->prev_g1[0] = g1[i - 2];
  194. q->prev_g1[1] = g1[i - 1];
  195. q->last_codebook_gain = qcelp_g12ga[g1[i - 1]];
  196. if (q->bitrate == RATE_QUARTER) {
  197. // Provide smoothing of the unvoiced excitation energy.
  198. gain[7] = gain[4];
  199. gain[6] = 0.4 * gain[3] + 0.6 * gain[4];
  200. gain[5] = gain[3];
  201. gain[4] = 0.8 * gain[2] + 0.2 * gain[3];
  202. gain[3] = 0.2 * gain[1] + 0.8 * gain[2];
  203. gain[2] = gain[1];
  204. gain[1] = 0.6 * gain[0] + 0.4 * gain[1];
  205. }
  206. } else if (q->bitrate != SILENCE) {
  207. if (q->bitrate == RATE_OCTAVE) {
  208. g1[0] = 2 * q->frame.cbgain[0] +
  209. av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
  210. subframes_count = 8;
  211. } else {
  212. assert(q->bitrate == I_F_Q);
  213. g1[0] = q->prev_g1[1];
  214. switch (q->erasure_count) {
  215. case 1 : break;
  216. case 2 : g1[0] -= 1; break;
  217. case 3 : g1[0] -= 2; break;
  218. default: g1[0] -= 6;
  219. }
  220. if (g1[0] < 0)
  221. g1[0] = 0;
  222. subframes_count = 4;
  223. }
  224. // This interpolation is done to produce smoother background noise.
  225. slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
  226. for (i = 1; i <= subframes_count; i++)
  227. gain[i - 1] = q->last_codebook_gain + slope * i;
  228. q->last_codebook_gain = gain[i - 2];
  229. q->prev_g1[0] = q->prev_g1[1];
  230. q->prev_g1[1] = g1[0];
  231. }
  232. }
  233. /**
  234. * If the received packet is Rate 1/4 a further sanity check is made of the
  235. * codebook gain.
  236. *
  237. * @param cbgain the unpacked cbgain array
  238. * @return -1 if the sanity check fails, 0 otherwise
  239. *
  240. * TIA/EIA/IS-733 2.4.8.7.3
  241. */
  242. static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
  243. {
  244. int i, diff, prev_diff = 0;
  245. for (i = 1; i < 5; i++) {
  246. diff = cbgain[i] - cbgain[i-1];
  247. if (FFABS(diff) > 10)
  248. return -1;
  249. else if (FFABS(diff - prev_diff) > 12)
  250. return -1;
  251. prev_diff = diff;
  252. }
  253. return 0;
  254. }
  255. /**
  256. * Compute the scaled codebook vector Cdn From INDEX and GAIN
  257. * for all rates.
  258. *
  259. * The specification lacks some information here.
  260. *
  261. * TIA/EIA/IS-733 has an omission on the codebook index determination
  262. * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
  263. * you have to subtract the decoded index parameter from the given scaled
  264. * codebook vector index 'n' to get the desired circular codebook index, but
  265. * it does not mention that you have to clamp 'n' to [0-9] in order to get
  266. * RI-compliant results.
  267. *
  268. * The reason for this mistake seems to be the fact they forgot to mention you
  269. * have to do these calculations per codebook subframe and adjust given
  270. * equation values accordingly.
