1/* 2 * Windows Media Audio Voice decoder. 3 * Copyright (c) 2009 Ronald S. Bultje 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 * @brief Windows Media Audio Voice compatible decoder 25 * @author Ronald S. Bultje <rsbultje@gmail.com> 26 */ 27 28#include <math.h> 29#include "avcodec.h" 30#include "get_bits.h" 31#include "put_bits.h" 32#include "wmavoice_data.h" 33#include "celp_math.h" 34#include "celp_filters.h" 35#include "acelp_vectors.h" 36#include "acelp_filters.h" 37#include "lsp.h" 38#include "libavutil/lzo.h" 39#include "avfft.h" 40#include "fft.h" 41 42#define MAX_BLOCKS 8 ///< maximum number of blocks per frame 43#define MAX_LSPS 16 ///< maximum filter order 44#define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple 45 ///< of 16 for ASM input buffer alignment 46#define MAX_FRAMES 3 ///< maximum number of frames per superframe 47#define MAX_FRAMESIZE 160 ///< maximum number of samples per frame 48#define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history 49#define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES) 50 ///< maximum number of samples per superframe 51#define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that 52 ///< was split over two packets 53#define VLC_NBITS 6 ///< number of bits to read per VLC iteration 54 55/** 56 * Frame type VLC coding. 57 */ 58static VLC frame_type_vlc; 59 60/** 61 * Adaptive codebook types. 62 */ 63enum { 64 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed) 65 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which 66 ///< we interpolate to get a per-sample pitch. 67 ///< Signal is generated using an asymmetric sinc 68 ///< window function 69 ///< @note see #wmavoice_ipol1_coeffs 70 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using 71 ///< a Hamming sinc window function 72 ///< @note see #wmavoice_ipol2_coeffs 73}; 74 75/** 76 * Fixed codebook types. 77 */ 78enum { 79 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence 80 ///< generated from a hardcoded (fixed) codebook 81 ///< with per-frame (low) gain values 82 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block 83 ///< gain values 84 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals, 85 ///< used in particular for low-bitrate streams 86 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in 87 ///< combinations of either single pulses or 88 ///< pulse pairs 89}; 90 91/** 92 * Description of frame types. 93 */ 94static const struct frame_type_desc { 95 uint8_t n_blocks; ///< amount of blocks per frame (each block 96 ///< (contains 160/#n_blocks samples) 97 uint8_t log_n_blocks; ///< log2(#n_blocks) 98 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*) 99 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*) 100 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs 101 ///< (rather than just one single pulse) 102 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES 103 uint16_t frame_size; ///< the amount of bits that make up the block 104 ///< data (per frame) 105} frame_descs[17] = { 106 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 }, 107 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 }, 108 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 }, 109 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 }, 110 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 }, 111 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 }, 112 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 }, 113 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 }, 114 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 }, 115 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 }, 116 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 }, 117 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 }, 118 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 }, 119 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 }, 120 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 }, 121 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 }, 122 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 } 123}; 124 125/** 126 * WMA Voice decoding context. 127 */ 128typedef struct { 129 /** 130 * @defgroup struct_global Global values 131 * Global values, specified in the stream header / extradata or used 132 * all over. 133 * @{ 134 */ 135 GetBitContext gb; ///< packet bitreader. During decoder init, 136 ///< it contains the extradata from the 137 ///< demuxer. During decoding, it contains 138 ///< packet data. 139 int8_t vbm_tree[25]; ///< converts VLC codes to frame type 140 141 int spillover_bitsize; ///< number of bits used to specify 142 ///< #spillover_nbits in the packet header 143 ///< = ceil(log2(ctx->block_align << 3)) 144 int history_nsamples; ///< number of samples in history for signal 145 ///< prediction (through ACB) 146 147 /* postfilter specific values */ 148 int do_apf; ///< whether to apply the averaged 149 ///< projection filter (APF) 150 int denoise_strength; ///< strength of denoising in Wiener filter 151 ///< [0-11] 152 int denoise_tilt_corr; ///< Whether to apply tilt correction to the 153 ///< Wiener filter coefficients (postfilter) 154 int dc_level; ///< Predicted amount of DC noise, based 155 ///< on which a DC removal filter is used 156 157 int lsps; ///< number of LSPs per frame [10 or 16] 158 int lsp_q_mode; ///< defines quantizer defaults [0, 1] 159 int lsp_def_mode; ///< defines different sets of LSP defaults 160 ///< [0, 1] 161 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded 162 ///< per-frame (independent coding) 163 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded 164 ///< per superframe (residual coding) 165 166 int min_pitch_val; ///< base value for pitch parsing code 167 int max_pitch_val; ///< max value + 1 for pitch parsing 168 int pitch_nbits; ///< number of bits used to specify the 169 ///< pitch value in the frame header 170 int block_pitch_nbits; ///< number of bits used to specify the 171 ///< first block's pitch value 172 int block_pitch_range; ///< range of the block pitch 173 int block_delta_pitch_nbits; ///< number of bits used to specify the 174 ///< delta pitch between this and the last 175 ///< block's pitch value, used in all but 176 ///< first block 177 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is 178 ///< from -this to +this-1) 179 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale 180 ///< conversion 181 182 /** 183 * @} 184 * @defgroup struct_packet Packet values 185 * Packet values, specified in the packet header or related to a packet. 186 * A packet is considered to be a single unit of data provided to this 187 * decoder by the demuxer. 188 * @{ 189 */ 190 int spillover_nbits; ///< number of bits of the previous packet's 191 ///< last superframe preceeding this 192 ///< packet's first full superframe (useful 193 ///< for re-synchronization also) 194 int has_residual_lsps; ///< if set, superframes contain one set of 195 ///< LSPs that cover all frames, encoded as 196 ///< independent and residual LSPs; if not 197 ///< set, each frame contains its own, fully 198 ///< independent, LSPs 199 int skip_bits_next; ///< number of bits to skip at the next call 200 ///< to #wmavoice_decode_packet() (since 201 ///< they're part of the previous superframe) 202 203 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE]; 204 ///< cache for superframe data split over 205 ///< multiple packets 206 int sframe_cache_size; ///< set to >0 if we have data from an 207 ///< (incomplete) superframe from a previous 208 ///< packet that spilled over in the current 209 ///< packet; specifies the amount of bits in 210 ///< #sframe_cache 211 PutBitContext pb; ///< bitstream writer for #sframe_cache 212 213 /** 214 * @} 215 * @defgroup struct_frame Frame and superframe values 216 * Superframe and frame data - these can change from frame to frame, 217 * although some of them do in that case serve as a cache / history for 218 * the next frame or superframe. 