3 * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet
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
23 * @file libavcodec/qcelpdec.c
25 * @author Reynaldo H. Verdejo Pinochet
26 * @remark FFmpeg merging spearheaded by Kenan Gillet
27 * @remark Development mentored by Benjamin Larson
34 #include "bitstream.h"
36 #include "qcelpdata.h"
38 #include "celp_math.h"
39 #include "celp_filters.h"
40 #include "acelp_vectors.h"
47 I_F_Q = -1, /*!< insufficient frame quality */
58 qcelp_packet_rate bitrate;
59 QCELPFrame frame; /*!< unpacked data frame */
61 uint8_t erasure_count;
62 uint8_t octave_count; /*!< count the consecutive RATE_OCTAVE frames */
64 float predictor_lspf[10];/*!< LSP predictor for RATE_OCTAVE and I_F_Q */
65 float pitch_synthesis_filter_mem[303];
66 float pitch_pre_filter_mem[303];
67 float rnd_fir_filter_mem[180];
68 float formant_mem[170];
69 float last_codebook_gain;
75 uint8_t warned_buf_mismatch_bitrate;
79 * Reconstructs LPC coefficients from the line spectral pair frequencies.
81 * TIA/EIA/IS-733 2.4.3.3.5
83 void ff_celp_lspf2lpc(const double *lspf, float *lpc);
86 * Initialize the speech codec according to the specification.
88 * TIA/EIA/IS-733 2.4.9
90 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
92 QCELPContext *q = avctx->priv_data;
95 avctx->sample_fmt = SAMPLE_FMT_FLT;
98 q->prev_lspf[i] = (i+1)/11.;
104 * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
105 * transmission codes of any bitrate and checks for badly received packets.
107 * @param q the context
108 * @param lspf line spectral pair frequencies
110 * @return 0 on success, -1 if the packet is badly received
112 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
114 static int decode_lspf(QCELPContext *q, float *lspf)
117 float tmp_lspf, smooth, erasure_coeff;
118 const float *predictors;
120 if(q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q)
122 predictors = (q->prev_bitrate != RATE_OCTAVE &&
123 q->prev_bitrate != I_F_Q ?
124 q->prev_lspf : q->predictor_lspf);
126 if(q->bitrate == RATE_OCTAVE)
132 q->predictor_lspf[i] =
133 lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
134 : -QCELP_LSP_SPREAD_FACTOR)
135 + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
136 + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
138 smooth = (q->octave_count < 10 ? .875 : 0.1);
141 erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
143 assert(q->bitrate == I_F_Q);
145 if(q->erasure_count > 1)
146 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
150 q->predictor_lspf[i] =
151 lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
152 + erasure_coeff * predictors[i];
157 // Check the stability of the LSP frequencies.
158 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
160 lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
162 lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
164 lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
166 // Low-pass filter the LSP frequencies.
167 ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
175 lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
176 lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
179 // Check for badly received packets.
180 if(q->bitrate == RATE_QUARTER)
182 if(lspf[9] <= .70 || lspf[9] >= .97)
185 if(fabs(lspf[i] - lspf[i-2]) < .08)
189 if(lspf[9] <= .66 || lspf[9] >= .985)
192 if (fabs(lspf[i] - lspf[i-4]) < .0931)
200 * Converts codebook transmission codes to GAIN and INDEX.
202 * @param q the context
203 * @param gain array holding the decoded gain
205 * TIA/EIA/IS-733 2.4.6.2
207 static void decode_gain_and_index(QCELPContext *q,
209 int i, subframes_count, g1[16];
212 if(q->bitrate >= RATE_QUARTER)
216 case RATE_FULL: subframes_count = 16; break;
217 case RATE_HALF: subframes_count = 4; break;
218 default: subframes_count = 5;
220 for(i=0; i<subframes_count; i++)
222 g1[i] = 4 * q->frame.cbgain[i];
223 if(q->bitrate == RATE_FULL && !((i+1) & 3))
225 g1[i] += av_clip((g1[i-1] + g1[i-2] + g1[i-3]) / 3 - 6, 0, 32);
228 gain[i] = qcelp_g12ga[g1[i]];
230 if(q->frame.cbsign[i])
233 q->frame.cindex[i] = (q->frame.cindex[i]-89) & 127;
237 q->prev_g1[0] = g1[i-2];
238 q->prev_g1[1] = g1[i-1];
239 q->last_codebook_gain = qcelp_g12ga[g1[i-1]];
241 if(q->bitrate == RATE_QUARTER)
243 // Provide smoothing of the unvoiced excitation energy.