  271. *
  272. * @param q the context
  273. * @param gain array holding the 4 pitch subframe gain values
  274. * @param cdn_vector array for the generated scaled codebook vector
  275. */
  276. static void compute_svector(QCELPContext *q, const float *gain,
  277. float *cdn_vector)
  278. {
  279. int i, j, k;
  280. uint16_t cbseed, cindex;
  281. float *rnd, tmp_gain, fir_filter_value;
  282. switch (q->bitrate) {
  283. case RATE_FULL:
  284. for (i = 0; i < 16; i++) {
  285. tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
  286. cindex = -q->frame.cindex[i];
  287. for (j = 0; j < 10; j++)
  288. *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
  289. }
  290. break;
  291. case RATE_HALF:
  292. for (i = 0; i < 4; i++) {
  293. tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
  294. cindex = -q->frame.cindex[i];
  295. for (j = 0; j < 40; j++)
  296. *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
  297. }
  298. break;
  299. case RATE_QUARTER:
  300. cbseed = (0x0003 & q->frame.lspv[4]) << 14 |
  301. (0x003F & q->frame.lspv[3]) << 8 |
  302. (0x0060 & q->frame.lspv[2]) << 1 |
  303. (0x0007 & q->frame.lspv[1]) << 3 |
  304. (0x0038 & q->frame.lspv[0]) >> 3;
  305. rnd = q->rnd_fir_filter_mem + 20;
  306. for (i = 0; i < 8; i++) {
  307. tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
  308. for (k = 0; k < 20; k++) {
  309. cbseed = 521 * cbseed + 259;
  310. *rnd = (int16_t) cbseed;
  311. // FIR filter
  312. fir_filter_value = 0.0;
  313. for (j = 0; j < 10; j++)
  314. fir_filter_value += qcelp_rnd_fir_coefs[j] *
  315. (rnd[-j] + rnd[-20+j]);
  316. fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
  317. *cdn_vector++ = tmp_gain * fir_filter_value;
  318. rnd++;
  319. }
  320. }
  321. memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160,
  322. 20 * sizeof(float));
  323. break;
  324. case RATE_OCTAVE:
  325. cbseed = q->first16bits;
  326. for (i = 0; i < 8; i++) {
  327. tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
  328. for (j = 0; j < 20; j++) {
  329. cbseed = 521 * cbseed + 259;
  330. *cdn_vector++ = tmp_gain * (int16_t) cbseed;
  331. }
  332. }
  333. break;
  334. case I_F_Q:
  335. cbseed = -44; // random codebook index
  336. for (i = 0; i < 4; i++) {
  337. tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
  338. for (j = 0; j < 40; j++)
  339. *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
  340. }
  341. break;
  342. case SILENCE:
  343. memset(cdn_vector, 0, 160 * sizeof(float));
  344. break;
  345. }
  346. }
  347. /**
  348. * Apply generic gain control.
  349. *
  350. * @param v_out output vector
  351. * @param v_in gain-controlled vector
  352. * @param v_ref vector to control gain of
  353. *
  354. * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6
  355. */
  356. static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in)
  357. {
  358. int i;
  359. for (i = 0; i < 160; i += 40) {
  360. float res = avpriv_scalarproduct_float_c(v_ref + i, v_ref + i, 40);
  361. ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40);
  362. }
  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 = 0.0; 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 || q->bitrate == SILENCE ||
  423. (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) {
  424. if (q->bitrate >= RATE_HALF) {
  425. // Compute gain & lag for the whole frame.
  426. for (i = 0; i < 4; i++) {
  427. q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
  428. q->pitch_lag[i] = q->frame.plag[i] + 16;
  429. }
  430. } else {
  431. float max_pitch_gain;
  432. if (q->bitrate == I_F_Q) {
  433. if (q->erasure_count < 3)
  434. max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
  435. else
  436. max_pitch_gain = 0.0;
  437. } else {
  438. assert(q->bitrate == SILENCE);
  439. max_pitch_gain = 1.0;
  440. }
  441. for (i = 0; i < 4; i++)
  442. q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
  443. memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
  444. }
  445. // pitch synthesis filter
  446. v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
  447. cdn_vector, q->pitch_gain,
  448. q->pitch_lag, q->frame.pfrac);
  449. // pitch prefilter update
  450. for (i = 0; i < 4; i++)
  451. q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
  452. v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
  453. v_synthesis_filtered,
  454. q->pitch_gain, q->pitch_lag,
  455. q->frame.pfrac);
  456. apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
  457. } else {
  458. memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17, 143 * sizeof(float));
  459. memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
  460. memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
  461. memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
  462. }
  463. }
  464. /**
  465. * Reconstruct LPC coefficients from the line spectral pair frequencies
  466. * and perform bandwidth expansion.
  467. *
  468. * @param lspf line spectral pair frequencies
  469. * @param lpc linear predictive coding coefficients
  470. *
  471. * @note: bandwidth_expansion_coeff could be precalculated into a table
  472. * but it seems to be slower on x86
  473. *
  474. * TIA/EIA/IS-733 2.4.3.3.5
  475. */
  476. static void lspf2lpc(const float *lspf, float *lpc)
  477. {
  478. double lsp[10];
  479. double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF;
  480. int i;
  481. for (i = 0; i < 10; i++)
  482. lsp[i] = cos(M_PI * lspf[i]);
  483. ff_acelp_lspd2lpc(lsp, lpc, 5);
  484. for (i = 0; i < 10; i++) {
  485. lpc[i] *= bandwidth_expansion_coeff;
  486. bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF;
  487. }
  488. }
  489. /**
  490. * Interpolate LSP frequencies and compute LPC coefficients
  491. * for a given bitrate & pitch subframe.
