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
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;
77 * Reconstructs LPC coefficients from the line spectral pair frequencies.
79 * TIA/EIA/IS-733 2.4.3.3.5
81 void ff_qcelp_lspf2lpc(const float *lspf, float *lpc);
83 static void weighted_vector_sumf(float *out, const float *in_a,
84 const float *in_b, float weight_coeff_a,
85 float weight_coeff_b, int length)
89 for(i=0; i<length; i++)
90 out[i] = weight_coeff_a * in_a[i]
91 + weight_coeff_b * in_b[i];
95 * Initialize the speech codec according to the specification.
97 * TIA/EIA/IS-733 2.4.9
99 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
101 QCELPContext *q = avctx->priv_data;
104 avctx->sample_fmt = SAMPLE_FMT_FLT;
107 q->prev_lspf[i] = (i+1)/11.;
113 * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
114 * transmission codes of any bitrate and checks for badly received packets.
116 * @param q the context
117 * @param lspf line spectral pair frequencies
119 * @return 0 on success, -1 if the packet is badly received
121 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
123 static int decode_lspf(QCELPContext *q, float *lspf)
126 float tmp_lspf, smooth, erasure_coeff;
127 const float *predictors;
129 if(q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q)
131 predictors = (q->prev_bitrate != RATE_OCTAVE &&
132 q->prev_bitrate != I_F_Q ?
133 q->prev_lspf : q->predictor_lspf);
135 if(q->bitrate == RATE_OCTAVE)
141 q->predictor_lspf[i] =
142 lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR
143 : -QCELP_LSP_SPREAD_FACTOR)
144 + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
145 + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
147 smooth = (q->octave_count < 10 ? .875 : 0.1);
150 erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
152 assert(q->bitrate == I_F_Q);
154 if(q->erasure_count > 1)
155 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
159 q->predictor_lspf[i] =
160 lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
161 + erasure_coeff * predictors[i];
166 // Check the stability of the LSP frequencies.
167 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
169 lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
171 lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
173 lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
175 // Low-pass filter the LSP frequencies.
176 weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
184 lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001;
185 lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001;
188 // Check for badly received packets.
189 if(q->bitrate == RATE_QUARTER)
191 if(lspf[9] <= .70 || lspf[9] >= .97)
194 if(fabs(lspf[i] - lspf[i-2]) < .08)
198 if(lspf[9] <= .66 || lspf[9] >= .985)
201 if (fabs(lspf[i] - lspf[i-4]) < .0931)
209 * Converts codebook transmission codes to GAIN and INDEX.
211 * @param q the context
212 * @param gain array holding the decoded gain
214 * TIA/EIA/IS-733 2.4.6.2
216 static void decode_gain_and_index(QCELPContext *q,
218 int i, subframes_count, g1[16];
221 if(q->bitrate >= RATE_QUARTER)
225 case RATE_FULL: subframes_count = 16; break;
226 case RATE_HALF: subframes_count = 4; break;
227 default: subframes_count = 5;
229 for(i=0; i<subframes_count; i++)
231 g1[i] = 4 * q->frame.cbgain[i];
232 if(q->bitrate == RATE_FULL && !((i+1) & 3))
234 g1[i] += av_clip((g1[i-1] + g1[i-2] + g1[i-3]) / 3 - 6, 0, 32);
237 gain[i] = qcelp_g12ga[g1[i]];
239 if(q->frame.cbsign[i])
242 q->frame.cindex[i] = (q->frame.cindex[i]-89) & 127;
246 q->prev_g1[0] = g1[i-2];
247 q->prev_g1[1] = g1[i-1];
248 q->last_codebook_gain = qcelp_g12ga[g1[i-1]];
250 if(q->bitrate == RATE_QUARTER)
252 // Provide smoothing of the unvoiced excitation energy.
