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"
46 I_F_Q = -1, /*!< insufficient frame quality */
57 qcelp_packet_rate bitrate;
58 QCELPFrame frame; /*!< unpacked data frame */
60 uint8_t erasure_count;
61 uint8_t octave_count; /*!< count the consecutive RATE_OCTAVE frames */
63 float predictor_lspf[10];/*!< LSP predictor for RATE_OCTAVE and I_F_Q */
64 float pitch_synthesis_filter_mem[303];
65 float pitch_pre_filter_mem[303];
66 float rnd_fir_filter_mem[180];
67 float formant_mem[170];
68 float last_codebook_gain;
74 uint8_t warned_buf_mismatch_bitrate;
78 * Reconstructs LPC coefficients from the line spectral pair frequencies.
80 * TIA/EIA/IS-733 2.4.3.3.5
82 void ff_qcelp_lspf2lpc(const float *lspf, float *lpc);
84 static void weighted_vector_sumf(float *out, const float *in_a,
85 const float *in_b, float weight_coeff_a,
86 float weight_coeff_b, int length)
90 for(i=0; i<length; i++)
91 out[i] = weight_coeff_a * in_a[i]
92 + weight_coeff_b * in_b[i];
96 * Initialize the speech codec according to the specification.
98 * TIA/EIA/IS-733 2.4.9
100 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
102 QCELPContext *q = avctx->priv_data;
105 avctx->sample_fmt = SAMPLE_FMT_FLT;
108 q->prev_lspf[i] = (i+1)/11.;
114 * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
115 * transmission codes of any bitrate and checks for badly received packets.
117 * @param q the context
118 * @param lspf line spectral pair frequencies
120 * @return 0 on success, -1 if the packet is badly received
122 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
124 static int decode_lspf(QCELPContext *q, float *lspf)
127 float tmp_lspf, smooth, erasure_coeff;
128 const float *predictors;
130 if(q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q)
132 predictors = (q->prev_bitrate != RATE_OCTAVE &&
133 q->prev_bitrate != I_F_Q ?
134 q->prev_lspf : q->predictor_lspf);
136 if(q->bitrate == RATE_OCTAVE)
142 q->predictor_lspf[i] =
143 lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
144 : -QCELP_LSP_SPREAD_FACTOR)
145 + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
146 + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
148 smooth = (q->octave_count < 10 ? .875 : 0.1);
151 erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
153 assert(q->bitrate == I_F_Q);
155 if(q->erasure_count > 1)
156 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
160 q->predictor_lspf[i] =
161 lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
162 + erasure_coeff * predictors[i];
167 // Check the stability of the LSP frequencies.
168 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
170 lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
172 lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
174 lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
176 // Low-pass filter the LSP frequencies.
177 weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
185 lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
186 lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
189 // Check for badly received packets.
190 if(q->bitrate == RATE_QUARTER)
192 if(lspf[9] <= .70 || lspf[9] >= .97)
195 if(fabs(lspf[i] - lspf[i-2]) < .08)
199 if(lspf[9] <= .66 || lspf[9] >= .985)
202 if (fabs(lspf[i] - lspf[i-4]) < .0931)
210 * Converts codebook transmission codes to GAIN and INDEX.
212 * @param q the context
213 * @param gain array holding the decoded gain
215 * TIA/EIA/IS-733 2.4.6.2
217 static void decode_gain_and_index(QCELPContext *q,
219 int i, subframes_count, g1[16];
222 if(q->bitrate >= RATE_QUARTER)
226 case RATE_FULL: subframes_count = 16; break;
227 case RATE_HALF: subframes_count = 4; break;
228 default: subframes_count = 5;
230 for(i=0; i<subframes_count; i++)
232 g1[i] = 4 * q->frame.cbgain[i];
233 if(q->bitrate == RATE_FULL && !((i+1) & 3))
235 g1[i] += av_clip((g1[i-1] + g1[i-2] + g1[i-3]) / 3 - 6, 0, 32);
238 gain[i] = qcelp_g12ga[g1[i]];
240 if(q->frame.cbsign[i])
243 q->frame.cindex[i] = (q->frame.cindex[i]-89) & 127;
247 q->prev_g1[0] = g1[i-2];
248 q->prev_g1[1] = g1[i-1];
249 q->last_codebook_gain = qcelp_g12ga[g1[i-1]];
251 if(q->bitrate == RATE_QUARTER)
253 // Provide smoothing of the unvoiced excitation energy.
