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
32 #include "bitstream.h"
35 #include "qcelpdata.h"
37 #include "celp_math.h"
38 #include "celp_filters.h"
43 static void weighted_vector_sumf(float *out, const float *in_a,
44 const float *in_b, float weight_coeff_a,
45 float weight_coeff_b, int length)
49 for(i=0; i<length; i++)
50 out[i] = weight_coeff_a * in_a[i]
51 + weight_coeff_b * in_b[i];
55 * Initialize the speech codec according to the specification.
57 * TIA/EIA/IS-733 2.4.9
59 static av_cold int qcelp_decode_init(AVCodecContext *avctx)
61 QCELPContext *q = avctx->priv_data;
64 avctx->sample_fmt = SAMPLE_FMT_FLT;
66 for (i = 0; i < 10; i++)
67 q->prev_lspf[i] = (i + 1) / 11.;
73 * Decodes the 10 quantized LSP frequencies from the LSPV/LSP
74 * transmission codes of any bitrate and checks for badly received packets.
76 * @param q the context
77 * @param lspf line spectral pair frequencies
79 * @return 0 on success, -1 if the packet is badly received
81 * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3
83 static int decode_lspf(QCELPContext *q, float *lspf)
88 if(q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q)
91 const float *predictors = (q->prev_bitrate != RATE_OCTAVE &&
92 q->prev_bitrate != I_F_Q ? q->prev_lspf
95 if(q->bitrate == RATE_OCTAVE)
101 q->predictor_lspf[i] =
102 lspf[i] = (q->lspv[i] ? QCELP_LSP_SPREAD_FACTOR
103 : -QCELP_LSP_SPREAD_FACTOR)
104 + predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR
105 + (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR)/11);
107 smooth = (q->octave_count < 10 ? .875 : 0.1);
110 float erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR;
112 assert(q->bitrate == I_F_Q);
114 if(q->erasure_count > 1)
115 erasure_coeff *= (q->erasure_count < 4 ? 0.9 : 0.7);
119 q->predictor_lspf[i] =
120 lspf[i] = (i + 1) * ( 1 - erasure_coeff)/11
121 + erasure_coeff * predictors[i];
126 // Check the stability of the LSP frequencies.
127 lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR);
129 lspf[i] = FFMAX(lspf[i], (lspf[i-1] + QCELP_LSP_SPREAD_FACTOR));
131 lspf[9] = FFMIN(lspf[9], (1.0 - QCELP_LSP_SPREAD_FACTOR));
133 lspf[i-1] = FFMIN(lspf[i-1], (lspf[i] - QCELP_LSP_SPREAD_FACTOR));
135 // Low-pass filter the LSP frequencies.
136 weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0-smooth, 10);
144 lspf[2*i+0] = tmp_lspf += qcelp_lspvq[i][q->lspv[i]][0] * 0.0001;
145 lspf[2*i+1] = tmp_lspf += qcelp_lspvq[i][q->lspv[i]][1] * 0.0001;
148 // Check for badly received packets.
149 if(q->bitrate == RATE_QUARTER)
151 if(lspf[9] <= .70 || lspf[9] >= .97)
154 if(fabs(lspf[i] - lspf[i-2]) < .08)
158 if(lspf[9] <= .66 || lspf[9] >= .985)
161 if (fabs(lspf[i] - lspf[i-4]) < .0931)
169 * If the received packet is Rate 1/4 a further sanity check is made of the
172 * @param cbgain the unpacked cbgain array
173 * @return -1 if the sanity check fails, 0 otherwise
175 * TIA/EIA/IS-733 2.4.8.7.3
177 static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain)
183 int diff = cbgain[i] - cbgain[i-1];
186 else if(FFABS(diff - prev_diff) > 12)
194 * Computes the scaled codebook vector Cdn From INDEX and GAIN
197 * The specification lacks some information here.
199 * TIA/EIA/IS-733 has an omission on the codebook index determination
200 * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says
201 * you have to subtract the decoded index parameter from the given scaled
202 * codebook vector index 'n' to get the desired circular codebook index, but
203 * it does not mention that you have to clamp 'n' to [0-9] in order to get
204 * RI-compliant results.
206 * The reason for this mistake seems to be the fact they forgot to mention you
207 * have to do these calculations per codebook subframe and adjust given
208 * equation values accordingly.