254 gain[6] = 0.4*gain[3] + 0.6*gain[4];
256 gain[4] = 0.8*gain[2] + 0.2*gain[3];
257 gain[3] = 0.2*gain[1] + 0.8*gain[2];
259 gain[1] = 0.6*gain[0] + 0.4*gain[1];
263 if(q->bitrate == RATE_OCTAVE)
265 g1[0] = 2 * q->frame.cbgain[0]
266 + av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54);
270 assert(q->bitrate == I_F_Q);
272 g1[0] = q->prev_g1[1];
273 switch(q->erasure_count)
276 case 2 : g1[0] -= 1; break;
277 case 3 : g1[0] -= 2; break;
284 // This interpolation is done to produce smoother background noise.
285 slope = 0.5*(qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count;
286 for(i=1; i<=subframes_count; i++)
287 gain[i-1] = q->last_codebook_gain + slope * i;
289 q->last_codebook_gain = gain[i-2];
290 q->prev_g1[0] = q->prev_g1[1];
291 q->prev_g1[1] = g1[0];
296 * If the received packet is Rate 1/4 a further sanity check is made of the
299 * @param cbgain the unpacked cbgain array
300 * @return -1 if the sanity check fails, 0 otherwise
302 * TIA/EIA/IS-733 2.4.8.7.3
304 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
306 int i, diff, prev_diff=0;
310 diff = cbgain[i] - cbgain[i-1];
313 else if(FFABS(diff - prev_diff) > 12)
321 * Computes the scaled codebook vector Cdn From INDEX and GAIN
324 * The specification lacks some information here.
326 * TIA/EIA/IS-733 has an omission on the codebook index determination
327 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
328 * you have to subtract the decoded index parameter from the given scaled
329 * codebook vector index 'n' to get the desired circular codebook index, but
330 * it does not mention that you have to clamp 'n' to [0-9] in order to get
331 * RI-compliant results.
333 * The reason for this mistake seems to be the fact they forgot to mention you
334 * have to do these calculations per codebook subframe and adjust given
335 * equation values accordingly.
337 * @param q the context
338 * @param gain array holding the 4 pitch subframe gain values
339 * @param cdn_vector array for the generated scaled codebook vector
341 static void compute_svector(QCELPContext *q, const float *gain,
345 uint16_t cbseed, cindex;
346 float *rnd, tmp_gain, fir_filter_value;
353 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
354 cindex = -q->frame.cindex[i];
356 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
362 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
363 cindex = -q->frame.cindex[i];
364 for (j = 0; j < 40; j++)
365 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
369 cbseed = (0x0003 & q->frame.lspv[4])<<14 |
370 (0x003F & q->frame.lspv[3])<< 8 |
371 (0x0060 & q->frame.lspv[2])<< 1 |
372 (0x0007 & q->frame.lspv[1])<< 3 |
373 (0x0038 & q->frame.lspv[0])>> 3 ;
374 rnd = q->rnd_fir_filter_mem + 20;
377 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
380 cbseed = 521 * cbseed + 259;
381 *rnd = (int16_t)cbseed;
384 fir_filter_value = 0.0;
386 fir_filter_value += qcelp_rnd_fir_coefs[j ]
387 * (rnd[-j ] + rnd[-20+j]);
389 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
390 *cdn_vector++ = tmp_gain * fir_filter_value;
394 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 20 * sizeof(float));
397 cbseed = q->first16bits;
400 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
403 cbseed = 521 * cbseed + 259;
404 *cdn_vector++ = tmp_gain * (int16_t)cbseed;
409 cbseed = -44; // random codebook index
412 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
414 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
421 * Apply generic gain control.
423 * @param v_out output vector
424 * @param v_in gain-controlled vector
425 * @param v_ref vector to control gain of
427 * FIXME: If v_ref is a zero vector, it energy is zero
428 * and the behavior of the gain control is
429 * undefined in the specs.
431 * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
433 static void apply_gain_ctrl(float *v_out, const float *v_ref,
439 for(i=0, j=0; i<4; i++)
441 scalefactor = ff_dot_productf(v_in + j, v_in + j, 40);
443 scalefactor = sqrt(ff_dot_productf(v_ref + j, v_ref + j, 40)
446 ff_log_missing_feature(NULL, "Zero energy for gain control", 1);
447 for(len=j+40; j<len; j++)
448 v_out[j] = scalefactor * v_in[j];
453 * Apply filter in pitch-subframe steps.