255 gain[6] = 0.4*gain[3] + 0.6*gain[4];
257 gain[4] = 0.8*gain[2] + 0.2*gain[3];
258 gain[3] = 0.2*gain[1] + 0.8*gain[2];
260 gain[1] = 0.6*gain[0] + 0.4*gain[1];
264 if(q->bitrate == RATE_OCTAVE)
266 g1[0] = 2 * q->frame.cbgain[0]
267 + av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
271 assert(q->bitrate == I_F_Q);
273 g1[0] = q->prev_g1[1];
274 switch(q->erasure_count)
277 case 2 : g1[0] -= 1; break;
278 case 3 : g1[0] -= 2; break;
285 // This interpolation is done to produce smoother background noise.
286 slope = 0.5*(qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
287 for(i=1; i<=subframes_count; i++)
288 gain[i-1] = q->last_codebook_gain + slope * i;
290 q->last_codebook_gain = gain[i-2];
291 q->prev_g1[0] = q->prev_g1[1];
292 q->prev_g1[1] = g1[0];
297 * If the received packet is Rate 1/4 a further sanity check is made of the
300 * @param cbgain the unpacked cbgain array
301 * @return -1 if the sanity check fails, 0 otherwise
303 * TIA/EIA/IS-733 2.4.8.7.3
305 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
307 int i, diff, prev_diff=0;
311 diff = cbgain[i] - cbgain[i-1];
314 else if(FFABS(diff - prev_diff) > 12)
322 * Computes the scaled codebook vector Cdn From INDEX and GAIN
325 * The specification lacks some information here.
327 * TIA/EIA/IS-733 has an omission on the codebook index determination
328 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
329 * you have to subtract the decoded index parameter from the given scaled
330 * codebook vector index 'n' to get the desired circular codebook index, but
331 * it does not mention that you have to clamp 'n' to [0-9] in order to get
332 * RI-compliant results.
334 * The reason for this mistake seems to be the fact they forgot to mention you
335 * have to do these calculations per codebook subframe and adjust given
336 * equation values accordingly.
338 * @param q the context
339 * @param gain array holding the 4 pitch subframe gain values
340 * @param cdn_vector array for the generated scaled codebook vector
342 static void compute_svector(QCELPContext *q, const float *gain,
346 uint16_t cbseed, cindex;
347 float *rnd, tmp_gain, fir_filter_value;
354 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
355 cindex = -q->frame.cindex[i];
357 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
363 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
364 cindex = -q->frame.cindex[i];
365 for (j = 0; j < 40; j++)
366 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
370 cbseed = (0x0003 & q->frame.lspv[4])<<14 |
371 (0x003F & q->frame.lspv[3])<< 8 |
372 (0x0060 & q->frame.lspv[2])<< 1 |
373 (0x0007 & q->frame.lspv[1])<< 3 |
374 (0x0038 & q->frame.lspv[0])>> 3 ;
375 rnd = q->rnd_fir_filter_mem + 20;
378 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
381 cbseed = 521 * cbseed + 259;
382 *rnd = (int16_t)cbseed;
385 fir_filter_value = 0.0;
387 fir_filter_value += qcelp_rnd_fir_coefs[j ]
388 * (rnd[-j ] + rnd[-20+j]);
390 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
391 *cdn_vector++ = tmp_gain * fir_filter_value;
395 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 20 * sizeof(float));
398 cbseed = q->first16bits;
401 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
404 cbseed = 521 * cbseed + 259;
405 *cdn_vector++ = tmp_gain * (int16_t)cbseed;
410 cbseed = -44; // random codebook index
413 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
415 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
419 memset(cdn_vector, 0, 160 * sizeof(float));
425 * Apply generic gain control.
427 * @param v_out output vector
428 * @param v_in gain-controlled vector
429 * @param v_ref vector to control gain of
431 * FIXME: If v_ref is a zero vector, it energy is zero
432 * and the behavior of the gain control is
433 * undefined in the specs.