219 * @{ 220 */ 221 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous 222 ///< superframe 223 int last_pitch_val; ///< pitch value of the previous frame 224 int last_acb_type; ///< frame type [0-2] of the previous frame 225 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val) 226 ///< << 16) / #MAX_FRAMESIZE 227 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE 228 229 int aw_idx_is_ext; ///< whether the AW index was encoded in 230 ///< 8 bits (instead of 6) 231 int aw_pulse_range; ///< the range over which #aw_pulse_set1() 232 ///< can apply the pulse, relative to the 233 ///< value in aw_first_pulse_off. The exact 234 ///< position of the first AW-pulse is within 235 ///< [pulse_off, pulse_off + this], and 236 ///< depends on bitstream values; [16 or 24] 237 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note 238 ///< that this number can be negative (in 239 ///< which case it basically means "zero") 240 int aw_first_pulse_off[2]; ///< index of first sample to which to 241 ///< apply AW-pulses, or -0xff if unset 242 int aw_next_pulse_off_cache; ///< the position (relative to start of the 243 ///< second block) at which pulses should 244 ///< start to be positioned, serves as a 245 ///< cache for pitch-adaptive window pulses 246 ///< between blocks 247 248 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is 249 ///< only used for comfort noise in #pRNG() 250 float gain_pred_err[6]; ///< cache for gain prediction 251 float excitation_history[MAX_SIGNAL_HISTORY]; 252 ///< cache of the signal of previous 253 ///< superframes, used as a history for 254 ///< signal generation 255 float synth_history[MAX_LSPS]; ///< see #excitation_history 256 /** 257 * @} 258 * @defgroup post_filter Postfilter values 259 * Varibales used for postfilter implementation, mostly history for 260 * smoothing and so on, and context variables for FFT/iFFT. 261 * @{ 262 */ 263 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the 264 ///< postfilter (for denoise filter) 265 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert 266 ///< transform, part of postfilter) 267 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi] 268 ///< range 269 float postfilter_agc; ///< gain control memory, used in 270 ///< #adaptive_gain_control() 271 float dcf_mem[2]; ///< DC filter history 272 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE]; 273 ///< zero filter output (i.e. excitation) 274 ///< by postfilter 275 float denoise_filter_cache[MAX_FRAMESIZE]; 276 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache 277 DECLARE_ALIGNED(16, float, tilted_lpcs_pf)[0x80]; 278 ///< aligned buffer for LPC tilting 279 DECLARE_ALIGNED(16, float, denoise_coeffs_pf)[0x80]; 280 ///< aligned buffer for denoise coefficients 281 DECLARE_ALIGNED(16, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16]; 282 ///< aligned buffer for postfilter speech 283 ///< synthesis 284 /** 285 * @} 286 */ 287} WMAVoiceContext; 288 289/** 290 * Sets up the variable bit mode (VBM) tree from container extradata. 291 * @param gb bit I/O context. 292 * The bit context (s->gb) should be loaded with byte 23-46 of the 293 * container extradata (i.e. the ones containing the VBM tree). 294 * @param vbm_tree pointer to array to which the decoded VBM tree will be 295 * written. 296 * @return 0 on success, <0 on error. 297 */ 298static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25]) 299{ 300 static const uint8_t bits[] = { 301 2, 2, 2, 4, 4, 4, 302 6, 6, 6, 8, 8, 8, 303 10, 10, 10, 12, 12, 12, 304 14, 14, 14, 14 305 }; 306 static const uint16_t codes[] = { 307 0x0000, 0x0001, 0x0002, // 00/01/10 308 0x000c, 0x000d, 0x000e, // 11+00/01/10 309 0x003c, 0x003d, 0x003e, // 1111+00/01/10 310 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10 311 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10 312 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10 313 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx 314 }; 315 int cntr[8], n, res; 316 317 memset(vbm_tree, 0xff, sizeof(vbm_tree)); 318 memset(cntr, 0, sizeof(cntr)); 319 for (n = 0; n < 17; n++) { 320 res = get_bits(gb, 3); 321 if (cntr[res] > 3) // should be >= 3 + (res == 7)) 322 return -1; 323 vbm_tree[res * 3 + cntr[res]++] = n; 324 } 325 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits), 326 bits, 1, 1, codes, 2, 2, 132); 327 return 0; 328} 329 330/** 331 * Set up decoder with parameters from demuxer (extradata etc.). 332 */ 333static av_cold int wmavoice_decode_init(AVCodecContext *ctx) 334{ 335 int n, flags, pitch_range, lsp16_flag; 336 WMAVoiceContext *s = ctx->priv_data; 337 338 /** 339 * Extradata layout: 340 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c), 341 * - byte 19-22: flags field (annoyingly in LE; see below for known 342 * values), 343 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits, 344 * rest is 0). 345 */ 346 if (ctx->extradata_size != 46) { 347 av_log(ctx, AV_LOG_ERROR, 348 "Invalid extradata size %d (should be 46)\n", 349 ctx->extradata_size); 350 return -1; 351 } 352 flags = AV_RL32(ctx->extradata + 18); 353 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align); 354 s->do_apf = flags & 0x1; 355 if (s->do_apf) { 356 ff_rdft_init(&s->rdft, 7, DFT_R2C); 357 ff_rdft_init(&s->irdft, 7, IDFT_C2R); 358 ff_dct_init(&s->dct, 6, DCT_I); 359 ff_dct_init(&s->dst, 6, DST_I); 360 361 ff_sine_window_init(s->cos, 256); 362 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0])); 363 for (n = 0; n < 255; n++) { 364 s->sin[n] = -s->sin[510 - n]; 365 s->cos[510 - n] = s->cos[n]; 366 } 367 } 368 s->denoise_strength = (flags >> 2) & 0xF; 369 if (s->denoise_strength >= 12) { 370 av_log(ctx, AV_LOG_ERROR, 371 "Invalid denoise filter strength %d (max=11)\n", 372 s->denoise_strength); 373 return -1; 374 } 375 s->denoise_tilt_corr = !!(flags & 0x40); 376 s->dc_level = (flags >> 7) & 0xF; 377 s->lsp_q_mode = !!(flags & 0x2000); 378 s->lsp_def_mode = !!(flags & 0x4000); 379 lsp16_flag = flags & 0x1000; 380 if (lsp16_flag) { 381 s->lsps = 16; 382 s->frame_lsp_bitsize = 34; 383 s->sframe_lsp_bitsize = 60; 384 } else { 385 s->lsps = 10; 386 s->frame_lsp_bitsize = 24; 387 s->sframe_lsp_bitsize = 48; 388 } 389 for (n = 0; n < s->lsps; n++) 390 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); 391 392 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3); 393 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) { 394 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n"); 395 return -1; 396 } 397 398 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8; 399 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8; 400 pitch_range = s->max_pitch_val - s->min_pitch_val; 401 s->pitch_nbits = av_ceil_log2(pitch_range); 402 s->last_pitch_val = 40; 403 s->last_acb_type = ACB_TYPE_NONE; 404 s->history_nsamples = s->max_pitch_val + 8; 405 406 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) { 407 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8, 408 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8; 409 410 av_log(ctx, AV_LOG_ERROR, 411 "Unsupported samplerate %d (min=%d, max=%d)\n", 412 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz 413 414 return -1; 415 } 416 417 s->block_conv_table[0] = s->min_pitch_val; 418 s->block_conv_table[1] = (pitch_range * 25) >> 6; 419 s->block_conv_table[2] = (pitch_range * 44) >> 6; 420 s->block_conv_table[3] = s->max_pitch_val - 1; 421 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF; 422 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange); 423 s->block_pitch_range = s->block_conv_table[2] + 424 s->block_conv_table[3] + 1 + 425 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val); 426 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range); 427 428 ctx->sample_fmt = SAMPLE_FMT_FLT; 429 430 return 0; 431} 432 433/** 434 * @defgroup postfilter Postfilter functions 435 * Postfilter functions (gain control, wiener denoise filter, DC filter, 436 * kalman smoothening, plus surrounding code to wrap it) 437 * @{ 438 */ 439/** 440 * Adaptive gain control (as used in postfilter). 441 * 442 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except 443 * that the energy here is calculated using sum(abs(...)), whereas the 444 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)). 