/src/ffmpeg/libavcodec/qcelpdec.c
Line | Count | Source |
1 | | /* |
2 | | * QCELP decoder |
3 | | * Copyright (c) 2007 Reynaldo H. Verdejo Pinochet |
4 | | * |
5 | | * This file is part of FFmpeg. |
6 | | * |
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. |
11 | | * |
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. |
16 | | * |
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 |
20 | | */ |
21 | | |
22 | | /** |
23 | | * @file |
24 | | * QCELP decoder |
25 | | * @author Reynaldo H. Verdejo Pinochet |
26 | | * @remark FFmpeg merging spearheaded by Kenan Gillet |
27 | | * @remark Development mentored by Benjamin Larson |
28 | | */ |
29 | | |
30 | | #include "libavutil/avassert.h" |
31 | | #include "libavutil/channel_layout.h" |
32 | | #include "libavutil/float_dsp.h" |
33 | | #include "avcodec.h" |
34 | | #include "codec_internal.h" |
35 | | #include "decode.h" |
36 | | #include "get_bits.h" |
37 | | #include "qcelpdata.h" |
38 | | #include "celp_filters.h" |
39 | | #include "acelp_filters.h" |
40 | | #include "acelp_vectors.h" |
41 | | #include "lsp.h" |
42 | | |
43 | | typedef enum { |
44 | | I_F_Q = -1, /**< insufficient frame quality */ |
45 | | SILENCE, |
46 | | RATE_OCTAVE, |
47 | | RATE_QUARTER, |
48 | | RATE_HALF, |
49 | | RATE_FULL |
50 | | } qcelp_packet_rate; |
51 | | |
52 | | typedef struct QCELPContext { |
53 | | GetBitContext gb; |
54 | | qcelp_packet_rate bitrate; |
55 | | QCELPFrame frame; /**< unpacked data frame */ |
56 | | |
57 | | uint8_t erasure_count; |
58 | | uint8_t octave_count; /**< count the consecutive RATE_OCTAVE frames */ |
59 | | float prev_lspf[10]; |
60 | | float predictor_lspf[10];/**< LSP predictor for RATE_OCTAVE and I_F_Q */ |
61 | | float pitch_synthesis_filter_mem[303]; |
62 | | float pitch_pre_filter_mem[303]; |
63 | | float rnd_fir_filter_mem[180]; |
64 | | float formant_mem[170]; |
65 | | float last_codebook_gain; |
66 | | int prev_g1[2]; |
67 | | int prev_bitrate; |
68 | | float pitch_gain[4]; |
69 | | uint8_t pitch_lag[4]; |
70 | | uint16_t first16bits; |
71 | | uint8_t warned_buf_mismatch_bitrate; |
72 | | |
73 | | /* postfilter */ |
74 | | float postfilter_synth_mem[10]; |
75 | | float postfilter_agc_mem; |
76 | | float postfilter_tilt_mem; |
77 | | } QCELPContext; |
78 | | |
79 | | /** |
80 | | * Initialize the speech codec according to the specification. |
81 | | * |
82 | | * TIA/EIA/IS-733 2.4.9 |
83 | | */ |
84 | | static av_cold int qcelp_decode_init(AVCodecContext *avctx) |
85 | 1.09k | { |
86 | 1.09k | QCELPContext *q = avctx->priv_data; |
87 | 1.09k | int i; |
88 | | |
89 | 1.09k | av_channel_layout_uninit(&avctx->ch_layout); |
90 | 1.09k | avctx->ch_layout = (AVChannelLayout)AV_CHANNEL_LAYOUT_MONO; |
91 | 1.09k | avctx->sample_fmt = AV_SAMPLE_FMT_FLT; |
92 | 1.09k | if (!avctx->sample_rate) |
93 | 727 | avctx->sample_rate = 8000; |
94 | | |
95 | 12.0k | for (i = 0; i < 10; i++) |
96 | 10.9k | q->prev_lspf[i] = (i + 1) / 11.0; |
97 | | |
98 | 1.09k | return 0; |
99 | 1.09k | } |
100 | | |
101 | | /** |
102 | | * Decode the 10 quantized LSP frequencies from the LSPV/LSP |
103 | | * transmission codes of any bitrate and check for badly received packets. |
104 | | * |
105 | | * @param q the context |
106 | | * @param lspf line spectral pair frequencies |
107 | | * |
108 | | * @return 0 on success, -1 if the packet is badly received |
109 | | * |
110 | | * TIA/EIA/IS-733 2.4.3.2.6.2-2, 2.4.8.7.3 |
111 | | */ |
112 | | static int decode_lspf(QCELPContext *q, float *lspf) |
113 | 248k | { |
114 | 248k | int i; |
115 | 248k | float tmp_lspf, smooth, erasure_coeff; |
116 | 248k | const float *predictors; |
117 | | |
118 | 248k | if (q->bitrate == RATE_OCTAVE || q->bitrate == I_F_Q) { |
119 | 210k | predictors = q->prev_bitrate != RATE_OCTAVE && |
120 | 205k | q->prev_bitrate != I_F_Q ? q->prev_lspf |
121 | 210k | : q->predictor_lspf; |
122 | | |