435 * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
437 static void apply_gain_ctrl(float *v_out, const float *v_ref,
443 for(i=0, j=0; i<4; i++)
445 scalefactor = ff_dot_productf(v_in + j, v_in + j, 40);
447 scalefactor = sqrt(ff_dot_productf(v_ref + j, v_ref + j, 40)
450 ff_log_missing_feature(NULL, "Zero energy for gain control", 1);
451 for(len=j+40; j<len; j++)
452 v_out[j] = scalefactor * v_in[j];
457 * Apply filter in pitch-subframe steps.
459 * @param memory buffer for the previous state of the filter
460 * - must be able to contain 303 elements
461 * - the 143 first elements are from the previous state
462 * - the next 160 are for output
463 * @param v_in input filter vector
464 * @param gain per-subframe gain array, each element is between 0.0 and 2.0
465 * @param lag per-subframe lag array, each element is
466 * - between 16 and 143 if its corresponding pfrac is 0,
467 * - between 16 and 139 otherwise
468 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
471 * @return filter output vector
473 static const float *do_pitchfilter(float memory[303], const float v_in[160],
474 const float gain[4], const uint8_t *lag,
475 const uint8_t pfrac[4])
478 float *v_lag, *v_out;
481 v_out = memory + 143; // Output vector starts at memory[143].
487 v_lag = memory + 143 + 40 * i - lag[i];
488 for(v_len=v_in+40; v_in<v_len; v_in++)
490 if(pfrac[i]) // If it is a fractional lag...
492 for(j=0, *v_out=0.; j<4; j++)
493 *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
497 *v_out = *v_in + gain[i] * *v_out;
504 memcpy(v_out, v_in, 40 * sizeof(float));
510 memmove(memory, memory + 160, 143 * sizeof(float));
515 * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
516 * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2
518 * @param q the context
519 * @param cdn_vector the scaled codebook vector
521 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
524 const float *v_synthesis_filtered, *v_pre_filtered;
526 if(q->bitrate >= RATE_HALF ||
527 q->bitrate == SILENCE ||
528 (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF)))
531 if(q->bitrate >= RATE_HALF)
534 // Compute gain & lag for the whole frame.
537 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
539 q->pitch_lag[i] = q->frame.plag[i] + 16;
543 float max_pitch_gain;
545 if (q->bitrate == I_F_Q)
547 if (q->erasure_count < 3)
548 max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1);
550 max_pitch_gain = 0.0;
553 assert(q->bitrate == SILENCE);
554 max_pitch_gain = 1.0;
557 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
559 memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
562 // pitch synthesis filter
563 v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
564 cdn_vector, q->pitch_gain,
565 q->pitch_lag, q->frame.pfrac);
567 // pitch prefilter update
569 q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
571 v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
572 v_synthesis_filtered,
573 q->pitch_gain, q->pitch_lag,
576 apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
579 memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
580 143 * sizeof(float));
581 memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
582 memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
583 memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
588 * Interpolates LSP frequencies and computes LPC coefficients
589 * for a given bitrate & pitch subframe.
591 * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2
593 * @param q the context
594 * @param curr_lspf LSP frequencies vector of the current frame
595 * @param lpc float vector for the resulting LPC
596 * @param subframe_num frame number in decoded stream
598 void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
599 const int subframe_num)
601 float interpolated_lspf[10];
604 if(q->bitrate >= RATE_QUARTER)
605 weight = 0.25 * (subframe_num + 1);
606 else if(q->bitrate == RATE_OCTAVE && !subframe_num)
613 weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
614 weight, 1.0 - weight, 10);
615 ff_qcelp_lspf2lpc(interpolated_lspf, lpc);
616 }else if(q->bitrate >= RATE_QUARTER ||
617 (q->bitrate == I_F_Q && !subframe_num))
618 ff_qcelp_lspf2lpc(curr_lspf, lpc);
619 else if(q->bitrate == SILENCE && !subframe_num)
620 ff_qcelp_lspf2lpc(q->prev_lspf, lpc);
623 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
627 case 35: return RATE_FULL;
628 case 17: return RATE_HALF;
629 case 8: return RATE_QUARTER;
630 case 4: return RATE_OCTAVE;
631 case 1: return SILENCE;
638 * Determine the bitrate from the frame size and/or the first byte of the frame.