445 * 446 * @param out output buffer for filtered samples 447 * @param in input buffer containing the samples as they are after the 448 * postfilter steps so far 449 * @param speech_synth input buffer containing speech synth before postfilter 450 * @param size input buffer size 451 * @param alpha exponential filter factor 452 * @param gain_mem pointer to filter memory (single float) 453 */ 454static void adaptive_gain_control(float *out, const float *in, 455 const float *speech_synth, 456 int size, float alpha, float *gain_mem) 457{ 458 int i; 459 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor; 460 float mem = *gain_mem; 461 462 for (i = 0; i < size; i++) { 463 speech_energy += fabsf(speech_synth[i]); 464 postfilter_energy += fabsf(in[i]); 465 } 466 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy; 467 468 for (i = 0; i < size; i++) { 469 mem = alpha * mem + gain_scale_factor; 470 out[i] = in[i] * mem; 471 } 472 473 *gain_mem = mem; 474} 475 476/** 477 * Kalman smoothing function. 478 * 479 * This function looks back pitch +/- 3 samples back into history to find 480 * the best fitting curve (that one giving the optimal gain of the two 481 * signals, i.e. the highest dot product between the two), and then 482 * uses that signal history to smoothen the output of the speech synthesis 483 * filter. 484 * 485 * @param s WMA Voice decoding context 486 * @param pitch pitch of the speech signal 487 * @param in input speech signal 488 * @param out output pointer for smoothened signal 489 * @param size input/output buffer size 490 * 491 * @returns -1 if no smoothening took place, e.g. because no optimal 492 * fit could be found, or 0 on success. 493 */ 494static int kalman_smoothen(WMAVoiceContext *s, int pitch, 495 const float *in, float *out, int size) 496{ 497 int n; 498 float optimal_gain = 0, dot; 499 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)], 500 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)], 501 *best_hist_ptr; 502 503 /* find best fitting point in history */ 504 do { 505 dot = ff_dot_productf(in, ptr, size); 506 if (dot > optimal_gain) { 507 optimal_gain = dot; 508 best_hist_ptr = ptr; 509 } 510 } while (--ptr >= end); 511 512 if (optimal_gain <= 0) 513 return -1; 514 dot = ff_dot_productf(best_hist_ptr, best_hist_ptr, size); 515 if (dot <= 0) // would be 1.0 516 return -1; 517 518 if (optimal_gain <= dot) { 519 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000 520 } else 521 dot = 0.625; 522 523 /* actual smoothing */ 524 for (n = 0; n < size; n++) 525 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]); 526 527 return 0; 528} 529 530/** 531 * Get the tilt factor of a formant filter from its transfer function 532 * @see #tilt_factor() in amrnbdec.c, which does essentially the same, 533 * but somehow (??) it does a speech synthesis filter in the 534 * middle, which is missing here 535 * 536 * @param lpcs LPC coefficients 537 * @param n_lpcs Size of LPC buffer 538 * @returns the tilt factor 539 */ 540static float tilt_factor(const float *lpcs, int n_lpcs) 541{ 542 float rh0, rh1; 543 544 rh0 = 1.0 + ff_dot_productf(lpcs, lpcs, n_lpcs); 545 rh1 = lpcs[0] + ff_dot_productf(lpcs, &lpcs[1], n_lpcs - 1); 546 547 return rh1 / rh0; 548} 549 550/** 551 * Derive denoise filter coefficients (in real domain) from the LPCs. 552 */ 553static void calc_input_response(WMAVoiceContext *s, float *lpcs, 554 int fcb_type, float *coeffs, int remainder) 555{ 556 float last_coeff, min = 15.0, max = -15.0; 557 float irange, angle_mul, gain_mul, range, sq; 558 int n, idx; 559 560 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */ 561 ff_rdft_calc(&s->rdft, lpcs); 562#define log_range(var, assign) do { \ 563 float tmp = log10f(assign); var = tmp; \ 564 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \ 565 } while (0) 566 log_range(last_coeff, lpcs[1] * lpcs[1]); 567 for (n = 1; n < 64; n++) 568 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] + 569 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]); 570 log_range(lpcs[0], lpcs[0] * lpcs[0]); 571#undef log_range 572 range = max - min; 573 lpcs[64] = last_coeff; 574 575 /* Now, use this spectrum to pick out these frequencies with higher 576 * (relative) power/energy (which we then take to be "not noise"), 577 * and set up a table (still in lpc[]) of (relative) gains per frequency. 578 * These frequencies will be maintained, while others ("noise") will be 579 * decreased in the filter output. */ 580 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63] 581 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) : 582 (5.0 / 14.7)); 583 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI); 584 for (n = 0; n <= 64; n++) { 585 float pow; 586 587 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1); 588 pow = wmavoice_denoise_power_table[s->denoise_strength][idx]; 589 lpcs[n] = angle_mul * pow; 590 591 /* 70.57 =~ 1/log10(1.0331663) */ 592 idx = (pow * gain_mul - 0.0295) * 70.570526123; 593 if (idx > 127) { // fallback if index falls outside table range 594 coeffs[n] = wmavoice_energy_table[127] * 595 powf(1.0331663, idx - 127); 596 } else 597 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)]; 598 } 599 600 /* calculate the Hilbert transform of the gains, which we do (since this 601 * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()). 602 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the 603 * "moment" of the LPCs in this filter. */ 604 ff_dct_calc(&s->dct, lpcs); 605 ff_dct_calc(&s->dst, lpcs); 606 607 /* Split out the coefficient indexes into phase/magnitude pairs */ 608 idx = 255 + av_clip(lpcs[64], -255, 255); 609 coeffs[0] = coeffs[0] * s->cos[idx]; 610 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255); 611 last_coeff = coeffs[64] * s->cos[idx]; 612 for (n = 63;; n--) { 613 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255); 614 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; 615 coeffs[n * 2] = coeffs[n] * s->cos[idx]; 616 617 if (!--n) break; 618 619 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255); 620 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx]; 621 coeffs[n * 2] = coeffs[n] * s->cos[idx]; 622 } 623 coeffs[1] = last_coeff; 624 625 /* move into real domain */ 626 ff_rdft_calc(&s->irdft, coeffs); 627 628 /* tilt correction and normalize scale */ 629 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder)); 630 if (s->denoise_tilt_corr) { 631 float tilt_mem = 0; 632 633 coeffs[remainder - 1] = 0; 634 ff_tilt_compensation(&tilt_mem, 635 -1.8 * tilt_factor(coeffs, remainder - 1), 636 coeffs, remainder); 637 } 638 sq = (1.0 / 64.0) * sqrtf(1 / ff_dot_productf(coeffs, coeffs, remainder)); 639 for (n = 0; n < remainder; n++) 640 coeffs[n] *= sq; 641} 642 643/** 644 * This function applies a Wiener filter on the (noisy) speech signal as 645 * a means to denoise it. 646 * 647 * - take RDFT of LPCs to get the power spectrum of the noise + speech; 648 * - using this power spectrum, calculate (for each frequency) the Wiener 649 * filter gain, which depends on the frequency power and desired level 650 * of noise subtraction (when set too high, this leads to artifacts) 651 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse 652 * of 4-8kHz); 653 * - by doing a phase shift, calculate the Hilbert transform of this array 654 * of per-frequency filter-gains to get the filtering coefficients; 655 * - smoothen/normalize/de-tilt these filter coefficients as desired; 656 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT 657 * to get the denoised speech signal; 658 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond 659 * the frame boundary) are saved and applied to subsequent frames by an 660 * overlap-add method (otherwise you get clicking-artifacts). 