2 * Real Audio 1.0 (14.4K) encoder
3 * Copyright (c) 2010 Francesco Lavra <francescolavra@interfree.it>
5 * This file is part of FFmpeg.
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.
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.
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
24 * Real Audio 1.0 (14.4K) encoder
25 * @author Francesco Lavra <francescolavra@interfree.it>
32 #include "celp_filters.h"
36 static av_cold int ra144_encode_init(AVCodecContext * avctx)
41 if (avctx->sample_fmt != AV_SAMPLE_FMT_S16) {
42 av_log(avctx, AV_LOG_ERROR, "invalid sample format\n");
45 if (avctx->channels != 1) {
46 av_log(avctx, AV_LOG_ERROR, "invalid number of channels: %d\n",
50 avctx->frame_size = NBLOCKS * BLOCKSIZE;
51 avctx->bit_rate = 8000;
52 ractx = avctx->priv_data;
53 ractx->lpc_coef[0] = ractx->lpc_tables[0];
54 ractx->lpc_coef[1] = ractx->lpc_tables[1];
56 ret = ff_lpc_init(&ractx->lpc_ctx, avctx->frame_size, LPC_ORDER,
57 FF_LPC_TYPE_LEVINSON);
62 static av_cold int ra144_encode_close(AVCodecContext *avctx)
64 RA144Context *ractx = avctx->priv_data;
65 ff_lpc_end(&ractx->lpc_ctx);
71 * Quantize a value by searching a sorted table for the element with the
74 * @param value value to quantize
75 * @param table array containing the quantization table
76 * @param size size of the quantization table
77 * @return index of the quantization table corresponding to the element with the
80 static int quantize(int value, const int16_t *table, unsigned int size)
82 unsigned int low = 0, high = size - 1;
85 int index = (low + high) >> 1;
86 int error = table[index] - value;
89 return table[high] + error > value ? low : high;
100 * Orthogonalize a vector to another vector
102 * @param v vector to orthogonalize
103 * @param u vector against which orthogonalization is performed
105 static void orthogonalize(float *v, const float *u)
108 float num = 0, den = 0;
110 for (i = 0; i < BLOCKSIZE; i++) {
115 for (i = 0; i < BLOCKSIZE; i++)
121 * Calculate match score and gain of an LPC-filtered vector with respect to
122 * input data, possibly othogonalizing it to up to 2 other vectors
124 * @param work array used to calculate the filtered vector
125 * @param coefs coefficients of the LPC filter
126 * @param vect original vector
127 * @param ortho1 first vector against which orthogonalization is performed
128 * @param ortho2 second vector against which orthogonalization is performed
129 * @param data input data
130 * @param score pointer to variable where match score is returned
131 * @param gain pointer to variable where gain is returned
133 static void get_match_score(float *work, const float *coefs, float *vect,
134 const float *ortho1, const float *ortho2,
135 const float *data, float *score, float *gain)
140 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
142 orthogonalize(work, ortho1);
144 orthogonalize(work, ortho2);
146 for (i = 0; i < BLOCKSIZE; i++) {
147 g += work[i] * work[i];
148 c += data[i] * work[i];
160 * Create a vector from the adaptive codebook at a given lag value
162 * @param vect array where vector is stored
163 * @param cb adaptive codebook
164 * @param lag lag value
166 static void create_adapt_vect(float *vect, const int16_t *cb, int lag)
170 cb += BUFFERSIZE - lag;
171 for (i = 0; i < FFMIN(BLOCKSIZE, lag); i++)
174 for (i = 0; i < BLOCKSIZE - lag; i++)
175 vect[lag + i] = cb[i];
180 * Search the adaptive codebook for the best entry and gain and remove its
181 * contribution from input data
183 * @param adapt_cb array from which the adaptive codebook is extracted
184 * @param work array used to calculate LPC-filtered vectors
185 * @param coefs coefficients of the LPC filter
186 * @param data input data
187 * @return index of the best entry of the adaptive codebook
189 static int adaptive_cb_search(const int16_t *adapt_cb, float *work,
190 const float *coefs, float *data)
193 float score, gain, best_score, best_gain;
194 float exc[BLOCKSIZE];
196 gain = best_score = 0;
197 for (i = BLOCKSIZE / 2; i <= BUFFERSIZE; i++) {
198 create_adapt_vect(exc, adapt_cb, i);
199 get_match_score(work, coefs, exc, NULL, NULL, data, &score, &gain);
200 if (score > best_score) {
210 * Re-calculate the filtered vector from the vector with maximum match score
211 * and remove its contribution from input data.