640 * @param avctx the AV codec context
641 * @param buf_size length of the buffer
642 * @param buf the bufffer
644 * @return the bitrate on success,
645 * I_F_Q if the bitrate cannot be satisfactorily determined
647 * TIA/EIA/IS-733 2.4.8.7.1
649 static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, const int buf_size,
652 qcelp_packet_rate bitrate;
654 if((bitrate = buf_size2bitrate(buf_size)) >= 0)
658 QCELPContext *q = avctx->priv_data;
659 if (!q->warned_buf_mismatch_bitrate)
661 av_log(avctx, AV_LOG_WARNING,
662 "Claimed bitrate and buffer size mismatch.\n");
663 q->warned_buf_mismatch_bitrate = 1;
666 }else if(bitrate < **buf)
668 av_log(avctx, AV_LOG_ERROR,
669 "Buffer is too small for the claimed bitrate.\n");
673 }else if((bitrate = buf_size2bitrate(buf_size + 1)) >= 0)
675 av_log(avctx, AV_LOG_WARNING,
676 "Bitrate byte is missing, guessing the bitrate from packet size.\n");
680 if(bitrate == SILENCE)
682 //FIXME: Remove experimental warning when tested with samples.
683 ff_log_ask_for_sample(avctx, "'Blank frame handling is experimental.");
688 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
691 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
695 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
696 const uint8_t *buf, int buf_size)
698 QCELPContext *q = avctx->priv_data;
699 float *outbuffer = data;
701 float quantized_lspf[10], lpc[10];
705 if((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q)
707 warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
711 if(q->bitrate == RATE_OCTAVE &&
712 (q->first16bits = AV_RB16(buf)) == 0xFFFF)
714 warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
718 if(q->bitrate > SILENCE)
720 const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
721 const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
722 + qcelp_unpacking_bitmaps_lengths[q->bitrate];
723 uint8_t *unpacked_data = (uint8_t *)&q->frame;
725 init_get_bits(&q->gb, buf, 8*buf_size);
727 memset(&q->frame, 0, sizeof(QCELPFrame));
729 for(; bitmaps < bitmaps_end; bitmaps++)
730 unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
732 // Check for erasures/blanks on rates 1, 1/4 and 1/8.
733 if(q->frame.reserved)
735 warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
738 if(q->bitrate == RATE_QUARTER &&
739 codebook_sanity_check_for_rate_quarter(q->frame.cbgain))
741 warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
745 if(q->bitrate >= RATE_HALF)
749 if(q->frame.pfrac[i] && q->frame.plag[i] >= 124)
751 warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
758 decode_gain_and_index(q, gain);
759 compute_svector(q, gain, outbuffer);
761 if(decode_lspf(q, quantized_lspf) < 0)
763 warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
768 apply_pitch_filters(q, outbuffer);
770 if(q->bitrate == I_F_Q)
775 decode_gain_and_index(q, gain);
776 compute_svector(q, gain, outbuffer);
777 decode_lspf(q, quantized_lspf);
778 apply_pitch_filters(q, outbuffer);
780 q->erasure_count = 0;
782 formant_mem = q->formant_mem + 10;
785 interpolate_lpc(q, quantized_lspf, lpc, i);
786 ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40,
790 memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
792 // FIXME: postfilter and final gain control should be here.
793 // TIA/EIA/IS-733 2.4.8.6
795 formant_mem = q->formant_mem + 10;
797 *outbuffer++ = av_clipf(*formant_mem++, QCELP_CLIP_LOWER_BOUND,
798 QCELP_CLIP_UPPER_BOUND);
800 memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
801 q->prev_bitrate = q->bitrate;
803 *data_size = 160 * sizeof(*outbuffer);
808 AVCodec qcelp_decoder =
811 .type = CODEC_TYPE_AUDIO,
812 .id = CODEC_ID_QCELP,
813 .init = qcelp_decode_init,
814 .decode = qcelp_decode_frame,
815 .priv_data_size = sizeof(QCELPContext),
816 .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),