245 gain[6] = 0.4*gain[3] + 0.6*gain[4];
247 gain[4] = 0.8*gain[2] + 0.2*gain[3];
248 gain[3] = 0.2*gain[1] + 0.8*gain[2];
250 gain[1] = 0.6*gain[0] + 0.4*gain[1];
254 if(q->bitrate == RATE_OCTAVE)
256 g1[0] = 2 * q->frame.cbgain[0]
257 + av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
261 assert(q->bitrate == I_F_Q);
263 g1[0] = q->prev_g1[1];
264 switch(q->erasure_count)
267 case 2 : g1[0] -= 1; break;
268 case 3 : g1[0] -= 2; break;
275 // This interpolation is done to produce smoother background noise.
276 slope = 0.5*(qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
277 for(i=1; i<=subframes_count; i++)
278 gain[i-1] = q->last_codebook_gain + slope * i;
280 q->last_codebook_gain = gain[i-2];
281 q->prev_g1[0] = q->prev_g1[1];
282 q->prev_g1[1] = g1[0];
287 * If the received packet is Rate 1/4 a further sanity check is made of the
290 * @param cbgain the unpacked cbgain array
291 * @return -1 if the sanity check fails, 0 otherwise
293 * TIA/EIA/IS-733 2.4.8.7.3
295 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
297 int i, diff, prev_diff=0;
301 diff = cbgain[i] - cbgain[i-1];
304 else if(FFABS(diff - prev_diff) > 12)
312 * Computes the scaled codebook vector Cdn From INDEX and GAIN
315 * The specification lacks some information here.
317 * TIA/EIA/IS-733 has an omission on the codebook index determination
318 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
319 * you have to subtract the decoded index parameter from the given scaled
320 * codebook vector index 'n' to get the desired circular codebook index, but
321 * it does not mention that you have to clamp 'n' to [0-9] in order to get
322 * RI-compliant results.
324 * The reason for this mistake seems to be the fact they forgot to mention you
325 * have to do these calculations per codebook subframe and adjust given
326 * equation values accordingly.
328 * @param q the context
329 * @param gain array holding the 4 pitch subframe gain values
330 * @param cdn_vector array for the generated scaled codebook vector
332 static void compute_svector(QCELPContext *q, const float *gain,
336 uint16_t cbseed, cindex;
337 float *rnd, tmp_gain, fir_filter_value;
344 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
345 cindex = -q->frame.cindex[i];
347 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
353 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
354 cindex = -q->frame.cindex[i];
355 for (j = 0; j < 40; j++)
356 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
360 cbseed = (0x0003 & q->frame.lspv[4])<<14 |
361 (0x003F & q->frame.lspv[3])<< 8 |
362 (0x0060 & q->frame.lspv[2])<< 1 |
363 (0x0007 & q->frame.lspv[1])<< 3 |
364 (0x0038 & q->frame.lspv[0])>> 3 ;
365 rnd = q->rnd_fir_filter_mem + 20;
368 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
371 cbseed = 521 * cbseed + 259;
372 *rnd = (int16_t)cbseed;
375 fir_filter_value = 0.0;
377 fir_filter_value += qcelp_rnd_fir_coefs[j ]
378 * (rnd[-j ] + rnd[-20+j]);
380 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
381 *cdn_vector++ = tmp_gain * fir_filter_value;
385 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 20 * sizeof(float));
388 cbseed = q->first16bits;
391 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
394 cbseed = 521 * cbseed + 259;
395 *cdn_vector++ = tmp_gain * (int16_t)cbseed;
400 cbseed = -44; // random codebook index
403 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
405 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
409 memset(cdn_vector, 0, 160 * sizeof(float));
415 * Apply generic gain control.
417 * @param v_out output vector
418 * @param v_in gain-controlled vector
419 * @param v_ref vector to control gain of
421 * FIXME: If v_ref is a zero vector, it energy is zero
422 * and the behavior of the gain control is
423 * undefined in the specs.
425 * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
427 static void apply_gain_ctrl(float *v_out, const float *v_ref,
433 for(i=0, j=0; i<4; i++)
435 scalefactor = ff_dot_productf(v_in + j, v_in + j, 40);
437 scalefactor = sqrt(ff_dot_productf(v_ref + j, v_ref + j, 40)
440 ff_log_missing_feature(NULL, "Zero energy for gain control", 1);
441 for(len=j+40; j<len; j++)
442 v_out[j] = scalefactor * v_in[j];
447 * Apply filter in pitch-subframe steps.