  492. *
  493. * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
  494. *
  495. * @param q the context
  496. * @param curr_lspf LSP frequencies vector of the current frame
  497. * @param lpc float vector for the resulting LPC
  498. * @param subframe_num frame number in decoded stream
  499. */
  500. static void interpolate_lpc(QCELPContext *q, const float *curr_lspf,
  501. float *lpc, const int subframe_num)
  502. {
  503. float interpolated_lspf[10];
  504. float weight;
  505. if (q->bitrate >= RATE_QUARTER)
  506. weight = 0.25 * (subframe_num + 1);
  507. else if (q->bitrate == RATE_OCTAVE && !subframe_num)
  508. weight = 0.625;
  509. else
  510. weight = 1.0;
  511. if (weight != 1.0) {
  512. ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
  513. weight, 1.0 - weight, 10);
  514. lspf2lpc(interpolated_lspf, lpc);
  515. } else if (q->bitrate >= RATE_QUARTER ||
  516. (q->bitrate == I_F_Q && !subframe_num))
  517. lspf2lpc(curr_lspf, lpc);
  518. else if (q->bitrate == SILENCE && !subframe_num)
  519. lspf2lpc(q->prev_lspf, lpc);
  520. }
  521. static qcelp_packet_rate buf_size2bitrate(const int buf_size)
  522. {
  523. switch (buf_size) {
  524. case 35: return RATE_FULL;
  525. case 17: return RATE_HALF;
  526. case 8: return RATE_QUARTER;
  527. case 4: return RATE_OCTAVE;
  528. case 1: return SILENCE;
  529. }
  530. return I_F_Q;
  531. }
  532. /**
  533. * Determine the bitrate from the frame size and/or the first byte of the frame.
  534. *
  535. * @param avctx the AV codec context
  536. * @param buf_size length of the buffer
  537. * @param buf the buffer
  538. *
  539. * @return the bitrate on success,
  540. * I_F_Q if the bitrate cannot be satisfactorily determined
  541. *
  542. * TIA/EIA/IS-733 2.4.8.7.1
  543. */
  544. static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx,
  545. const int buf_size,
  546. const uint8_t **buf)
  547. {
  548. qcelp_packet_rate bitrate;
  549. if ((bitrate = buf_size2bitrate(buf_size)) >= 0) {
  550. if (bitrate > **buf) {
  551. QCELPContext *q = avctx->priv_data;
  552. if (!q->warned_buf_mismatch_bitrate) {
  553. av_log(avctx, AV_LOG_WARNING,
  554. "Claimed bitrate and buffer size mismatch.\n");
  555. q->warned_buf_mismatch_bitrate = 1;
  556. }
  557. bitrate = **buf;
  558. } else if (bitrate < **buf) {
  559. av_log(avctx, AV_LOG_ERROR,
  560. "Buffer is too small for the claimed bitrate.\n");
  561. return I_F_Q;
  562. }
  563. (*buf)++;
  564. } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) {
  565. av_log(avctx, AV_LOG_WARNING,
  566. "Bitrate byte is missing, guessing the bitrate from packet size.\n");
  567. } else
  568. return I_F_Q;
  569. if (bitrate == SILENCE) {
  570. // FIXME: Remove this warning when tested with samples.