213 create_adapt_vect(exc, adapt_cb, best_vect);
214 ff_celp_lp_synthesis_filterf(work, coefs, exc, BLOCKSIZE, LPC_ORDER);
215 for (i = 0; i < BLOCKSIZE; i++)
216 data[i] -= best_gain * work[i];
217 return best_vect - BLOCKSIZE / 2 + 1;
222 * Find the best vector of a fixed codebook by applying an LPC filter to
223 * codebook entries, possibly othogonalizing them to up to 2 other vectors and
224 * matching the results with input data
226 * @param work array used to calculate the filtered vectors
227 * @param coefs coefficients of the LPC filter
228 * @param cb fixed codebook
229 * @param ortho1 first vector against which orthogonalization is performed
230 * @param ortho2 second vector against which orthogonalization is performed
231 * @param data input data
232 * @param idx pointer to variable where the index of the best codebook entry is
234 * @param gain pointer to variable where the gain of the best codebook entry is
237 static void find_best_vect(float *work, const float *coefs,
238 const int8_t cb[][BLOCKSIZE], const float *ortho1,
239 const float *ortho2, float *data, int *idx,
243 float g, score, best_score;
244 float vect[BLOCKSIZE];
246 *idx = *gain = best_score = 0;
247 for (i = 0; i < FIXED_CB_SIZE; i++) {
248 for (j = 0; j < BLOCKSIZE; j++)
250 get_match_score(work, coefs, vect, ortho1, ortho2, data, &score, &g);
251 if (score > best_score) {
261 * Search the two fixed codebooks for the best entry and gain
263 * @param work array used to calculate LPC-filtered vectors
264 * @param coefs coefficients of the LPC filter
265 * @param data input data
266 * @param cba_idx index of the best entry of the adaptive codebook
267 * @param cb1_idx pointer to variable where the index of the best entry of the
268 * first fixed codebook is returned
269 * @param cb2_idx pointer to variable where the index of the best entry of the
270 * second fixed codebook is returned
272 static void fixed_cb_search(float *work, const float *coefs, float *data,
273 int cba_idx, int *cb1_idx, int *cb2_idx)
277 float cba_vect[BLOCKSIZE], cb1_vect[BLOCKSIZE];
278 float vect[BLOCKSIZE];
281 * The filtered vector from the adaptive codebook can be retrieved from
282 * work, because this function is called just after adaptive_cb_search().
285 memcpy(cba_vect, work, sizeof(cba_vect));
287 find_best_vect(work, coefs, ff_cb1_vects, cba_idx ? cba_vect : NULL, NULL,
288 data, cb1_idx, &gain);
291 * Re-calculate the filtered vector from the vector with maximum match score
292 * and remove its contribution from input data.
295 for (i = 0; i < BLOCKSIZE; i++)
296 vect[i] = ff_cb1_vects[*cb1_idx][i];
297 ff_celp_lp_synthesis_filterf(work, coefs, vect, BLOCKSIZE, LPC_ORDER);
299 orthogonalize(work, cba_vect);
300 for (i = 0; i < BLOCKSIZE; i++)
301 data[i] -= gain * work[i];
302 memcpy(cb1_vect, work, sizeof(cb1_vect));
307 find_best_vect(work, coefs, ff_cb2_vects, cba_idx ? cba_vect : NULL,
308 ortho_cb1 ? cb1_vect : NULL, data, cb2_idx, &gain);
313 * Encode a subblock of the current frame
315 * @param ractx encoder context
316 * @param sblock_data input data of the subblock
317 * @param lpc_coefs coefficients of the LPC filter
318 * @param rms RMS of the reflection coefficients
319 * @param pb pointer to PutBitContext of the current frame
321 static void ra144_encode_subblock(RA144Context *ractx,
322 const int16_t *sblock_data,
323 const int16_t *lpc_coefs, unsigned int rms,
326 float data[BLOCKSIZE], work[LPC_ORDER + BLOCKSIZE];
327 float coefs[LPC_ORDER];
328 float zero[BLOCKSIZE], cba[BLOCKSIZE], cb1[BLOCKSIZE], cb2[BLOCKSIZE];
329 int16_t cba_vect[BLOCKSIZE];
330 int cba_idx, cb1_idx, cb2_idx, gain;
333 float error, best_error;
335 for (i = 0; i < LPC_ORDER; i++) {
336 work[i] = ractx->curr_sblock[BLOCKSIZE + i];
337 coefs[i] = lpc_coefs[i] * (1/4096.0);
341 * Calculate the zero-input response of the LPC filter and subtract it from
344 memset(data, 0, sizeof(data));
345 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, data, BLOCKSIZE,
347 for (i = 0; i < BLOCKSIZE; i++) {
348 zero[i] = work[LPC_ORDER + i];
349 data[i] = sblock_data[i] - zero[i];
353 * Codebook search is performed without taking into account the contribution
354 * of the previous subblock, since it has been just subtracted from input
357 memset(work, 0, LPC_ORDER * sizeof(*work));
359 cba_idx = adaptive_cb_search(ractx->adapt_cb, work + LPC_ORDER, coefs,
363 * The filtered vector from the adaptive codebook can be retrieved from
364 * work, see implementation of adaptive_cb_search().