449 * @param memory buffer for the previous state of the filter
450 * - must be able to contain 303 elements
451 * - the 143 first elements are from the previous state
452 * - the next 160 are for output
453 * @param v_in input filter vector
454 * @param gain per-subframe gain array, each element is between 0.0 and 2.0
455 * @param lag per-subframe lag array, each element is
456 * - between 16 and 143 if its corresponding pfrac is 0,
457 * - between 16 and 139 otherwise
458 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
461 * @return filter output vector
463 static const float *do_pitchfilter(float memory[303], const float v_in[160],
464 const float gain[4], const uint8_t *lag,
465 const uint8_t pfrac[4])
468 float *v_lag, *v_out;
471 v_out = memory + 143; // Output vector starts at memory[143].
477 v_lag = memory + 143 + 40 * i - lag[i];
478 for(v_len=v_in+40; v_in<v_len; v_in++)
480 if(pfrac[i]) // If it is a fractional lag...
482 for(j=0, *v_out=0.; j<4; j++)
483 *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
487 *v_out = *v_in + gain[i] * *v_out;
494 memcpy(v_out, v_in, 40 * sizeof(float));
500 memmove(memory, memory + 160, 143 * sizeof(float));
505 * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
506 * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
508 * @param q the context
509 * @param cdn_vector the scaled codebook vector
511 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
514 const float *v_synthesis_filtered, *v_pre_filtered;
516 if(q->bitrate >= RATE_HALF ||
517 q->bitrate == SILENCE ||
518 (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF)))
521 if(q->bitrate >= RATE_HALF)
524 // Compute gain & lag for the whole frame.
527 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
529 q->pitch_lag[i] = q->frame.plag[i] + 16;
533 float max_pitch_gain;
535 if (q->bitrate == I_F_Q)
537 if (q->erasure_count < 3)
538 max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
540 max_pitch_gain = 0.0;
543 assert(q->bitrate == SILENCE);
544 max_pitch_gain = 1.0;
547 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
549 memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
552 // pitch synthesis filter
553 v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
554 cdn_vector, q->pitch_gain,
555 q->pitch_lag, q->frame.pfrac);
557 // pitch prefilter update
559 q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
561 v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
562 v_synthesis_filtered,
563 q->pitch_gain, q->pitch_lag,
566 apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
569 memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
570 143 * sizeof(float));
571 memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
572 memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
573 memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
578 * Reconstructs LPC coefficients from the line spectral pair frequencies
579 * and performs bandwidth expansion.
581 * @param lspf line spectral pair frequencies
582 * @param lpc linear predictive coding coefficients
584 * @note: bandwith_expansion_coeff could be precalculated into a table
585 * but it seems to be slower on x86
587 * TIA/EIA/IS-733 2.4.3.3.5
589 void lspf2lpc(const float *lspf, float *lpc)
592 double bandwith_expansion_coeff = QCELP_BANDWITH_EXPANSION_COEFF;
596 lsf[i] = cos(M_PI * lspf[i]);
598 ff_celp_lspf2lpc(lsf, lpc);
602 lpc[i] *= bandwith_expansion_coeff;
603 bandwith_expansion_coeff *= QCELP_BANDWITH_EXPANSION_COEFF;
608 * Interpolates LSP frequencies and computes LPC coefficients
609 * for a given bitrate & pitch subframe.
611 * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
613 * @param q the context
614 * @param curr_lspf LSP frequencies vector of the current frame
615 * @param lpc float vector for the resulting LPC
616 * @param subframe_num frame number in decoded stream
618 void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
619 const int subframe_num)
621 float interpolated_lspf[10];
624 if(q->bitrate >= RATE_QUARTER)
625 weight = 0.25 * (subframe_num + 1);
626 else if(q->bitrate == RATE_OCTAVE && !subframe_num)
633 ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
634 weight, 1.0 - weight, 10);
635 lspf2lpc(interpolated_lspf, lpc);
636 }else if(q->bitrate >= RATE_QUARTER ||
637 (q->bitrate == I_F_Q && !subframe_num))
638 lspf2lpc(curr_lspf, lpc);
639 else if(q->bitrate == SILENCE && !subframe_num)
640 lspf2lpc(q->prev_lspf, lpc);
643 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
647 case 35: return RATE_FULL;
648 case 17: return RATE_HALF;
649 case 8: return RATE_QUARTER;
650 case 4: return RATE_OCTAVE;
651 case 1: return SILENCE;
658 * Determine the bitrate from the frame size and/or the first byte of the frame.