123 | 210k | if (q->bitrate == RATE_OCTAVE) { |
124 | 4.84k | q->octave_count++; |
125 | | |
126 | 53.2k | for (i = 0; i < 10; i++) { |
127 | 48.4k | q->predictor_lspf[i] = |
128 | 48.4k | lspf[i] = (q->frame.lspv[i] ? QCELP_LSP_SPREAD_FACTOR |
129 | 48.4k | : -QCELP_LSP_SPREAD_FACTOR) + |
130 | 48.4k | predictors[i] * QCELP_LSP_OCTAVE_PREDICTOR + |
131 | 48.4k | (i + 1) * ((1 - QCELP_LSP_OCTAVE_PREDICTOR) / 11); |
132 | 48.4k | } |
133 | 4.84k | smooth = q->octave_count < 10 ? .875 : 0.1; |
134 | 205k | } else { |
135 | 205k | erasure_coeff = QCELP_LSP_OCTAVE_PREDICTOR; |
136 | | |
137 | 205k | av_assert2(q->bitrate == I_F_Q); |
138 | | |
139 | 205k | if (q->erasure_count > 1) |
140 | 200k | erasure_coeff *= q->erasure_count < 4 ? 0.9 : 0.7; |
141 | | |
142 | 2.26M | for (i = 0; i < 10; i++) { |
143 | 2.05M | q->predictor_lspf[i] = |
144 | 2.05M | lspf[i] = (i + 1) * (1 - erasure_coeff) / 11 + |
145 | 2.05M | erasure_coeff * predictors[i]; |
146 | 2.05M | } |
147 | 205k | smooth = 0.125; |
148 | 205k | } |
149 | | |
150 | | // Check the stability of the LSP frequencies. |
151 | 210k | lspf[0] = FFMAX(lspf[0], QCELP_LSP_SPREAD_FACTOR); |
152 | 2.10M | for (i = 1; i < 10; i++) |
153 | 1.89M | lspf[i] = FFMAX(lspf[i], lspf[i - 1] + QCELP_LSP_SPREAD_FACTOR); |
154 | | |
155 | 210k | lspf[9] = FFMIN(lspf[9], 1.0 - QCELP_LSP_SPREAD_FACTOR); |
156 | 2.10M | for (i = 9; i > 0; i--) |
157 | 1.89M | lspf[i - 1] = FFMIN(lspf[i - 1], lspf[i] - QCELP_LSP_SPREAD_FACTOR); |
158 | | |
159 | | // Low-pass filter the LSP frequencies. |
160 | 210k | ff_weighted_vector_sumf(lspf, lspf, q->prev_lspf, smooth, 1.0 - smooth, 10); |
161 | 210k | } else { |
162 | 37.4k | q->octave_count = 0; |
163 | | |
164 | 37.4k | tmp_lspf = 0.0; |
165 | 224k | for (i = 0; i < 5; i++) { |
166 | 187k | lspf[2 * i + 0] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][0] * 0.0001; |
167 | 187k | lspf[2 * i + 1] = tmp_lspf += qcelp_lspvq[i][q->frame.lspv[i]][1] * 0.0001; |
168 | 187k | } |
169 | | |
170 | | // Check for badly received packets. |
171 | 37.4k | if (q->bitrate == RATE_QUARTER) { |
172 | 2.33k | if (lspf[9] <= .70 || lspf[9] >= .97) |
173 | 715 | return -1; |
174 | 12.3k | for (i = 3; i < 10; i++) |
175 | 11.0k | if (fabs(lspf[i] - lspf[i - 2]) < .08) |
176 | 397 | return -1; |
177 | 35.1k | } else { |
178 | 35.1k | if (lspf[9] <= .66 || lspf[9] >= .985) |
179 | 23.4k | return -1; |
180 | 81.2k | for (i = 4; i < 10; i++) |
181 | 69.7k | if (fabs(lspf[i] - lspf[i - 4]) < .0931) |
182 | 196 | return -1; |
183 | 11.7k | } |
184 | 37.4k | } |
185 | 223k | return 0; |
186 | 248k | } |
187 | | |
188 | | /** |
189 | | * Convert codebook transmission codes to GAIN and INDEX. |
190 | | * |
191 | | * @param q the context |
192 | | * @param gain array holding the decoded gain |
193 | | * |
194 | | * TIA/EIA/IS-733 2.4.6.2 |
195 | | */ |
196 | | static void decode_gain_and_index(QCELPContext *q, float *gain) |
197 | 248k | { |
198 | 248k | int i, subframes_count, g1[16]; |
199 | 248k | float slope; |
200 | | |
201 | 248k | if (q->bitrate >= RATE_QUARTER) { |
202 | 5.44k | switch (q->bitrate) { |
203 | 1.27k | case RATE_FULL: subframes_count = 16; break; |
204 | 1.83k | case RATE_HALF: subframes_count = 4; break; |
205 | 2.33k | default: subframes_count = 5; |
206 | 5.44k | } |
207 | 44.9k | for (i = 0; i < subframes_count; i++) { |
208 | 39.4k | g1[i] = 4 * q->frame.cbgain[i]; |
209 | 39.4k | if (q->bitrate == RATE_FULL && !((i + 1) & 3)) { |
210 | 5.11k | g1[i] += av_clip((g1[i - 1] + g1[i - 2] + g1[i - 3]) / 3 - 6, 0, 32); |
211 | 5.11k | } |
212 | | |
213 | 39.4k | gain[i] = qcelp_g12ga[g1[i]]; |
214 | | |
215 | 39.4k | if (q->frame.cbsign[i]) { |
216 | 10.5k | gain[i] = -gain[i]; |
217 | 10.5k | q->frame.cindex[i] = (q->frame.cindex[i] - 89) & 127; |
218 | 10.5k | } |
219 | 39.4k | } |
220 | | |
221 | 5.44k | q->prev_g1[0] = g1[i - 2]; |