661 * 662 * @param s WMA Voice decoding context 663 * @param s fcb_type Frame (codebook) type 664 * @param synth_pf input: the noisy speech signal, output: denoised speech 665 * data; should be 16-byte aligned (for ASM purposes) 666 * @param size size of the speech data 667 * @param lpcs LPCs used to synthesize this frame's speech data 668 */ 669static void wiener_denoise(WMAVoiceContext *s, int fcb_type, 670 float *synth_pf, int size, 671 const float *lpcs) 672{ 673 int remainder, lim, n; 674 675 if (fcb_type != FCB_TYPE_SILENCE) { 676 float *tilted_lpcs = s->tilted_lpcs_pf, 677 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0; 678 679 tilted_lpcs[0] = 1.0; 680 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps); 681 memset(&tilted_lpcs[s->lsps + 1], 0, 682 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1)); 683 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps), 684 tilted_lpcs, s->lsps + 2); 685 686 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame 687 * size is applied to the next frame. All input beyond this is zero, 688 * and thus all output beyond this will go towards zero, hence we can 689 * limit to min(size-1, 127-size) as a performance consideration. */ 690 remainder = FFMIN(127 - size, size - 1); 691 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder); 692 693 /* apply coefficients (in frequency spectrum domain), i.e. complex 694 * number multiplication */ 695 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size)); 696 ff_rdft_calc(&s->rdft, synth_pf); 697 ff_rdft_calc(&s->rdft, coeffs); 698 synth_pf[0] *= coeffs[0]; 699 synth_pf[1] *= coeffs[1]; 700 for (n = 1; n < 64; n++) { 701 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1]; 702 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1]; 703 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1]; 704 } 705 ff_rdft_calc(&s->irdft, synth_pf); 706 } 707 708 /* merge filter output with the history of previous runs */ 709 if (s->denoise_filter_cache_size) { 710 lim = FFMIN(s->denoise_filter_cache_size, size); 711 for (n = 0; n < lim; n++) 712 synth_pf[n] += s->denoise_filter_cache[n]; 713 s->denoise_filter_cache_size -= lim; 714 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size], 715 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size); 716 } 717 718 /* move remainder of filter output into a cache for future runs */ 719 if (fcb_type != FCB_TYPE_SILENCE) { 720 lim = FFMIN(remainder, s->denoise_filter_cache_size); 721 for (n = 0; n < lim; n++) 722 s->denoise_filter_cache[n] += synth_pf[size + n]; 723 if (lim < remainder) { 724 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim], 725 sizeof(s->denoise_filter_cache[0]) * (remainder - lim)); 726 s->denoise_filter_cache_size = remainder; 727 } 728 } 729} 730 731/** 732 * Averaging projection filter, the postfilter used in WMAVoice. 733 * 734 * This uses the following steps: 735 * - A zero-synthesis filter (generate excitation from synth signal) 736 * - Kalman smoothing on excitation, based on pitch 737 * - Re-synthesized smoothened output 738 * - Iterative Wiener denoise filter 739 * - Adaptive gain filter 740 * - DC filter 741 * 742 * @param s WMAVoice decoding context 743 * @param synth Speech synthesis output (before postfilter) 744 * @param samples Output buffer for filtered samples 745 * @param size Buffer size of synth & samples 746 * @param lpcs Generated LPCs used for speech synthesis 747 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses) 748 * @param pitch Pitch of the input signal 749 */ 750static void postfilter(WMAVoiceContext *s, const float *synth, 751 float *samples, int size, 752 const float *lpcs, float *zero_exc_pf, 753 int fcb_type, int pitch) 754{ 755 float synth_filter_in_buf[MAX_FRAMESIZE / 2], 756 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16], 757 *synth_filter_in = zero_exc_pf; 758 759 assert(size <= MAX_FRAMESIZE / 2); 760 761 /* generate excitation from input signal */ 762 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps); 763 764 if (fcb_type >= FCB_TYPE_AW_PULSES && 765 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size)) 766 synth_filter_in = synth_filter_in_buf; 767 768 /* re-synthesize speech after smoothening, and keep history */ 769 ff_celp_lp_synthesis_filterf(synth_pf, lpcs, 770 synth_filter_in, size, s->lsps); 771 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps], 772 sizeof(synth_pf[0]) * s->lsps); 773 774 wiener_denoise(s, fcb_type, synth_pf, size, lpcs); 775 776 adaptive_gain_control(samples, synth_pf, synth, size, 0.99, 777 &s->postfilter_agc); 778 779 if (s->dc_level > 8) { 780 /* remove ultra-low frequency DC noise / highpass filter; 781 * coefficients are identical to those used in SIPR decoding, 782 * and very closely resemble those used in AMR-NB decoding. */ 783 ff_acelp_apply_order_2_transfer_function(samples, samples, 784 (const float[2]) { -1.99997, 1.0 }, 785 (const float[2]) { -1.9330735188, 0.93589198496 }, 786 0.93980580475, s->dcf_mem, size); 787 } 788} 789/** 790 * @} 791 */ 792 793/** 794 * Dequantize LSPs 795 * @param lsps output pointer to the array that will hold the LSPs 796 * @param num number of LSPs to be dequantized 797 * @param values quantized values, contains n_stages values 798 * @param sizes range (i.e. max value) of each quantized value 799 * @param n_stages number of dequantization runs 800 * @param table dequantization table to be used 801 * @param mul_q LSF multiplier 802 * @param base_q base (lowest) LSF values 803 */ 804static void dequant_lsps(double *lsps, int num, 805 const uint16_t *values, 806 const uint16_t *sizes, 807 int n_stages, const uint8_t *table, 808 const double *mul_q, 809 const double *base_q) 810{ 811 int n, m; 812 813 memset(lsps, 0, num * sizeof(*lsps)); 814 for (n = 0; n < n_stages; n++) { 815 const uint8_t *t_off = &table[values[n] * num]; 816 double base = base_q[n], mul = mul_q[n]; 817 818 for (m = 0; m < num; m++) 819 lsps[m] += base + mul * t_off[m]; 820 821 table += sizes[n] * num; 822 } 823} 824 825/** 826 * @defgroup lsp_dequant LSP dequantization routines 827 * LSP dequantization routines, for 10/16LSPs and independent/residual coding. 828 * @note we assume enough bits are available, caller should check. 829 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits; 830 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits. 831 * @{ 832 */ 833/** 834 * Parse 10 independently-coded LSPs. 835 */ 836static void dequant_lsp10i(GetBitContext *gb, double *lsps) 837{ 838 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 }; 839 static const double mul_lsf[4] = { 840 5.2187144800e-3, 1.4626986422e-3, 841 9.6179549166e-4, 1.1325736225e-3 842 }; 843 static const double base_lsf[4] = { 844 M_PI * -2.15522e-1, M_PI * -6.1646e-2, 845 M_PI * -3.3486e-2, M_PI * -5.7408e-2 846 }; 847 uint16_t v[4]; 848 849 v[0] = get_bits(gb, 8); 850 v[1] = get_bits(gb, 6); 851 v[2] = get_bits(gb, 5); 852 v[3] = get_bits(gb, 5); 853 854 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i, 855 mul_lsf, base_lsf); 856} 857 858/** 859 * Parse 10 independently-coded LSPs, and then derive the tables to 860 * generate LSPs for the other frames from them (residual coding). 861 */ 862static void dequant_lsp10r(GetBitContext *gb, 863 double *i_lsps, const double *old, 864 double *a1, double *a2, int q_mode) 865{ 866 static const uint16_t vec_sizes[3] = { 128, 64, 64 }; 867 static const double mul_lsf[3] = { 868 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3 869 }; 870 static const double base_lsf[3] = { 871 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2 872 }; 873 const float (*ipol_tab)[2][10] = q_mode ? 874 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a; 875 uint16_t interpol, v[3]; 876 int n; 877 878 dequant_lsp10i(gb, i_lsps); 879 880 interpol = get_bits(gb, 5); 881 v[0] = get_bits(gb, 7); 882 v[1] = get_bits(gb, 6); 883 v[2] = get_bits(gb, 6); 884 885 for (n = 0; n < 10; n++) { 886 double delta = old[n] - i_lsps[n]; 887 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; 888 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; 889 } 890 891 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r, 892 mul_lsf, base_lsf); 893} 894 895/** 896 * Parse 16 independently-coded LSPs. 897 */ 898static void dequant_lsp16i(GetBitContext *gb, double *lsps) 899{ 900 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 }; 901 static const double mul_lsf[5] = { 902 3.3439586280e-3, 6.9908173703e-4, 903 3.3216608306e-3, 1.0334960326e-3, 904 3.1899104283e-3 905 }; 906 static const double base_lsf[5] = { 907 M_PI * -1.27576e-1, M_PI * -2.4292e-2, 908 M_PI * -1.28094e-1, M_PI * -3.2128e-2, 909 M_PI * -1.29816e-1 910 }; 911 uint16_t v[5]; 912 913 v[0] = get_bits(gb, 8); 914 v[1] = get_bits(gb, 6); 915 v[2] = get_bits(gb, 7); 916 v[3] = get_bits(gb, 6); 917 v[4] = get_bits(gb, 7); 918 919 dequant_lsps( lsps, 5, v, vec_sizes, 2, 920 wmavoice_dq_lsp16i1, mul_lsf, base_lsf); 921 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2, 922 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]); 923 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1, 924 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]); 925} 926 927/** 928 * Parse 16 independently-coded LSPs, and then derive the tables to 929 * generate LSPs for the other frames from them (residual coding). 