366 memcpy(cba, work + LPC_ORDER, sizeof(cba));
368 ff_copy_and_dup(cba_vect, ractx->adapt_cb, cba_idx + BLOCKSIZE / 2 - 1);
369 m[0] = (ff_irms(cba_vect) * rms) >> 12;
371 fixed_cb_search(work + LPC_ORDER, coefs, data, cba_idx, &cb1_idx, &cb2_idx);
372 for (i = 0; i < BLOCKSIZE; i++) {
373 cb1[i] = ff_cb1_vects[cb1_idx][i];
374 cb2[i] = ff_cb2_vects[cb2_idx][i];
376 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb1, BLOCKSIZE,
378 memcpy(cb1, work + LPC_ORDER, sizeof(cb1));
379 m[1] = (ff_cb1_base[cb1_idx] * rms) >> 8;
380 ff_celp_lp_synthesis_filterf(work + LPC_ORDER, coefs, cb2, BLOCKSIZE,
382 memcpy(cb2, work + LPC_ORDER, sizeof(cb2));
383 m[2] = (ff_cb2_base[cb2_idx] * rms) >> 8;
384 best_error = FLT_MAX;
386 for (n = 0; n < 256; n++) {
387 g[1] = ((ff_gain_val_tab[n][1] * m[1]) >> ff_gain_exp_tab[n]) *
389 g[2] = ((ff_gain_val_tab[n][2] * m[2]) >> ff_gain_exp_tab[n]) *
393 g[0] = ((ff_gain_val_tab[n][0] * m[0]) >> ff_gain_exp_tab[n]) *
395 for (i = 0; i < BLOCKSIZE; i++) {
396 data[i] = zero[i] + g[0] * cba[i] + g[1] * cb1[i] +
398 error += (data[i] - sblock_data[i]) *
399 (data[i] - sblock_data[i]);
402 for (i = 0; i < BLOCKSIZE; i++) {
403 data[i] = zero[i] + g[1] * cb1[i] + g[2] * cb2[i];
404 error += (data[i] - sblock_data[i]) *
405 (data[i] - sblock_data[i]);
408 if (error < best_error) {
413 put_bits(pb, 7, cba_idx);
414 put_bits(pb, 8, gain);
415 put_bits(pb, 7, cb1_idx);
416 put_bits(pb, 7, cb2_idx);
417 ff_subblock_synthesis(ractx, lpc_coefs, cba_idx, cb1_idx, cb2_idx, rms,
422 static int ra144_encode_frame(AVCodecContext *avctx, uint8_t *frame,
423 int buf_size, void *data)
425 static const uint8_t sizes[LPC_ORDER] = {64, 32, 32, 16, 16, 8, 8, 8, 8, 4};
426 static const uint8_t bit_sizes[LPC_ORDER] = {6, 5, 5, 4, 4, 3, 3, 3, 3, 2};
429 int32_t lpc_data[NBLOCKS * BLOCKSIZE];
430 int32_t lpc_coefs[LPC_ORDER][MAX_LPC_ORDER];
431 int shift[LPC_ORDER];
432 int16_t block_coefs[NBLOCKS][LPC_ORDER];
433 int lpc_refl[LPC_ORDER]; /**< reflection coefficients of the frame */
434 unsigned int refl_rms[NBLOCKS]; /**< RMS of the reflection coefficients */
438 if (buf_size < FRAMESIZE) {
439 av_log(avctx, AV_LOG_ERROR, "output buffer too small\n");
442 ractx = avctx->priv_data;
445 * Since the LPC coefficients are calculated on a frame centered over the
446 * fourth subframe, to encode a given frame, data from the next frame is
447 * needed. In each call to this function, the previous frame (whose data are
448 * saved in the encoder context) is encoded, and data from the current frame
449 * are saved in the encoder context to be used in the next function call.