222 | 5.44k | q->prev_g1[1] = g1[i - 1]; |
223 | 5.44k | q->last_codebook_gain = qcelp_g12ga[g1[i - 1]]; |
224 | | |
225 | 5.44k | if (q->bitrate == RATE_QUARTER) { |
226 | | // Provide smoothing of the unvoiced excitation energy. |
227 | 2.33k | gain[7] = gain[4]; |
228 | 2.33k | gain[6] = 0.4 * gain[3] + 0.6 * gain[4]; |
229 | 2.33k | gain[5] = gain[3]; |
230 | 2.33k | gain[4] = 0.8 * gain[2] + 0.2 * gain[3]; |
231 | 2.33k | gain[3] = 0.2 * gain[1] + 0.8 * gain[2]; |
232 | 2.33k | gain[2] = gain[1]; |
233 | 2.33k | gain[1] = 0.6 * gain[0] + 0.4 * gain[1]; |
234 | 2.33k | } |
235 | 242k | } else if (q->bitrate != SILENCE) { |
236 | 210k | if (q->bitrate == RATE_OCTAVE) { |
237 | 4.84k | g1[0] = 2 * q->frame.cbgain[0] + |
238 | 4.84k | av_clip((q->prev_g1[0] + q->prev_g1[1]) / 2 - 5, 0, 54); |
239 | 4.84k | subframes_count = 8; |
240 | 205k | } else { |
241 | 205k | av_assert2(q->bitrate == I_F_Q); |
242 | | |
243 | 205k | g1[0] = q->prev_g1[1]; |
244 | 205k | switch (q->erasure_count) { |
245 | 4.40k | case 1 : break; |
246 | 2.54k | case 2 : g1[0] -= 1; break; |
247 | 2.03k | case 3 : g1[0] -= 2; break; |
248 | 196k | default: g1[0] -= 6; |
249 | 205k | } |
250 | 205k | if (g1[0] < 0) |
251 | 198k | g1[0] = 0; |
252 | 205k | subframes_count = 4; |
253 | 205k | } |
254 | | // This interpolation is done to produce smoother background noise. |
255 | 210k | slope = 0.5 * (qcelp_g12ga[g1[0]] - q->last_codebook_gain) / subframes_count; |
256 | 1.07M | for (i = 1; i <= subframes_count; i++) |
257 | 861k | gain[i - 1] = q->last_codebook_gain + slope * i; |
258 | | |
259 | 210k | q->last_codebook_gain = gain[i - 2]; |
260 | 210k | q->prev_g1[0] = q->prev_g1[1]; |
261 | 210k | q->prev_g1[1] = g1[0]; |
262 | 210k | } |
263 | 248k | } |
264 | | |
265 | | /** |
266 | | * If the received packet is Rate 1/4 a further sanity check is made of the |
267 | | * codebook gain. |
268 | | * |
269 | | * @param cbgain the unpacked cbgain array |
270 | | * @return -1 if the sanity check fails, 0 otherwise |
271 | | * |
272 | | * TIA/EIA/IS-733 2.4.8.7.3 |
273 | | */ |
274 | | static int codebook_sanity_check_for_rate_quarter(const uint8_t *cbgain) |
275 | 2.83k | { |
276 | 2.83k | int i, diff, prev_diff = 0; |
277 | | |
278 | 12.9k | for (i = 1; i < 5; i++) { |
279 | 10.5k | diff = cbgain[i] - cbgain[i-1]; |
280 | 10.5k | if (FFABS(diff) > 10) |
281 | 252 | return -1; |
282 | 10.3k | else if (FFABS(diff - prev_diff) > 12) |
283 | 244 | return -1; |
284 | 10.0k | prev_diff = diff; |
285 | 10.0k | } |
286 | 2.33k | return 0; |
287 | 2.83k | } |
288 | | |
289 | | /** |
290 | | * Compute the scaled codebook vector Cdn From INDEX and GAIN |
291 | | * for all rates. |
292 | | * |
293 | | * The specification lacks some information here. |
294 | | * |
295 | | * TIA/EIA/IS-733 has an omission on the codebook index determination |
296 | | * formula for RATE_FULL and RATE_HALF frames at section 2.4.8.1.1. It says |
297 | | * you have to subtract the decoded index parameter from the given scaled |
298 | | * codebook vector index 'n' to get the desired circular codebook index, but |
299 | | * it does not mention that you have to clamp 'n' to [0-9] in order to get |
300 | | * RI-compliant results. |
301 | | * |
302 | | * The reason for this mistake seems to be the fact they forgot to mention you |
303 | | * have to do these calculations per codebook subframe and adjust given |
304 | | * equation values accordingly. |
305 | | * |
306 | | * @param q the context |
307 | | * @param gain array holding the 4 pitch subframe gain values |
308 | | * @param cdn_vector array for the generated scaled codebook vector |
309 | | */ |
310 | | static void compute_svector(QCELPContext *q, const float *gain, |
311 | | float *cdn_vector) |
312 | 248k | { |
313 | 248k | int i, j, k; |
314 | 248k | uint16_t cbseed, cindex; |
315 | 248k | float *rnd, tmp_gain, fir_filter_value; |