930 */ 931static void dequant_lsp16r(GetBitContext *gb, 932 double *i_lsps, const double *old, 933 double *a1, double *a2, int q_mode) 934{ 935 static const uint16_t vec_sizes[3] = { 128, 128, 128 }; 936 static const double mul_lsf[3] = { 937 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3 938 }; 939 static const double base_lsf[3] = { 940 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2 941 }; 942 const float (*ipol_tab)[2][16] = q_mode ? 943 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a; 944 uint16_t interpol, v[3]; 945 int n; 946 947 dequant_lsp16i(gb, i_lsps); 948 949 interpol = get_bits(gb, 5); 950 v[0] = get_bits(gb, 7); 951 v[1] = get_bits(gb, 7); 952 v[2] = get_bits(gb, 7); 953 954 for (n = 0; n < 16; n++) { 955 double delta = old[n] - i_lsps[n]; 956 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n]; 957 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n]; 958 } 959 960 dequant_lsps( a2, 10, v, vec_sizes, 1, 961 wmavoice_dq_lsp16r1, mul_lsf, base_lsf); 962 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1, 963 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]); 964 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1, 965 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]); 966} 967 968/** 969 * @} 970 * @defgroup aw Pitch-adaptive window coding functions 971 * The next few functions are for pitch-adaptive window coding. 972 * @{ 973 */ 974/** 975 * Parse the offset of the first pitch-adaptive window pulses, and 976 * the distribution of pulses between the two blocks in this frame. 977 * @param s WMA Voice decoding context private data 978 * @param gb bit I/O context 979 * @param pitch pitch for each block in this frame 980 */ 981static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb, 982 const int *pitch) 983{ 984 static const int16_t start_offset[94] = { 985 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11, 986 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26, 987 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43, 988 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67, 989 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91, 990 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115, 991 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139, 992 141, 143, 145, 147, 149, 151, 153, 155, 157, 159 993 }; 994 int bits, offset; 995 996 /* position of pulse */ 997 s->aw_idx_is_ext = 0; 998 if ((bits = get_bits(gb, 6)) >= 54) { 999 s->aw_idx_is_ext = 1; 1000 bits += (bits - 54) * 3 + get_bits(gb, 2); 1001 } 1002 1003 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count 1004 * the distribution of the pulses in each block contained in this frame. */ 1005 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16; 1006 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ; 1007 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0]; 1008 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2; 1009 offset += s->aw_n_pulses[0] * pitch[0]; 1010 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1]; 1011 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2; 1012 1013 /* if continuing from a position before the block, reset position to 1014 * start of block (when corrected for the range over which it can be 1015 * spread in aw_pulse_set1()). */ 1016 if (start_offset[bits] < MAX_FRAMESIZE / 2) { 1017 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0) 1018 s->aw_first_pulse_off[1] -= pitch[1]; 1019 if (start_offset[bits] < 0) 1020 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0) 1021 s->aw_first_pulse_off[0] -= pitch[0]; 1022 } 1023} 1024 1025/** 1026 * Apply second set of pitch-adaptive window pulses. 1027 * @param s WMA Voice decoding context private data 1028 * @param gb bit I/O context 1029 * @param block_idx block index in frame [0, 1] 1030 * @param fcb structure containing fixed codebook vector info 1031 */ 1032static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb, 1033 int block_idx, AMRFixed *fcb) 1034{ 1035 uint16_t use_mask[7]; // only 5 are used, rest is padding 1036 /* in this function, idx is the index in the 80-bit (+ padding) use_mask 1037 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits 1038 * of idx are the position of the bit within a particular item in the 1039 * array (0 being the most significant bit, and 15 being the least 1040 * significant bit), and the remainder (>> 4) is the index in the 1041 * use_mask[]-array. This is faster and uses less memory than using a 1042 * 80-byte/80-int array. */ 1043 int pulse_off = s->aw_first_pulse_off[block_idx], 1044 pulse_start, n, idx, range, aidx, start_off = 0; 1045 1046 /* set offset of first pulse to within this block */ 1047 if (s->aw_n_pulses[block_idx] > 0) 1048 while (pulse_off + s->aw_pulse_range < 1) 1049 pulse_off += fcb->pitch_lag; 1050 1051 /* find range per pulse */ 1052 if (s->aw_n_pulses[0] > 0) { 1053 if (block_idx == 0) { 1054 range = 32; 1055 } else /* block_idx = 1 */ { 1056 range = 8; 1057 if (s->aw_n_pulses[block_idx] > 0) 1058 pulse_off = s->aw_next_pulse_off_cache; 1059 } 1060 } else 1061 range = 16; 1062 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0; 1063 1064 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly, 1065 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus 1066 * we exclude that range from being pulsed again in this function. */ 1067 memset( use_mask, -1, 5 * sizeof(use_mask[0])); 1068 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0])); 1069 if (s->aw_n_pulses[block_idx] > 0) 1070 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) { 1071 int excl_range = s->aw_pulse_range; // always 16 or 24 1072 uint16_t *use_mask_ptr = &use_mask[idx >> 4]; 1073 int first_sh = 16 - (idx & 15); 1074 *use_mask_ptr++ &= 0xFFFF << first_sh; 1075 excl_range -= first_sh; 1076 if (excl_range >= 16) { 1077 *use_mask_ptr++ = 0; 1078 *use_mask_ptr &= 0xFFFF >> (excl_range - 16); 1079 } else 1080 *use_mask_ptr &= 0xFFFF >> excl_range; 1081 } 1082 1083 /* find the 'aidx'th offset that is not excluded */ 1084 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4); 1085 for (n = 0; n <= aidx; pulse_start++) { 1086 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ; 1087 if (idx >= MAX_FRAMESIZE / 2) { // find from zero 1088 if (use_mask[0]) idx = 0x0F; 1089 else if (use_mask[1]) idx = 0x1F; 1090 else if (use_mask[2]) idx = 0x2F; 1091 else if (use_mask[3]) idx = 0x3F; 1092 else if (use_mask[4]) idx = 0x4F; 1093 else return; 1094 idx -= av_log2_16bit(use_mask[idx >> 4]); 1095 } 1096 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) { 1097 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15)); 1098 n++; 1099 start_off = idx; 1100 } 1101 } 1102 1103 fcb->x[fcb->n] = start_off; 1104 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0; 1105 fcb->n++; 1106 1107 /* set offset for next block, relative to start of that block */ 1108 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag; 1109 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0; 1110} 1111 1112/** 1113 * Apply first set of pitch-adaptive window pulses. 