451 for (i = 0; i < (2 * BLOCKSIZE + BLOCKSIZE / 2); i++) {
452 lpc_data[i] = ractx->curr_block[BLOCKSIZE + BLOCKSIZE / 2 + i];
453 energy += (lpc_data[i] * lpc_data[i]) >> 4;
455 for (i = 2 * BLOCKSIZE + BLOCKSIZE / 2; i < NBLOCKS * BLOCKSIZE; i++) {
456 lpc_data[i] = *((int16_t *)data + i - 2 * BLOCKSIZE - BLOCKSIZE / 2) >>
458 energy += (lpc_data[i] * lpc_data[i]) >> 4;
460 energy = ff_energy_tab[quantize(ff_t_sqrt(energy >> 5) >> 10, ff_energy_tab,
463 ff_lpc_calc_coefs(&ractx->lpc_ctx, lpc_data, NBLOCKS * BLOCKSIZE, LPC_ORDER,
464 LPC_ORDER, 16, lpc_coefs, shift, FF_LPC_TYPE_LEVINSON,
465 0, ORDER_METHOD_EST, 12, 0);
466 for (i = 0; i < LPC_ORDER; i++)
467 block_coefs[NBLOCKS - 1][i] = -(lpc_coefs[LPC_ORDER - 1][i] <<
468 (12 - shift[LPC_ORDER - 1]));
471 * TODO: apply perceptual weighting of the input speech through bandwidth
472 * expansion of the LPC filter.
475 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
477 * The filter is unstable: use the coefficients of the previous frame.
479 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[1]);
480 if (ff_eval_refl(lpc_refl, block_coefs[NBLOCKS - 1], avctx)) {
481 /* the filter is still unstable. set reflection coeffs to zero. */
482 memset(lpc_refl, 0, sizeof(lpc_refl));
485 init_put_bits(&pb, frame, buf_size);
486 for (i = 0; i < LPC_ORDER; i++) {
487 idx = quantize(lpc_refl[i], ff_lpc_refl_cb[i], sizes[i]);
488 put_bits(&pb, bit_sizes[i], idx);
489 lpc_refl[i] = ff_lpc_refl_cb[i][idx];
491 ractx->lpc_refl_rms[0] = ff_rms(lpc_refl);
492 ff_eval_coefs(ractx->lpc_coef[0], lpc_refl);
493 refl_rms[0] = ff_interp(ractx, block_coefs[0], 1, 1, ractx->old_energy);
494 refl_rms[1] = ff_interp(ractx, block_coefs[1], 2,
495 energy <= ractx->old_energy,
496 ff_t_sqrt(energy * ractx->old_energy) >> 12);
497 refl_rms[2] = ff_interp(ractx, block_coefs[2], 3, 0, energy);
498 refl_rms[3] = ff_rescale_rms(ractx->lpc_refl_rms[0], energy);
499 ff_int_to_int16(block_coefs[NBLOCKS - 1], ractx->lpc_coef[0]);
500 put_bits(&pb, 5, quantize(energy, ff_energy_tab, 32));
501 for (i = 0; i < NBLOCKS; i++)
502 ra144_encode_subblock(ractx, ractx->curr_block + i * BLOCKSIZE,
503 block_coefs[i], refl_rms[i], &pb);
505 ractx->old_energy = energy;
506 ractx->lpc_refl_rms[1] = ractx->lpc_refl_rms[0];
507 FFSWAP(unsigned int *, ractx->lpc_coef[0], ractx->lpc_coef[1]);
508 for (i = 0; i < NBLOCKS * BLOCKSIZE; i++)
509 ractx->curr_block[i] = *((int16_t *)data + i) >> 2;
514 AVCodec ff_ra_144_encoder = {
516 .type = AVMEDIA_TYPE_AUDIO,
517 .id = CODEC_ID_RA_144,
518 .priv_data_size = sizeof(RA144Context),
519 .init = ra144_encode_init,
520 .encode = ra144_encode_frame,
521 .close = ra144_encode_close,
522 .sample_fmts = (const enum AVSampleFormat[]){ AV_SAMPLE_FMT_S16,
523 AV_SAMPLE_FMT_NONE },
524 .long_name = NULL_IF_CONFIG_SMALL("RealAudio 1.0 (14.4K)"),