210 * @param q the context
211 * @param gain array holding the 4 pitch subframe gain values
212 * @param cdn_vector array for the generated scaled codebook vector
214 static void compute_svector(const QCELPContext *q, const float *gain,
218 uint16_t cbseed, cindex;
219 float *rnd, tmp_gain, fir_filter_value;
226 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
227 cindex = -q->cindex[i];
229 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cindex++ & 127];
235 tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO;
236 cindex = -q->cindex[i];
237 for (j = 0; j < 40; j++)
238 *cdn_vector++ = tmp_gain * qcelp_rate_half_codebook[cindex++ & 127];
242 cbseed = (0x0003 & q->lspv[4])<<14 |
243 (0x003F & q->lspv[3])<< 8 |
244 (0x0060 & q->lspv[2])<< 1 |
245 (0x0007 & q->lspv[1])<< 3 |
246 (0x0038 & q->lspv[0])>> 3 ;
247 rnd = q->rnd_fir_filter_mem + 20;
250 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
253 cbseed = 521 * cbseed + 259;
254 *rnd = (int16_t)cbseed;
257 fir_filter_value = 0.0;
259 fir_filter_value += qcelp_rnd_fir_coefs[j ]
260 * (rnd[-j ] + rnd[-20+j]);
262 fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10];
263 *cdn_vector++ = tmp_gain * fir_filter_value;
267 memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, 20 * sizeof(float));
270 cbseed = q->first16bits;
273 tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0);
276 cbseed = 521 * cbseed + 259;
277 *cdn_vector++ = tmp_gain * (int16_t)cbseed;
282 cbseed = -44; // random codebook index
285 tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO;
287 *cdn_vector++ = tmp_gain * qcelp_rate_full_codebook[cbseed++ & 127];
294 * Apply generic gain control.
296 * @param v_out output vector
297 * @param v_in gain-controlled vector
298 * @param v_ref vector to control gain of
300 * FIXME: If v_ref is a zero vector, it energy is zero
301 * and the behavior of the gain control is
302 * undefined in the specs.
304 * TIA/EIA/IS-733 2.4.8.3-2/3/4/5, 2.4.8.6
306 static void apply_gain_ctrl(float *v_out, const float *v_ref,
312 for(i=0, j=0; i<4; i++)
314 scalefactor = ff_dot_productf(v_in + j, v_in + j, 40);
316 scalefactor = sqrt(ff_dot_productf(v_ref + j, v_ref + j, 40)
319 av_log_missing_feature(NULL, "Zero energy for gain control", 1);
320 for(len=j+40; j<len; j++)
321 v_out[j] = scalefactor * v_in[j];
326 * Apply filter in pitch-subframe steps.
328 * @param memory buffer for the previous state of the filter
329 * - must be able to contain 303 elements
330 * - the 143 first elements are from the previous state
331 * - the next 160 are for output
332 * @param v_in input filter vector
333 * @param gain per-subframe gain array, each element is between 0.0 and 2.0
334 * @param lag per-subframe lag array, each element is
335 * - between 16 and 143 if its corresponding pfrac is 0,
336 * - between 16 and 139 otherwise
337 * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0
340 * @return filter output vector
342 static const float *do_pitchfilter(float memory[303], const float v_in[160],
343 const float gain[4], const uint8_t *lag,
344 const uint8_t pfrac[4])
347 float *v_lag, *v_out;
350 v_out = memory + 143; // Output vector starts at memory[143].
356 v_lag = memory + 143 + 40 * i - lag[i];
357 for(v_len=v_in+40; v_in<v_len; v_in++)
359 if(pfrac[i]) // If it is a fractional lag...
361 for(j=0, *v_out=0.; j<4; j++)
362 *v_out += qcelp_hammsinc_table[j] * (v_lag[j-4] + v_lag[3-j]);
366 *v_out = *v_in + gain[i] * *v_out;
373 memcpy(v_out, v_in, 40 * sizeof(float));
379 memmove(memory, memory + 160, 143 * sizeof(float));
384 * Interpolates LSP frequencies and computes LPC coefficients
385 * for a given bitrate & pitch subframe.
387 * TIA/EIA/IS-733 2.4.3.3.4
389 * @param q the context
390 * @param curr_lspf LSP frequencies vector of the current frame
391 * @param lpc float vector for the resulting LPC
392 * @param subframe_num frame number in decoded stream
394 void interpolate_lpc(QCELPContext *q, const float *curr_lspf, float *lpc,
395 const int subframe_num)
397 float interpolated_lspf[10];
400 if(q->bitrate >= RATE_QUARTER)
401 weight = 0.25 * (subframe_num + 1);
402 else if(q->bitrate == RATE_OCTAVE && !subframe_num)
409 weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf,
410 weight, 1.0 - weight, 10);
411 qcelp_lspf2lpc(interpolated_lspf, lpc);
412 }else if(q->bitrate >= RATE_QUARTER || (q->bitrate == I_F_Q && !subframe_num))
413 qcelp_lspf2lpc(curr_lspf, lpc);
416 static int buf_size2bitrate(const int buf_size)
435 static void warn_insufficient_frame_quality(AVCodecContext *avctx,
438 av_log(avctx, AV_LOG_WARNING, "Frame #%d, IFQ: %s\n", avctx->frame_number,
442 AVCodec qcelp_decoder =
445 .type = CODEC_TYPE_AUDIO,
446 .id = CODEC_ID_QCELP,
447 .init = qcelp_decode_init,
448 .decode = qcelp_decode_frame,
449 .priv_data_size = sizeof(QCELPContext),
450 .long_name = NULL_IF_CONFIG_SMALL("QCELP / PureVoice"),