316 | | |
317 | 248k | switch (q->bitrate) { |
318 | 1.27k | case RATE_FULL: |
319 | 21.7k | for (i = 0; i < 16; i++) { |
320 | 20.4k | tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; |
321 | 20.4k | cindex = -q->frame.cindex[i]; |
322 | 224k | for (j = 0; j < 10; j++) |
323 | 204k | *cdn_vector++ = tmp_gain * |
324 | 204k | qcelp_rate_full_codebook[cindex++ & 127]; |
325 | 20.4k | } |
326 | 1.27k | break; |
327 | 1.83k | case RATE_HALF: |
328 | 9.16k | for (i = 0; i < 4; i++) { |
329 | 7.33k | tmp_gain = gain[i] * QCELP_RATE_HALF_CODEBOOK_RATIO; |
330 | 7.33k | cindex = -q->frame.cindex[i]; |
331 | 300k | for (j = 0; j < 40; j++) |
332 | 293k | *cdn_vector++ = tmp_gain * |
333 | 293k | qcelp_rate_half_codebook[cindex++ & 127]; |
334 | 7.33k | } |
335 | 1.83k | break; |
336 | 2.33k | case RATE_QUARTER: |
337 | 2.33k | cbseed = (0x0003 & q->frame.lspv[4]) << 14 | |
338 | 2.33k | (0x003F & q->frame.lspv[3]) << 8 | |
339 | 2.33k | (0x0060 & q->frame.lspv[2]) << 1 | |
340 | 2.33k | (0x0007 & q->frame.lspv[1]) << 3 | |
341 | 2.33k | (0x0038 & q->frame.lspv[0]) >> 3; |
342 | 2.33k | rnd = q->rnd_fir_filter_mem + 20; |
343 | 21.0k | for (i = 0; i < 8; i++) { |
344 | 18.6k | tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); |
345 | 392k | for (k = 0; k < 20; k++) { |
346 | 373k | cbseed = 521 * cbseed + 259; |
347 | 373k | *rnd = (int16_t) cbseed; |
348 | | |
349 | | // FIR filter |
350 | 373k | fir_filter_value = 0.0; |
351 | 4.11M | for (j = 0; j < 10; j++) |
352 | 3.73M | fir_filter_value += qcelp_rnd_fir_coefs[j] * |
353 | 3.73M | (rnd[-j] + rnd[-20+j]); |
354 | | |
355 | 373k | fir_filter_value += qcelp_rnd_fir_coefs[10] * rnd[-10]; |
356 | 373k | *cdn_vector++ = tmp_gain * fir_filter_value; |
357 | 373k | rnd++; |
358 | 373k | } |
359 | 18.6k | } |
360 | 2.33k | memcpy(q->rnd_fir_filter_mem, q->rnd_fir_filter_mem + 160, |
361 | 2.33k | 20 * sizeof(float)); |
362 | 2.33k | break; |
363 | 4.84k | case RATE_OCTAVE: |
364 | 4.84k | cbseed = q->first16bits; |
365 | 43.5k | for (i = 0; i < 8; i++) { |
366 | 38.7k | tmp_gain = gain[i] * (QCELP_SQRT1887 / 32768.0); |
367 | 813k | for (j = 0; j < 20; j++) { |
368 | 774k | cbseed = 521 * cbseed + 259; |
369 | 774k | *cdn_vector++ = tmp_gain * (int16_t) cbseed; |
370 | 774k | } |
371 | 38.7k | } |
372 | 4.84k | break; |
373 | 205k | case I_F_Q: |
374 | 205k | cbseed = -44; // random codebook index |
375 | 1.02M | for (i = 0; i < 4; i++) { |
376 | 822k | tmp_gain = gain[i] * QCELP_RATE_FULL_CODEBOOK_RATIO; |
377 | 33.7M | for (j = 0; j < 40; j++) |
378 | 32.9M | *cdn_vector++ = tmp_gain * |
379 | 32.9M | qcelp_rate_full_codebook[cbseed++ & 127]; |
380 | 822k | } |
381 | 205k | break; |
382 | 32.0k | case SILENCE: |
383 | 32.0k | memset(cdn_vector, 0, 160 * sizeof(float)); |
384 | 32.0k | break; |
385 | 248k | } |
386 | 248k | } |
387 | | |
388 | | /** |
389 | | * Apply generic gain control. |
390 | | * |
391 | | * @param v_out output vector |
392 | | * @param v_in gain-controlled vector |
393 | | * @param v_ref vector to control gain of |
394 | | * |
395 | | * TIA/EIA/IS-733 2.4.8.3, 2.4.8.6 |
396 | | */ |
397 | | static void apply_gain_ctrl(float *v_out, const float *v_ref, const float *v_in) |
398 | 12.1k | { |
399 | 12.1k | int i; |
400 | | |
401 | 60.6k | for (i = 0; i < 160; i += 40) { |
402 | 48.5k | float res = ff_scalarproduct_float_c(v_ref + i, v_ref + i, 40); |
403 | 48.5k | ff_scale_vector_to_given_sum_of_squares(v_out + i, v_in + i, res, 40); |
404 | 48.5k | } |
405 | 12.1k | } |
406 | | |
407 | | /** |
408 | | * Apply filter in pitch-subframe steps. |
409 | | * |
410 | | * @param memory buffer for the previous state of the filter |
411 | | * - must be able to contain 303 elements |
412 | | * - the 143 first elements are from the previous state |
413 | | * - the next 160 are for output |
414 | | * @param v_in input filter vector |