455 * @param memory buffer for the previous state of the filter
456 * - must be able to contain 303 elements
457 * - the 143 first elements are from the previous state
458 * - the next 160 are for output
459 * @param v_in input filter vector
460 * @param gain per-subframe gain array, each element is between 0.0 and 2.0
461 * @param lag per-subframe lag array, each element is
462 * - between 16 and 143 if its corresponding pfrac is 0,
463 * - between 16 and 139 otherwise
464 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
467 * @return filter output vector
469 static const float *do_pitchfilter(float memory[303], const float v_in[160],
470 const float gain[4], const uint8_t *lag,
471 const uint8_t pfrac[4])
474 float *v_lag, *v_out;
477 v_out = memory + 143; // Output vector starts at memory[143].
483 v_lag = memory + 143 + 40 * i - lag[i];
484 for(v_len=v_in+40; v_in<v_len; v_in++)
486 if(pfrac[i]) // If it is a fractional lag...
488 for(j=0, *v_out=0.; j<4; j++)
489 *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
493 *v_out = *v_in + gain[i] * *v_out;
500 memcpy(v_out, v_in, 40 * sizeof(float));
506 memmove(memory, memory + 160, 143 * sizeof(float));
511 * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector.
512 * TIA/EIA/IS-733 2.4.5.2
514 * @param q the context
515 * @param cdn_vector the scaled codebook vector
517 static void apply_pitch_filters(QCELPContext *q, float *cdn_vector)
520 const float *v_synthesis_filtered, *v_pre_filtered;
522 if(q->bitrate >= RATE_HALF ||
523 (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF)))
526 if(q->bitrate >= RATE_HALF)
529 // Compute gain & lag for the whole frame.
532 q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0;
534 q->pitch_lag[i] = q->frame.plag[i] + 16;
538 float max_pitch_gain = q->erasure_count < 3 ? 0.9 - 0.3 * (q->erasure_count - 1) : 0.0;
540 q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain);
542 memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac));
545 // pitch synthesis filter
546 v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem,
547 cdn_vector, q->pitch_gain,
548 q->pitch_lag, q->frame.pfrac);
550 // pitch prefilter update
552 q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0);
554 v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem,
555 v_synthesis_filtered,
556 q->pitch_gain, q->pitch_lag,
559 apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered);
562 memcpy(q->pitch_synthesis_filter_mem, cdn_vector + 17,
563 143 * sizeof(float));
564 memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float));
565 memset(q->pitch_gain, 0, sizeof(q->pitch_gain));
566 memset(q->pitch_lag, 0, sizeof(q->pitch_lag));
571 * Interpolates LSP frequencies and computes LPC coefficients
572 * for a given bitrate & pitch subframe.
574 * TIA/EIA/IS-733 2.4.3.3.4
576 * @param q the context
577 * @param curr_lspf LSP frequencies vector of the current frame
578 * @param lpc float vector for the resulting LPC
579 * @param subframe_num frame number in decoded stream
581 void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
582 const int subframe_num)
584 float interpolated_lspf[10];
587 if(q->bitrate >= RATE_QUARTER)
588 weight = 0.25 * (subframe_num + 1);
589 else if(q->bitrate == RATE_OCTAVE && !subframe_num)
596 weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
597 weight, 1.0 - weight, 10);
598 ff_qcelp_lspf2lpc(interpolated_lspf, lpc);
599 }else if(q->bitrate >= RATE_QUARTER ||
600 (q->bitrate == I_F_Q && !subframe_num))
601 ff_qcelp_lspf2lpc(curr_lspf, lpc);
604 static qcelp_packet_rate buf_size2bitrate(const int buf_size)
608 case 35: return RATE_FULL;
609 case 17: return RATE_HALF;
610 case 8: return RATE_QUARTER;
611 case 4: return RATE_OCTAVE;
612 case 1: return SILENCE;
619 * Determine the bitrate from the frame size and/or the first byte of the frame.