1114 * @param s WMA Voice decoding context private data 1115 * @param gb bit I/O context 1116 * @param block_idx block index in frame [0, 1] 1117 * @param fcb storage location for fixed codebook pulse info 1118 */ 1119static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb, 1120 int block_idx, AMRFixed *fcb) 1121{ 1122 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx)); 1123 float v; 1124 1125 if (s->aw_n_pulses[block_idx] > 0) { 1126 int n, v_mask, i_mask, sh, n_pulses; 1127 1128 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each 1129 n_pulses = 3; 1130 v_mask = 8; 1131 i_mask = 7; 1132 sh = 4; 1133 } else { // 4 pulses, 1:sign + 2:index each 1134 n_pulses = 4; 1135 v_mask = 4; 1136 i_mask = 3; 1137 sh = 3; 1138 } 1139 1140 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) { 1141 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0; 1142 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n + 1143 s->aw_first_pulse_off[block_idx]; 1144 while (fcb->x[fcb->n] < 0) 1145 fcb->x[fcb->n] += fcb->pitch_lag; 1146 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2) 1147 fcb->n++; 1148 } 1149 } else { 1150 int num2 = (val & 0x1FF) >> 1, delta, idx; 1151 1152 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; } 1153 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; } 1154 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; } 1155 else { delta = 7; idx = num2 + 1 - 3 * 75; } 1156 v = (val & 0x200) ? -1.0 : 1.0; 1157 1158 fcb->no_repeat_mask |= 3 << fcb->n; 1159 fcb->x[fcb->n] = idx - delta; 1160 fcb->y[fcb->n] = v; 1161 fcb->x[fcb->n + 1] = idx; 1162 fcb->y[fcb->n + 1] = (val & 1) ? -v : v; 1163 fcb->n += 2; 1164 } 1165} 1166 1167/** 1168 * @} 1169 * 1170 * Generate a random number from frame_cntr and block_idx, which will lief 1171 * in the range [0, 1000 - block_size] (so it can be used as an index in a 1172 * table of size 1000 of which you want to read block_size entries). 1173 * 1174 * @param frame_cntr current frame number 1175 * @param block_num current block index 1176 * @param block_size amount of entries we want to read from a table 1177 * that has 1000 entries 1178 * @return a (non-)random number in the [0, 1000 - block_size] range. 1179 */ 1180static int pRNG(int frame_cntr, int block_num, int block_size) 1181{ 1182 /* array to simplify the calculation of z: 1183 * y = (x % 9) * 5 + 6; 1184 * z = (49995 * x) / y; 1185 * Since y only has 9 values, we can remove the division by using a 1186 * LUT and using FASTDIV-style divisions. For each of the 9 values 1187 * of y, we can rewrite z as: 1188 * z = x * (49995 / y) + x * ((49995 % y) / y) 1189 * In this table, each col represents one possible value of y, the 1190 * first number is 49995 / y, and the second is the FASTDIV variant 1191 * of 49995 % y / y. */ 1192 static const unsigned int div_tbl[9][2] = { 1193 { 8332, 3 * 715827883U }, // y = 6 1194 { 4545, 0 * 390451573U }, // y = 11 1195 { 3124, 11 * 268435456U }, // y = 16 1196 { 2380, 15 * 204522253U }, // y = 21 1197 { 1922, 23 * 165191050U }, // y = 26 1198 { 1612, 23 * 138547333U }, // y = 31 1199 { 1388, 27 * 119304648U }, // y = 36 1200 { 1219, 16 * 104755300U }, // y = 41 1201 { 1086, 39 * 93368855U } // y = 46 1202 }; 1203 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr; 1204 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6, 1205 // so this is effectively a modulo (%) 1206 y = x - 9 * MULH(477218589, x); // x % 9 1207 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1])); 1208 // z = x * 49995 / (y * 5 + 6) 1209 return z % (1000 - block_size); 1210} 1211 1212/** 1213 * Parse hardcoded signal for a single block. 1214 * @note see #synth_block(). 1215 */ 1216static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb, 1217 int block_idx, int size, 1218 const struct frame_type_desc *frame_desc, 1219 float *excitation) 1220{ 1221 float gain; 1222 int n, r_idx; 1223 1224 assert(size <= MAX_FRAMESIZE); 1225 1226 /* Set the offset from which we start reading wmavoice_std_codebook */ 1227 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) { 1228 r_idx = pRNG(s->frame_cntr, block_idx, size); 1229 gain = s->silence_gain; 1230 } else /* FCB_TYPE_HARDCODED */ { 1231 r_idx = get_bits(gb, 8); 1232 gain = wmavoice_gain_universal[get_bits(gb, 6)]; 1233 } 1234 1235 /* Clear gain prediction parameters */ 1236 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err)); 1237 1238 /* Apply gain to hardcoded codebook and use that as excitation signal */ 1239 for (n = 0; n < size; n++) 1240 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain; 1241} 1242 1243/** 1244 * Parse FCB/ACB signal for a single block. 1245 * @note see #synth_block(). 1246 */ 1247static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb, 1248 int block_idx, int size, 1249 int block_pitch_sh2, 1250 const struct frame_type_desc *frame_desc, 1251 float *excitation) 1252{ 1253 static const float gain_coeff[6] = { 1254 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458 1255 }; 1256 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain; 1257 int n, idx, gain_weight; 1258 AMRFixed fcb; 1259 1260 assert(size <= MAX_FRAMESIZE / 2); 1261 memset(pulses, 0, sizeof(*pulses) * size); 1262 1263 fcb.pitch_lag = block_pitch_sh2 >> 2; 1264 fcb.pitch_fac = 1.0; 1265 fcb.no_repeat_mask = 0; 1266 fcb.n = 0; 1267 1268 /* For the other frame types, this is where we apply the innovation 1269 * (fixed) codebook pulses of the speech signal. */ 1270 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { 1271 aw_pulse_set1(s, gb, block_idx, &fcb); 1272 aw_pulse_set2(s, gb, block_idx, &fcb); 1273 } else /* FCB_TYPE_EXC_PULSES */ { 1274 int offset_nbits = 5 - frame_desc->log_n_blocks; 1275 1276 fcb.no_repeat_mask = -1; 1277 /* similar to ff_decode_10_pulses_35bits(), but with single pulses 1278 * (instead of double) for a subset of pulses */ 1279 for (n = 0; n < 5; n++) { 1280 float sign; 1281 int pos1, pos2; 1282 1283 sign = get_bits1(gb) ? 1.0 : -1.0; 1284 pos1 = get_bits(gb, offset_nbits); 1285 fcb.x[fcb.n] = n + 5 * pos1; 1286 fcb.y[fcb.n++] = sign; 1287 if (n < frame_desc->dbl_pulses) { 1288 pos2 = get_bits(gb, offset_nbits); 1289 fcb.x[fcb.n] = n + 5 * pos2; 1290 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign; 1291 } 1292 } 1293 } 1294 ff_set_fixed_vector(pulses, &fcb, 1.0, size); 1295 1296 /* Calculate gain for adaptive & fixed codebook signal. 1297 * see ff_amr_set_fixed_gain(). */ 1298 idx = get_bits(gb, 7); 1299 fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) - 1300 5.2409161640 + wmavoice_gain_codebook_fcb[idx]); 1301 acb_gain = wmavoice_gain_codebook_acb[idx]; 1302 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx], 1303 -2.9957322736 /* log(0.05) */, 1304 1.6094379124 /* log(5.0) */); 1305 1306 gain_weight = 8 >> frame_desc->log_n_blocks; 1307 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err, 1308 sizeof(*s->gain_pred_err) * (6 - gain_weight)); 1309 for (n = 0; n < gain_weight; n++) 1310 s->gain_pred_err[n] = pred_err; 1311 1312 /* Calculation of adaptive codebook */ 1313 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) { 1314 int len; 1315 for (n = 0; n < size; n += len) { 1316 int next_idx_sh16; 1317 int abs_idx = block_idx * size + n; 1318 int pitch_sh16 = (s->last_pitch_val << 16) + 1319 s->pitch_diff_sh16 * abs_idx; 1320 int pitch = (pitch_sh16 + 0x6FFF) >> 16; 1321 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000; 1322 idx = idx_sh16 >> 16; 1323 if (s->pitch_diff_sh16) { 1324 if (s->pitch_diff_sh16 > 0) { 1325 next_idx_sh16 = (idx_sh16) &~ 0xFFFF; 1326 } else 1327 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF; 1328 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8, 1329 1, size - n); 1330 } else 1331 len = size; 1332 1333 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch], 1334 wmavoice_ipol1_coeffs, 17, 1335 idx, 9, len); 1336 } 1337 } else /* ACB_TYPE_HAMMING */ { 1338 int block_pitch = block_pitch_sh2 >> 2; 1339 idx = block_pitch_sh2 & 3; 1340 if (idx) { 1341 ff_acelp_interpolatef(excitation, &excitation[-block_pitch], 1342 wmavoice_ipol2_coeffs, 4, 1343 idx, 8, size); 1344 } else 1345 av_memcpy_backptr(excitation, sizeof(float) * block_pitch, 1346 sizeof(float) * size); 1347 } 1348 1349 /* Interpolate ACB/FCB and use as excitation signal */ 1350 ff_weighted_vector_sumf(excitation, excitation, pulses, 1351 acb_gain, fcb_gain, size); 1352} 1353 1354/** 1355 * Parse data in a single block. 1356 * @note we assume enough bits are available, caller should check. 