415 | | * @param gain per-subframe gain array, each element is between 0.0 and 2.0 |
416 | | * @param lag per-subframe lag array, each element is |
417 | | * - between 16 and 143 if its corresponding pfrac is 0, |
418 | | * - between 16 and 139 otherwise |
419 | | * @param pfrac per-subframe boolean array, 1 if the lag is fractional, 0 |
420 | | * otherwise |
421 | | * |
422 | | * @return filter output vector |
423 | | */ |
424 | | static const float *do_pitchfilter(float memory[303], const float v_in[160], |
425 | | const float gain[4], const uint8_t *lag, |
426 | | const uint8_t pfrac[4]) |
427 | 24.2k | { |
428 | 24.2k | int i, j; |
429 | 24.2k | float *v_lag, *v_out; |
430 | 24.2k | const float *v_len; |
431 | | |
432 | 24.2k | v_out = memory + 143; // Output vector starts at memory[143]. |
433 | | |
434 | 121k | for (i = 0; i < 4; i++) { |
435 | 97.0k | if (gain[i]) { |
436 | 28.2k | v_lag = memory + 143 + 40 * i - lag[i]; |
437 | 1.15M | for (v_len = v_in + 40; v_in < v_len; v_in++) { |
438 | 1.13M | if (pfrac[i]) { // If it is a fractional lag... |
439 | 3.11M | for (j = 0, *v_out = 0.0; j < 4; j++) |
440 | 2.49M | *v_out += qcelp_hammsinc_table[j] * |
441 | 2.49M | (v_lag[j - 4] + v_lag[3 - j]); |
442 | 622k | } else |
443 | 507k | *v_out = *v_lag; |
444 | | |
445 | 1.13M | *v_out = *v_in + gain[i] * *v_out; |
446 | | |
447 | 1.13M | v_lag++; |
448 | 1.13M | v_out++; |
449 | 1.13M | } |
450 | 68.7k | } else { |
451 | 68.7k | memcpy(v_out, v_in, 40 * sizeof(float)); |
452 | 68.7k | v_in += 40; |
453 | 68.7k | v_out += 40; |
454 | 68.7k | } |
455 | 97.0k | } |
456 | | |
457 | 24.2k | memmove(memory, memory + 160, 143 * sizeof(float)); |
458 | 24.2k | return memory + 143; |
459 | 24.2k | } |
460 | | |
461 | | /** |
462 | | * Apply pitch synthesis filter and pitch prefilter to the scaled codebook vector. |
463 | | * TIA/EIA/IS-733 2.4.5.2, 2.4.8.7.2 |
464 | | * |
465 | | * @param q the context |
466 | | * @param cdn_vector the scaled codebook vector |
467 | | */ |
468 | | static void apply_pitch_filters(QCELPContext *q, float *cdn_vector) |
469 | 223k | { |
470 | 223k | int i; |
471 | 223k | const float *v_synthesis_filtered, *v_pre_filtered; |
472 | | |
473 | 223k | if (q->bitrate >= RATE_HALF || q->bitrate == SILENCE || |
474 | 211k | (q->bitrate == I_F_Q && (q->prev_bitrate >= RATE_HALF))) { |
475 | | |
476 | 12.1k | if (q->bitrate >= RATE_HALF) { |
477 | | // Compute gain & lag for the whole frame. |
478 | 14.8k | for (i = 0; i < 4; i++) { |
479 | 11.9k | q->pitch_gain[i] = q->frame.plag[i] ? (q->frame.pgain[i] + 1) * 0.25 : 0.0; |
480 | | |
481 | 11.9k | q->pitch_lag[i] = q->frame.plag[i] + 16; |
482 | 11.9k | } |
483 | 9.14k | } else { |
484 | 9.14k | float max_pitch_gain; |
485 | | |
486 | 9.14k | if (q->bitrate == I_F_Q) { |
487 | 604 | if (q->erasure_count < 3) |
488 | 604 | max_pitch_gain = 0.9 - 0.3 * (q->erasure_count - 1); |
489 | 0 | else |
490 | 0 | max_pitch_gain = 0.0; |
491 | 8.54k | } else { |
492 | 8.54k | av_assert2(q->bitrate == SILENCE); |
493 | 8.54k | max_pitch_gain = 1.0; |
494 | 8.54k | } |
495 | 45.7k | for (i = 0; i < 4; i++) |
496 | 36.5k | q->pitch_gain[i] = FFMIN(q->pitch_gain[i], max_pitch_gain); |
497 | | |
498 | 9.14k | memset(q->frame.pfrac, 0, sizeof(q->frame.pfrac)); |
499 | 9.14k | } |
500 | | |
501 | | // pitch synthesis filter |
502 | 12.1k | v_synthesis_filtered = do_pitchfilter(q->pitch_synthesis_filter_mem, |
503 | 12.1k | cdn_vector, q->pitch_gain, |
504 | 12.1k | q->pitch_lag, q->frame.pfrac); |
505 | | |
506 | | // pitch prefilter update |
507 | 60.6k | for (i = 0; i < 4; i++) |
508 | 48.5k | q->pitch_gain[i] = 0.5 * FFMIN(q->pitch_gain[i], 1.0); |
509 | | |
510 | 12.1k | v_pre_filtered = do_pitchfilter(q->pitch_pre_filter_mem, |
511 | 12.1k | v_synthesis_filtered, |
512 | 12.1k | q->pitch_gain, q->pitch_lag, |
513 | 12.1k | q->frame.pfrac); |
514 | | |