660 * @param avctx the AV codec context
661 * @param buf_size length of the buffer
662 * @param buf the bufffer
664 * @return the bitrate on success,
665 * I_F_Q if the bitrate cannot be satisfactorily determined
667 * TIA/EIA/IS-733 2.4.8.7.1
669 static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, const int buf_size,
672 qcelp_packet_rate bitrate;
674 if((bitrate = buf_size2bitrate(buf_size)) >= 0)
678 QCELPContext *q = avctx->priv_data;
679 if (!q->warned_buf_mismatch_bitrate)
681 av_log(avctx, AV_LOG_WARNING,
682 "Claimed bitrate and buffer size mismatch.\n");
683 q->warned_buf_mismatch_bitrate = 1;
686 }else if(bitrate < **buf)
688 av_log(avctx, AV_LOG_ERROR,
689 "Buffer is too small for the claimed bitrate.\n");
693 }else if((bitrate = buf_size2bitrate(buf_size + 1)) >= 0)
695 av_log(avctx, AV_LOG_WARNING,
696 "Bitrate byte is missing, guessing the bitrate from packet size.\n");
700 if(bitrate == SILENCE)
702 //FIXME: Remove experimental warning when tested with samples.
703 ff_log_ask_for_sample(avctx, "'Blank frame handling is experimental.");
708 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
711 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
715 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
716 const uint8_t *buf, int buf_size)
718 QCELPContext *q = avctx->priv_data;
719 float *outbuffer = data;
721 float quantized_lspf[10], lpc[10];
725 if((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q)
727 warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
731 if(q->bitrate == RATE_OCTAVE &&
732 (q->first16bits = AV_RB16(buf)) == 0xFFFF)
734 warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
738 if(q->bitrate > SILENCE)
740 const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
741 const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
742 + qcelp_unpacking_bitmaps_lengths[q->bitrate];
743 uint8_t *unpacked_data = (uint8_t *)&q->frame;
745 init_get_bits(&q->gb, buf, 8*buf_size);
747 memset(&q->frame, 0, sizeof(QCELPFrame));
749 for(; bitmaps < bitmaps_end; bitmaps++)
750 unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
752 // Check for erasures/blanks on rates 1, 1/4 and 1/8.
753 if(q->frame.reserved)
755 warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
758 if(q->bitrate == RATE_QUARTER &&
759 codebook_sanity_check_for_rate_quarter(q->frame.cbgain))
761 warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
765 if(q->bitrate >= RATE_HALF)
769 if(q->frame.pfrac[i] && q->frame.plag[i] >= 124)
771 warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
778 decode_gain_and_index(q, gain);
779 compute_svector(q, gain, outbuffer);
781 if(decode_lspf(q, quantized_lspf) < 0)
783 warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
788 apply_pitch_filters(q, outbuffer);
790 if(q->bitrate == I_F_Q)
795 decode_gain_and_index(q, gain);
796 compute_svector(q, gain, outbuffer);
797 decode_lspf(q, quantized_lspf);
798 apply_pitch_filters(q, outbuffer);
800 q->erasure_count = 0;
802 formant_mem = q->formant_mem + 10;
805 interpolate_lpc(q, quantized_lspf, lpc, i);
806 ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40,
810 memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
812 // FIXME: postfilter and final gain control should be here.
813 // TIA/EIA/IS-733 2.4.8.6
815 formant_mem = q->formant_mem + 10;
817 *outbuffer++ = av_clipf(*formant_mem++, QCELP_CLIP_LOWER_BOUND,
818 QCELP_CLIP_UPPER_BOUND);
820 memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
821 q->prev_bitrate = q->bitrate;
823 *data_size = 160 * sizeof(*outbuffer);
828 AVCodec qcelp_decoder =
831 .type = CODEC_TYPE_AUDIO,
832 .id = CODEC_ID_QCELP,
833 .init = qcelp_decode_init,
834 .decode = qcelp_decode_frame,
835 .priv_data_size = sizeof(QCELPContext),
836 .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),