621 * @param avctx the AV codec context
622 * @param buf_size length of the buffer
623 * @param buf the bufffer
625 * @return the bitrate on success,
626 * I_F_Q if the bitrate cannot be satisfactorily determined
628 * TIA/EIA/IS-733 2.4.8.7.1
630 static int determine_bitrate(AVCodecContext *avctx, const int buf_size,
633 qcelp_packet_rate bitrate;
635 if((bitrate = buf_size2bitrate(buf_size)) >= 0)
639 av_log(avctx, AV_LOG_WARNING,
640 "Claimed bitrate and buffer size mismatch.\n");
642 }else if(bitrate < **buf)
644 av_log(avctx, AV_LOG_ERROR,
645 "Buffer is too small for the claimed bitrate.\n");
649 }else if((bitrate = buf_size2bitrate(buf_size + 1)) >= 0)
651 av_log(avctx, AV_LOG_WARNING,
652 "Bitrate byte is missing, guessing the bitrate from packet size.\n");
656 if(bitrate == SILENCE)
658 // FIXME: the decoder should not handle SILENCE frames as I_F_Q frames
659 ff_log_missing_feature(avctx, "Blank frame", 1);
665 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
668 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
672 static int qcelp_decode_frame(AVCodecContext *avctx, void *data, int *data_size,
673 const uint8_t *buf, int buf_size)
675 QCELPContext *q = avctx->priv_data;
676 float *outbuffer = data;
678 float quantized_lspf[10], lpc[10];
682 if((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q)
684 warn_insufficient_frame_quality(avctx, "bitrate cannot be determined.");
688 if(q->bitrate == RATE_OCTAVE &&
689 (q->first16bits = AV_RB16(buf)) == 0xFFFF)
691 warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on.");
695 if(q->bitrate > SILENCE)
697 const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate];
698 const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate]
699 + qcelp_unpacking_bitmaps_lengths[q->bitrate];
700 uint8_t *unpacked_data = (uint8_t *)&q->frame;
702 init_get_bits(&q->gb, buf, 8*buf_size);
704 memset(&q->frame, 0, sizeof(QCELPFrame));
706 for(; bitmaps < bitmaps_end; bitmaps++)
707 unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos;
709 // Check for erasures/blanks on rates 1, 1/4 and 1/8.
710 if(q->frame.reserved)
712 warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area.");
715 if(q->bitrate == RATE_QUARTER &&
716 codebook_sanity_check_for_rate_quarter(q->frame.cbgain))
718 warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed.");
722 if(q->bitrate >= RATE_HALF)
726 if(q->frame.pfrac[i] && q->frame.plag[i] >= 124)
728 warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter.");
735 decode_gain_and_index(q, gain);
736 compute_svector(q, gain, outbuffer);
738 if(decode_lspf(q, quantized_lspf) < 0)
740 warn_insufficient_frame_quality(avctx, "Badly received packets in frame.");
745 apply_pitch_filters(q, outbuffer);
747 if(q->bitrate == I_F_Q)
752 decode_gain_and_index(q, gain);
753 compute_svector(q, gain, outbuffer);
754 decode_lspf(q, quantized_lspf);
755 apply_pitch_filters(q, outbuffer);
757 q->erasure_count = 0;
759 formant_mem = q->formant_mem + 10;
762 interpolate_lpc(q, quantized_lspf, lpc, i);
763 ff_celp_lp_synthesis_filterf(formant_mem, lpc, outbuffer + i * 40, 40,
767 memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float));
769 // FIXME: postfilter and final gain control should be here.
770 // TIA/EIA/IS-733 2.4.8.6
772 formant_mem = q->formant_mem + 10;
774 *outbuffer++ = av_clipf(*formant_mem++, QCELP_CLIP_LOWER_BOUND,
775 QCELP_CLIP_UPPER_BOUND);
777 memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf));
778 q->prev_bitrate = q->bitrate;
780 *data_size = 160 * sizeof(*outbuffer);
785 AVCodec qcelp_decoder =
788 .type = CODEC_TYPE_AUDIO,
789 .id = CODEC_ID_QCELP,
790 .init = qcelp_decode_init,
791 .decode = qcelp_decode_frame,
792 .priv_data_size = sizeof(QCELPContext),
793 .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),