515 | 12.1k | apply_gain_ctrl(cdn_vector, v_synthesis_filtered, v_pre_filtered); |
516 | 211k | } else { |
517 | 211k | memcpy(q->pitch_synthesis_filter_mem, |
518 | 211k | cdn_vector + 17, 143 * sizeof(float)); |
519 | 211k | memcpy(q->pitch_pre_filter_mem, cdn_vector + 17, 143 * sizeof(float)); |
520 | 211k | memset(q->pitch_gain, 0, sizeof(q->pitch_gain)); |
521 | 211k | memset(q->pitch_lag, 0, sizeof(q->pitch_lag)); |
522 | 211k | } |
523 | 223k | } |
524 | | |
525 | | /** |
526 | | * Reconstruct LPC coefficients from the line spectral pair frequencies |
527 | | * and perform bandwidth expansion. |
528 | | * |
529 | | * @param lspf line spectral pair frequencies |
530 | | * @param lpc linear predictive coding coefficients |
531 | | * |
532 | | * @note: bandwidth_expansion_coeff could be precalculated into a table |
533 | | * but it seems to be slower on x86 |
534 | | * |
535 | | * TIA/EIA/IS-733 2.4.3.3.5 |
536 | | */ |
537 | | static void lspf2lpc(const float *lspf, float *lpc) |
538 | 235k | { |
539 | 235k | double lsp[10]; |
540 | 235k | double bandwidth_expansion_coeff = QCELP_BANDWIDTH_EXPANSION_COEFF; |
541 | 235k | int i; |
542 | | |
543 | 2.59M | for (i = 0; i < 10; i++) |
544 | 2.35M | lsp[i] = cos(M_PI * lspf[i]); |
545 | | |
546 | 235k | ff_acelp_lspd2lpc(lsp, lpc, 5); |
547 | | |
548 | 2.59M | for (i = 0; i < 10; i++) { |
549 | 2.35M | lpc[i] *= bandwidth_expansion_coeff; |
550 | 2.35M | bandwidth_expansion_coeff *= QCELP_BANDWIDTH_EXPANSION_COEFF; |
551 | 2.35M | } |
552 | 235k | } |
553 | | |
554 | | /** |
555 | | * Interpolate LSP frequencies and compute LPC coefficients |
556 | | * for a given bitrate & pitch subframe. |
557 | | * |
558 | | * TIA/EIA/IS-733 2.4.3.3.4, 2.4.8.7.2 |
559 | | * |
560 | | * @param q the context |
561 | | * @param curr_lspf LSP frequencies vector of the current frame |
562 | | * @param lpc float vector for the resulting LPC |
563 | | * @param subframe_num frame number in decoded stream |
564 | | */ |
565 | | static void interpolate_lpc(QCELPContext *q, const float *curr_lspf, |
566 | | float *lpc, const int subframe_num) |
567 | 893k | { |
568 | 893k | float interpolated_lspf[10]; |
569 | 893k | float weight; |
570 | | |
571 | 893k | if (q->bitrate >= RATE_QUARTER) |
572 | 16.8k | weight = 0.25 * (subframe_num + 1); |
573 | 876k | else if (q->bitrate == RATE_OCTAVE && !subframe_num) |
574 | 4.84k | weight = 0.625; |
575 | 871k | else |
576 | 871k | weight = 1.0; |
577 | | |
578 | 893k | if (weight != 1.0) { |
579 | 17.4k | ff_weighted_vector_sumf(interpolated_lspf, curr_lspf, q->prev_lspf, |
580 | 17.4k | weight, 1.0 - weight, 10); |
581 | 17.4k | lspf2lpc(interpolated_lspf, lpc); |
582 | 875k | } else if (q->bitrate >= RATE_QUARTER || |
583 | 871k | (q->bitrate == I_F_Q && !subframe_num)) |
584 | 209k | lspf2lpc(curr_lspf, lpc); |
585 | 665k | else if (q->bitrate == SILENCE && !subframe_num) |
586 | 8.54k | lspf2lpc(q->prev_lspf, lpc); |
587 | 893k | } |
588 | | |
589 | | static qcelp_packet_rate buf_size2bitrate(const int buf_size) |
590 | 242k | { |
591 | 242k | switch (buf_size) { |
592 | 1.55k | case 35: return RATE_FULL; |
593 | 2.33k | case 17: return RATE_HALF; |
594 | 3.75k | case 8: return RATE_QUARTER; |
595 | 6.45k | case 4: return RATE_OCTAVE; |
596 | 202k | case 1: return SILENCE; |
597 | 242k | } |
598 | | |
599 | 26.3k | return I_F_Q; |
600 | 242k | } |
601 | | |
602 | | /** |
603 | | * Determine the bitrate from the frame size and/or the first byte of the frame. |
604 | | * |
605 | | * @param avctx the AV codec context |
606 | | * @param buf_size length of the buffer |
607 | | * @param buf the buffer |
608 | | * |
609 | | * @return the bitrate on success, |
610 | | * I_F_Q if the bitrate cannot be satisfactorily determined |
611 | | * |
612 | | * TIA/EIA/IS-733 2.4.8.7.1 |
613 | | */ |
614 | | static qcelp_packet_rate determine_bitrate(AVCodecContext *avctx, |