  571. avpriv_request_sample(avctx, "Blank frame handling");
  572. }
  573. return bitrate;
  574. }
  575. static void warn_insufficient_frame_quality(AVCodecContext *avctx,
  576. const char *message)
  577. {
  578. av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n",
  579. avctx->frame_number, message);
  580. }
  581. static void postfilter(QCELPContext *q, float *samples, float *lpc)
  582. {
  583. static const float pow_0_775[10] = {
  584. 0.775000, 0.600625, 0.465484, 0.360750, 0.279582,
  585. 0.216676, 0.167924, 0.130141, 0.100859, 0.078166
  586. }, pow_0_625[10] = {
  587. 0.625000, 0.390625, 0.244141, 0.152588, 0.095367,
  588. 0.059605, 0.037253, 0.023283, 0.014552, 0.009095
  589. };
  590. float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160];
  591. int n;
  592. for (n = 0; n < 10; n++) {
  593. lpc_s[n] = lpc[n] * pow_0_625[n];
  594. lpc_p[n] = lpc[n] * pow_0_775[n];
  595. }
  596. ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s,
  597. q->formant_mem + 10, 160, 10);
  598. memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10);
  599. ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10);
  600. memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10);
  601. ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160);
  602. ff_adaptive_gain_control(samples, pole_out + 10,
  603. avpriv_scalarproduct_float_c(q->formant_mem + 10,
  604. q->formant_mem + 10,
  605. 160),
  606. 160, 0.9375, &q->postfilter_agc_mem);
  607. }
  608. static int qcelp_decode_frame(AVCodecContext *avctx, void *data,
  609. int *got_frame_ptr, AVPacket *avpkt)
  610. {
  611. const uint8_t *buf = avpkt->data;
  612. int buf_size = avpkt->size;
  613. QCELPContext *q = avctx->priv_data;
  614. AVFrame *frame = data;
  615. float *outbuffer;
  616. int i, ret;
  617. float quantized_lspf[10], lpc[10];
  618. float gain[16];
  619. float *formant_mem;
  620. /* get output buffer */
  621. frame->nb_samples = 160;
  622. if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) {
  623. av_log(avctx, AV_LOG_ERROR, "get_buffer() failed\n");
  624. return ret;
  625. }
  626. outbuffer = (float *)frame->data[0];
  627. if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) {
  628. warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
  629. goto erasure;
  630. }
  631. if (q->bitrate == RATE_OCTAVE &&
  632. (q->first16bits = AV_RB16(buf)) == 0xFFFF) {
  633. warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
  634. goto erasure;
  635. }
  636. if (q->bitrate > SILENCE) {
  637. const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
  638. const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] +
  639. qcelp_unpacking_bitmaps_lengths[q->bitrate];
  640. uint8_t *unpacked_data = (uint8_t *)&q->frame;
  641. bitstream_init8(&q->bc, buf, buf_size);
  642. memset(&q->frame, 0, sizeof(QCELPFrame));
  643. for (; bitmaps < bitmaps_end; bitmaps++)
  644. unpacked_data[bitmaps->index] |= bitstream_read(&q->bc, bitmaps->bitlen) << bitmaps->bitpos;
  645. // Check for erasures/blanks on rates 1, 1/4 and 1/8.
  646. if (q->frame.reserved) {
  647. warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
  648. goto erasure;
  649. }
  650. if (q->bitrate == RATE_QUARTER &&
  651. codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) {
  652. warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
  653. goto erasure;
  654. }
  655. if (q->bitrate >= RATE_HALF) {
  656. for (i = 0; i < 4; i++) {
  657. if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) {
  658. warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
  659. goto erasure;
  660. }
  661. }
  662. }
  663. }
  664. decode_gain_and_index(q, gain);
  665. compute_svector(q, gain, outbuffer);
  666. if (decode_lspf(q, quantized_lspf) < 0) {
  667. warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
  668. goto erasure;
  669. }
  670. apply_pitch_filters(q, outbuffer);
  671. if (q->bitrate == I_F_Q) {
  672. erasure:
  673. q->bitrate = I_F_Q;
  674. q->erasure_count++;
  675. decode_gain_and_index(q, gain);
  676. compute_svector(q, gain, outbuffer);
  677. decode_lspf(q, quantized_lspf);
  678. apply_pitch_filters(q, outbuffer);
  679. } else
  680. q->erasure_count = 0;
  681. formant_mem = q->formant_mem + 10;
  682. for (i = 0; i < 4; i++) {
  683. interpolate_lpc(q, quantized_lspf, lpc, i);
  684. ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40, 10);
  685. formant_mem += 40;
  686. }
  687. // postfilter, as per TIA/EIA/IS-733 2.4.8.6
  688. postfilter(q, outbuffer, lpc);
  689. memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
  690. memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
  691. q->prev_bitrate = q->bitrate;
  692. *got_frame_ptr = 1;
  693. return buf_size;
  694. }
  695. AVCodec ff_qcelp_decoder = {
  696. .name = "qcelp",
  697. .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),
  698. .type = AVMEDIA_TYPE_AUDIO,
  699. .id = AV_CODEC_ID_QCELP,
  700. .init = qcelp_decode_init,
  701. .decode = qcelp_decode_frame,
  702. .capabilities = AV_CODEC_CAP_DR1,
  703. .priv_data_size = sizeof(QCELPContext),
  704. };