1357 * 1358 * @param s WMA Voice decoding context private data 1359 * @param gb bit I/O context 1360 * @param block_idx index of the to-be-read block 1361 * @param size amount of samples to be read in this block 1362 * @param block_pitch_sh2 pitch for this block << 2 1363 * @param lsps LSPs for (the end of) this frame 1364 * @param prev_lsps LSPs for the last frame 1365 * @param frame_desc frame type descriptor 1366 * @param excitation target memory for the ACB+FCB interpolated signal 1367 * @param synth target memory for the speech synthesis filter output 1368 * @return 0 on success, <0 on error. 1369 */ 1370static void synth_block(WMAVoiceContext *s, GetBitContext *gb, 1371 int block_idx, int size, 1372 int block_pitch_sh2, 1373 const double *lsps, const double *prev_lsps, 1374 const struct frame_type_desc *frame_desc, 1375 float *excitation, float *synth) 1376{ 1377 double i_lsps[MAX_LSPS]; 1378 float lpcs[MAX_LSPS]; 1379 float fac; 1380 int n; 1381 1382 if (frame_desc->acb_type == ACB_TYPE_NONE) 1383 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation); 1384 else 1385 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2, 1386 frame_desc, excitation); 1387 1388 /* convert interpolated LSPs to LPCs */ 1389 fac = (block_idx + 0.5) / frame_desc->n_blocks; 1390 for (n = 0; n < s->lsps; n++) // LSF -> LSP 1391 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n])); 1392 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); 1393 1394 /* Speech synthesis */ 1395 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps); 1396} 1397 1398/** 1399 * Synthesize output samples for a single frame. 1400 * @note we assume enough bits are available, caller should check. 1401 * 1402 * @param ctx WMA Voice decoder context 1403 * @param gb bit I/O context (s->gb or one for cross-packet superframes) 1404 * @param frame_idx Frame number within superframe [0-2] 1405 * @param samples pointer to output sample buffer, has space for at least 160 1406 * samples 1407 * @param lsps LSP array 1408 * @param prev_lsps array of previous frame's LSPs 1409 * @param excitation target buffer for excitation signal 1410 * @param synth target buffer for synthesized speech data 1411 * @return 0 on success, <0 on error. 1412 */ 1413static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx, 1414 float *samples, 1415 const double *lsps, const double *prev_lsps, 1416 float *excitation, float *synth) 1417{ 1418 WMAVoiceContext *s = ctx->priv_data; 1419 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val; 1420 int pitch[MAX_BLOCKS], last_block_pitch; 1421 1422 /* Parse frame type ("frame header"), see frame_descs */ 1423 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], 1424 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks; 1425 1426 if (bd_idx < 0) { 1427 av_log(ctx, AV_LOG_ERROR, 1428 "Invalid frame type VLC code, skipping\n"); 1429 return -1; 1430 } 1431 1432 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */ 1433 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) { 1434 /* Pitch is provided per frame, which is interpreted as the pitch of 1435 * the last sample of the last block of this frame. We can interpolate 1436 * the pitch of other blocks (and even pitch-per-sample) by gradually 1437 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */ 1438 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1; 1439 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1; 1440 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits); 1441 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1); 1442 if (s->last_acb_type == ACB_TYPE_NONE || 1443 20 * abs(cur_pitch_val - s->last_pitch_val) > 1444 (cur_pitch_val + s->last_pitch_val)) 1445 s->last_pitch_val = cur_pitch_val; 1446 1447 /* pitch per block */ 1448 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { 1449 int fac = n * 2 + 1; 1450 1451 pitch[n] = (MUL16(fac, cur_pitch_val) + 1452 MUL16((n_blocks_x2 - fac), s->last_pitch_val) + 1453 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2; 1454 } 1455 1456 /* "pitch-diff-per-sample" for calculation of pitch per sample */ 1457 s->pitch_diff_sh16 = 1458 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE; 1459 } 1460 1461 /* Global gain (if silence) and pitch-adaptive window coordinates */ 1462 switch (frame_descs[bd_idx].fcb_type) { 1463 case FCB_TYPE_SILENCE: 1464 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)]; 1465 break; 1466 case FCB_TYPE_AW_PULSES: 1467 aw_parse_coords(s, gb, pitch); 1468 break; 1469 } 1470 1471 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) { 1472 int bl_pitch_sh2; 1473 1474 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */ 1475 switch (frame_descs[bd_idx].acb_type) { 1476 case ACB_TYPE_HAMMING: { 1477 /* Pitch is given per block. Per-block pitches are encoded as an 1478 * absolute value for the first block, and then delta values 1479 * relative to this value) for all subsequent blocks. The scale of 1480 * this pitch value is semi-logaritmic compared to its use in the 1481 * decoder, so we convert it to normal scale also. */ 1482 int block_pitch, 1483 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2, 1484 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1, 1485 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1; 1486 1487 if (n == 0) { 1488 block_pitch = get_bits(gb, s->block_pitch_nbits); 1489 } else 1490 block_pitch = last_block_pitch - s->block_delta_pitch_hrange + 1491 get_bits(gb, s->block_delta_pitch_nbits); 1492 /* Convert last_ so that any next delta is within _range */ 1493 last_block_pitch = av_clip(block_pitch, 1494 s->block_delta_pitch_hrange, 1495 s->block_pitch_range - 1496 s->block_delta_pitch_hrange); 1497 1498 /* Convert semi-log-style scale back to normal scale */ 1499 if (block_pitch < t1) { 1500 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch; 1501 } else { 1502 block_pitch -= t1; 1503 if (block_pitch < t2) { 1504 bl_pitch_sh2 = 1505 (s->block_conv_table[1] << 2) + (block_pitch << 1); 1506 } else { 1507 block_pitch -= t2; 1508 if (block_pitch < t3) { 1509 bl_pitch_sh2 = 1510 (s->block_conv_table[2] + block_pitch) << 2; 1511 } else 1512 bl_pitch_sh2 = s->block_conv_table[3] << 2; 1513 } 1514 } 1515 pitch[n] = bl_pitch_sh2 >> 2; 1516 break; 1517 } 1518 1519 case ACB_TYPE_ASYMMETRIC: { 1520 bl_pitch_sh2 = pitch[n] << 2; 1521 break; 1522 } 1523 1524 default: // ACB_TYPE_NONE has no pitch 1525 bl_pitch_sh2 = 0; 1526 break; 1527 } 1528 1529 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2, 1530 lsps, prev_lsps, &frame_descs[bd_idx], 1531 &excitation[n * block_nsamples], 1532 &synth[n * block_nsamples]); 1533 } 1534 1535 /* Averaging projection filter, if applicable. Else, just copy samples 1536 * from synthesis buffer */ 1537 if (s->do_apf) { 1538 double i_lsps[MAX_LSPS]; 1539 float lpcs[MAX_LSPS]; 1540 1541 for (n = 0; n < s->lsps; n++) // LSF -> LSP 1542 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n])); 1543 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); 1544 postfilter(s, synth, samples, 80, lpcs, 1545 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx], 1546 frame_descs[bd_idx].fcb_type, pitch[0]); 1547 1548 for (n = 0; n < s->lsps; n++) // LSF -> LSP 1549 i_lsps[n] = cos(lsps[n]); 1550 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1); 1551 postfilter(s, &synth[80], &samples[80], 80, lpcs, 1552 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80], 1553 frame_descs[bd_idx].fcb_type, pitch[0]); 1554 } else 1555 memcpy(samples, synth, 160 * sizeof(synth[0])); 1556 1557 /* Cache values for next frame */ 1558 s->frame_cntr++; 1559 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%) 1560 s->last_acb_type = frame_descs[bd_idx].acb_type; 1561 switch (frame_descs[bd_idx].acb_type) { 1562 case ACB_TYPE_NONE: 1563 s->last_pitch_val = 0; 1564 break; 1565 case ACB_TYPE_ASYMMETRIC: 1566 s->last_pitch_val = cur_pitch_val; 1567 break; 1568 case ACB_TYPE_HAMMING: 1569 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1]; 1570 break; 1571 } 1572 1573 return 0; 1574} 1575 1576/** 1577 * Ensure minimum value for first item, maximum value for last value, 1578 * proper spacing between each value and proper ordering. 