615 | | const int buf_size, |
616 | | const uint8_t **buf) |
617 | 223k | { |
618 | 223k | qcelp_packet_rate bitrate; |
619 | | |
620 | 223k | if ((bitrate = buf_size2bitrate(buf_size)) >= 0) { |
621 | 204k | if (bitrate > **buf) { |
622 | 1.08k | QCELPContext *q = avctx->priv_data; |
623 | 1.08k | if (!q->warned_buf_mismatch_bitrate) { |
624 | 90 | av_log(avctx, AV_LOG_WARNING, |
625 | 90 | "Claimed bitrate and buffer size mismatch.\n"); |
626 | 90 | q->warned_buf_mismatch_bitrate = 1; |
627 | 90 | } |
628 | 1.08k | bitrate = **buf; |
629 | 203k | } else if (bitrate < **buf) { |
630 | 171k | av_log(avctx, AV_LOG_ERROR, |
631 | 171k | "Buffer is too small for the claimed bitrate.\n"); |
632 | 171k | return I_F_Q; |
633 | 171k | } |
634 | 32.2k | (*buf)++; |
635 | 32.2k | } else if ((bitrate = buf_size2bitrate(buf_size + 1)) >= 0) { |
636 | 11.9k | av_log(avctx, AV_LOG_WARNING, |
637 | 11.9k | "Bitrate byte missing, guessing bitrate from packet size.\n"); |
638 | 11.9k | } else |
639 | 7.16k | return I_F_Q; |
640 | | |
641 | 44.1k | if (bitrate == SILENCE) { |
642 | | // FIXME: Remove this warning when tested with samples. |
643 | 32.0k | avpriv_request_sample(avctx, "Blank frame handling"); |
644 | 32.0k | } |
645 | 44.1k | return bitrate; |
646 | 223k | } |
647 | | |
648 | | static void warn_insufficient_frame_quality(AVCodecContext *avctx, |
649 | | const char *message) |
650 | 205k | { |
651 | 205k | av_log(avctx, AV_LOG_WARNING, "Frame #%"PRId64", IFQ: %s\n", |
652 | 205k | avctx->frame_num, message); |
653 | 205k | } |
654 | | |
655 | | static void postfilter(QCELPContext *q, float *samples, float *lpc) |
656 | 223k | { |
657 | 223k | static const float pow_0_775[10] = { |
658 | 223k | 0.775000, 0.600625, 0.465484, 0.360750, 0.279582, |
659 | 223k | 0.216676, 0.167924, 0.130141, 0.100859, 0.078166 |
660 | 223k | }, pow_0_625[10] = { |
661 | 223k | 0.625000, 0.390625, 0.244141, 0.152588, 0.095367, |
662 | 223k | 0.059605, 0.037253, 0.023283, 0.014552, 0.009095 |
663 | 223k | }; |
664 | 223k | float lpc_s[10], lpc_p[10], pole_out[170], zero_out[160]; |
665 | 223k | int n; |
666 | | |
667 | 2.45M | for (n = 0; n < 10; n++) { |
668 | 2.23M | lpc_s[n] = lpc[n] * pow_0_625[n]; |
669 | 2.23M | lpc_p[n] = lpc[n] * pow_0_775[n]; |
670 | 2.23M | } |
671 | | |
672 | 223k | ff_celp_lp_zero_synthesis_filterf(zero_out, lpc_s, |
673 | 223k | q->formant_mem + 10, 160, 10); |
674 | 223k | memcpy(pole_out, q->postfilter_synth_mem, sizeof(float) * 10); |
675 | 223k | ff_celp_lp_synthesis_filterf(pole_out + 10, lpc_p, zero_out, 160, 10); |
676 | 223k | memcpy(q->postfilter_synth_mem, pole_out + 160, sizeof(float) * 10); |
677 | | |
678 | 223k | ff_tilt_compensation(&q->postfilter_tilt_mem, 0.3, pole_out + 10, 160); |
679 | | |
680 | 223k | ff_adaptive_gain_control(samples, pole_out + 10, |
681 | 223k | ff_scalarproduct_float_c(q->formant_mem + 10, |
682 | 223k | q->formant_mem + 10, |
683 | 223k | 160), |
684 | 223k | 160, 0.9375, &q->postfilter_agc_mem); |
685 | 223k | } |
686 | | |
687 | | static int qcelp_decode_frame(AVCodecContext *avctx, AVFrame *frame, |
688 | | int *got_frame_ptr, AVPacket *avpkt) |
689 | 223k | { |
690 | 223k | const uint8_t *buf = avpkt->data; |
691 | 223k | int buf_size = avpkt->size; |
692 | 223k | QCELPContext *q = avctx->priv_data; |
693 | 223k | float *outbuffer; |
694 | 223k | int i, ret; |
695 | 223k | float quantized_lspf[10], lpc[10]; |
696 | 223k | float gain[16]; |
697 | 223k | float *formant_mem; |
698 | | |
699 | | /* get output buffer */ |
700 | 223k | frame->nb_samples = 160; |
701 | 223k | if ((ret = ff_get_buffer(avctx, frame, 0)) < 0) |
702 | 0 | return ret; |
703 | 223k | outbuffer = (float *)frame->data[0]; |
704 | | |
705 | 223k | if ((q->bitrate = determine_bitrate(avctx, buf_size, &buf)) == I_F_Q) { |
706 | 179k | warn_insufficient_frame_quality(avctx, "Bitrate cannot be determined."); |
707 | 179k | goto erasure; |
708 | 179k | } |
709 | | |
710 | 44.1k | if (q->bitrate == RATE_OCTAVE && |
711 | 5.77k | (q->first16bits = AV_RB16(buf)) == 0xFFFF) { |
712 | 232 | warn_insufficient_frame_quality(avctx, "Bitrate is 1/8 and first 16 bits are on."); |
713 | 232 | goto erasure; |
714 | 232 | } |
715 | | |
716 | 43.9k | if (q->bitrate > SILENCE) { |
717 | 11.9k | const QCELPBitmap *bitmaps = qcelp_unpacking_bitmaps_per_rate[q->bitrate]; |
718 | 11.9k | const QCELPBitmap *bitmaps_end = qcelp_unpacking_bitmaps_per_rate[q->bitrate] + |
719 | 11.9k | qcelp_unpacking_bitmaps_lengths[q->bitrate]; |
720 | 11.9k | uint8_t *unpacked_data = (uint8_t *)&q->frame; |
721 | | |
722 | 11.9k | if ((ret = init_get_bits8(&q->gb, buf, buf_size)) < 0) |
723 | 0 | return ret; |
724 | | |
725 | 11.9k | memset(&q->frame, 0, sizeof(QCELPFrame)); |
726 | | |
727 | 307k | for (; bitmaps < bitmaps_end; bitmaps++) |
728 | 295k | unpacked_data[bitmaps->index] |= get_bits(&q->gb, bitmaps->bitlen) << bitmaps->bitpos; |
729 | | |
730 | | // Check for erasures/blanks on rates 1, 1/4 and 1/8. |
731 | 11.9k | if (q->frame.reserved) { |
732 | 905 | warn_insufficient_frame_quality(avctx, "Wrong data in reserved frame area."); |
733 | 905 | goto erasure; |
734 | 905 | } |
735 | 11.0k | if (q->bitrate == RATE_QUARTER && |
736 | 2.83k | codebook_sanity_check_for_rate_quarter(q->frame.cbgain)) { |
737 | 496 | warn_insufficient_frame_quality(avctx, "Codebook gain sanity check failed."); |
738 | 496 | goto erasure; |
739 | 496 | } |
740 | | |
741 | 10.5k | if (q->bitrate >= RATE_HALF) { |
742 | 16.4k | for (i = 0; i < 4; i++) { |
743 | 13.3k | if (q->frame.pfrac[i] && q->frame.plag[i] >= 124) { |
744 | 247 | warn_insufficient_frame_quality(avctx, "Cannot initialize pitch filter."); |
745 | 247 | goto erasure; |
746 | 247 | } |
747 | 13.3k | } |
748 | 3.35k | } |
749 | 10.5k | } |
750 | | |
751 | 42.3k | decode_gain_and_index(q, gain); |
752 | 42.3k | compute_svector(q, gain, outbuffer); |
753 | | |
754 | 42.3k | if (decode_lspf(q, quantized_lspf) < 0) { |
755 | 24.7k | warn_insufficient_frame_quality(avctx, "Badly received packets in frame."); |
756 | 24.7k | goto erasure; |
757 | 24.7k | } |
758 | | |
759 | 17.5k | apply_pitch_filters(q, outbuffer); |
760 | | |
761 | 17.5k | if (q->bitrate == I_F_Q) { |
762 | 205k | erasure: |
763 | 205k | q->bitrate = I_F_Q; |
764 | 205k | q->erasure_count++; |
765 | 205k | decode_gain_and_index(q, gain); |
766 | 205k | compute_svector(q, gain, outbuffer); |
767 | 205k | decode_lspf(q, quantized_lspf); |
768 | 205k | apply_pitch_filters(q, outbuffer); |
769 | 205k | } else |
770 | 17.5k | q->erasure_count = 0; |
771 | | |
772 | 223k | formant_mem = q->formant_mem + 10; |
773 | 1.11M | for (i = 0; i < 4; i++) { |
774 | 893k | interpolate_lpc(q, quantized_lspf, lpc, i); |
775 | 893k | ff_celp_lp_synthesis_filterf(formant_mem, lpc, |
776 | 893k | outbuffer + i * 40, 40, 10); |
777 | 893k | formant_mem += 40; |
778 | 893k | } |
779 | | |
780 | | // postfilter, as per TIA/EIA/IS-733 2.4.8.6 |
781 | 223k | postfilter(q, outbuffer, lpc); |
782 | | |
783 | 223k | memcpy(q->formant_mem, q->formant_mem + 160, 10 * sizeof(float)); |
784 | | |
785 | 223k | memcpy(q->prev_lspf, quantized_lspf, sizeof(q->prev_lspf)); |
786 | 223k | q->prev_bitrate = q->bitrate; |
787 | | |
788 | 223k | *got_frame_ptr = 1; |
789 | | |
790 | 223k | return buf_size; |
791 | 17.5k | } |
792 | | |
793 | | const FFCodec ff_qcelp_decoder = { |
794 | | .p.name = "qcelp", |
795 | | CODEC_LONG_NAME("QCELP / PureVoice"), |
796 | | .p.type = AVMEDIA_TYPE_AUDIO, |
797 | | .p.id = AV_CODEC_ID_QCELP, |
798 | | .init = qcelp_decode_init, |
799 | | FF_CODEC_DECODE_CB(qcelp_decode_frame), |
800 | | .p.capabilities = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF, |
801 | | .priv_data_size = sizeof(QCELPContext), |
802 | | }; |