1579 * 1580 * @param lsps array of LSPs 1581 * @param num size of LSP array 1582 * 1583 * @note basically a double version of #ff_acelp_reorder_lsf(), might be 1584 * useful to put in a generic location later on. Parts are also 1585 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(), 1586 * which is in float. 1587 */ 1588static void stabilize_lsps(double *lsps, int num) 1589{ 1590 int n, m, l; 1591 1592 /* set minimum value for first, maximum value for last and minimum 1593 * spacing between LSF values. 1594 * Very similar to ff_set_min_dist_lsf(), but in double. */ 1595 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI); 1596 for (n = 1; n < num; n++) 1597 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI); 1598 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI); 1599 1600 /* reorder (looks like one-time / non-recursed bubblesort). 1601 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */ 1602 for (n = 1; n < num; n++) { 1603 if (lsps[n] < lsps[n - 1]) { 1604 for (m = 1; m < num; m++) { 1605 double tmp = lsps[m]; 1606 for (l = m - 1; l >= 0; l--) { 1607 if (lsps[l] <= tmp) break; 1608 lsps[l + 1] = lsps[l]; 1609 } 1610 lsps[l + 1] = tmp; 1611 } 1612 break; 1613 } 1614 } 1615} 1616 1617/** 1618 * Test if there's enough bits to read 1 superframe. 1619 * 1620 * @param orig_gb bit I/O context used for reading. This function 1621 * does not modify the state of the bitreader; it 1622 * only uses it to copy the current stream position 1623 * @param s WMA Voice decoding context private data 1624 * @return -1 if unsupported, 1 on not enough bits or 0 if OK. 1625 */ 1626static int check_bits_for_superframe(GetBitContext *orig_gb, 1627 WMAVoiceContext *s) 1628{ 1629 GetBitContext s_gb, *gb = &s_gb; 1630 int n, need_bits, bd_idx; 1631 const struct frame_type_desc *frame_desc; 1632 1633 /* initialize a copy */ 1634 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits); 1635 skip_bits_long(gb, get_bits_count(orig_gb)); 1636 assert(get_bits_left(gb) == get_bits_left(orig_gb)); 1637 1638 /* superframe header */ 1639 if (get_bits_left(gb) < 14) 1640 return 1; 1641 if (!get_bits1(gb)) 1642 return -1; // WMAPro-in-WMAVoice superframe 1643 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe 1644 if (s->has_residual_lsps) { // residual LSPs (for all frames) 1645 if (get_bits_left(gb) < s->sframe_lsp_bitsize) 1646 return 1; 1647 skip_bits_long(gb, s->sframe_lsp_bitsize); 1648 } 1649 1650 /* frames */ 1651 for (n = 0; n < MAX_FRAMES; n++) { 1652 int aw_idx_is_ext = 0; 1653 1654 if (!s->has_residual_lsps) { // independent LSPs (per-frame) 1655 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1; 1656 skip_bits_long(gb, s->frame_lsp_bitsize); 1657 } 1658 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)]; 1659 if (bd_idx < 0) 1660 return -1; // invalid frame type VLC code 1661 frame_desc = &frame_descs[bd_idx]; 1662 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) { 1663 if (get_bits_left(gb) < s->pitch_nbits) 1664 return 1; 1665 skip_bits_long(gb, s->pitch_nbits); 1666 } 1667 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) { 1668 skip_bits(gb, 8); 1669 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { 1670 int tmp = get_bits(gb, 6); 1671 if (tmp >= 0x36) { 1672 skip_bits(gb, 2); 1673 aw_idx_is_ext = 1; 1674 } 1675 } 1676 1677 /* blocks */ 1678 if (frame_desc->acb_type == ACB_TYPE_HAMMING) { 1679 need_bits = s->block_pitch_nbits + 1680 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits; 1681 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) { 1682 need_bits = 2 * !aw_idx_is_ext; 1683 } else 1684 need_bits = 0; 1685 need_bits += frame_desc->frame_size; 1686 if (get_bits_left(gb) < need_bits) 1687 return 1; 1688 skip_bits_long(gb, need_bits); 1689 } 1690 1691 return 0; 1692} 1693 1694/** 1695 * Synthesize output samples for a single superframe. If we have any data 1696 * cached in s->sframe_cache, that will be used instead of whatever is loaded 1697 * in s->gb. 1698 * 1699 * WMA Voice superframes contain 3 frames, each containing 160 audio samples, 1700 * to give a total of 480 samples per frame. See #synth_frame() for frame 1701 * parsing. In addition to 3 frames, superframes can also contain the LSPs 1702 * (if these are globally specified for all frames (residually); they can 1703 * also be specified individually per-frame. See the s->has_residual_lsps 1704 * option), and can specify the number of samples encoded in this superframe 1705 * (if less than 480), usually used to prevent blanks at track boundaries. 1706 * 1707 * @param ctx WMA Voice decoder context 1708 * @param samples pointer to output buffer for voice samples 1709 * @param data_size pointer containing the size of #samples on input, and the 1710 * amount of #samples filled on output 1711 * @return 0 on success, <0 on error or 1 if there was not enough data to 1712 * fully parse the superframe 1713 */ 1714static int synth_superframe(AVCodecContext *ctx, 1715 float *samples, int *data_size) 1716{ 1717 WMAVoiceContext *s = ctx->priv_data; 1718 GetBitContext *gb = &s->gb, s_gb; 1719 int n, res, n_samples = 480; 1720 double lsps[MAX_FRAMES][MAX_LSPS]; 1721 const double *mean_lsf = s->lsps == 16 ? 1722 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode]; 1723 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12]; 1724 float synth[MAX_LSPS + MAX_SFRAMESIZE]; 1725 1726 memcpy(synth, s->synth_history, 1727 s->lsps * sizeof(*synth)); 1728 memcpy(excitation, s->excitation_history, 1729 s->history_nsamples * sizeof(*excitation)); 1730 1731 if (s->sframe_cache_size > 0) { 1732 gb = &s_gb; 1733 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size); 1734 s->sframe_cache_size = 0; 1735 } 1736 1737 if ((res = check_bits_for_superframe(gb, s)) == 1) return 1; 1738 1739 /* First bit is speech/music bit, it differentiates between WMAVoice 1740 * speech samples (the actual codec) and WMAVoice music samples, which 1741 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in 1742 * the wild yet. */ 1743 if (!get_bits1(gb)) { 1744 av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1); 1745 return -1; 1746 } 1747 1748 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */ 1749 if (get_bits1(gb)) { 1750 if ((n_samples = get_bits(gb, 12)) > 480) { 1751 av_log(ctx, AV_LOG_ERROR, 1752 "Superframe encodes >480 samples (%d), not allowed\n", 1753 n_samples); 1754 return -1; 1755 } 1756 } 1757 /* Parse LSPs, if global for the superframe (can also be per-frame). */ 1758 if (s->has_residual_lsps) { 1759 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2]; 1760 1761 for (n = 0; n < s->lsps; n++) 1762 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n]; 1763 1764 if (s->lsps == 10) { 1765 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); 1766 } else /* s->lsps == 16 */ 1767 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode); 1768 1769 for (n = 0; n < s->lsps; n++) { 1770 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]); 1771 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]); 1772 lsps[2][n] += mean_lsf[n]; 1773 } 1774 for (n = 0; n < 3; n++) 1775 stabilize_lsps(lsps[n], s->lsps); 1776 } 1777 1778 /* Parse frames, optionally preceeded by per-frame (independent) LSPs. */ 1779 for (n = 0; n < 3; n++) { 1780 if (!s->has_residual_lsps) { 1781 int m; 1782 1783 if (s->lsps == 10) { 1784 dequant_lsp10i(gb, lsps[n]); 1785 } else /* s->lsps == 16 */ 1786 dequant_lsp16i(gb, lsps[n]); 1787 1788 for (m = 0; m < s->lsps; m++) 1789 lsps[n][m] += mean_lsf[m]; 1790 stabilize_lsps(lsps[n], s->lsps); 1791 } 1792 1793 if ((res = synth_frame(ctx, gb, n, 1794 &samples[n * MAX_FRAMESIZE], 1795 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1], 1796 &excitation[s->history_nsamples + n * MAX_FRAMESIZE], 1797 &synth[s->lsps + n * MAX_FRAMESIZE]))) 1798 return res; 1799 } 1800 1801 /* Statistics? FIXME - we don't check for length, a slight overrun 1802 * will be caught by internal buffer padding, and anything else 1803 * will be skipped, not read. */ 1804 if (get_bits1(gb)) { 1805 res = get_bits(gb, 4); 1806 skip_bits(gb, 10 * (res + 1)); 1807 } 1808 1809 /* Specify nr. of output samples */ 1810 *data_size = n_samples * sizeof(float); 1811 1812 /* Update history */ 1813 memcpy(s->prev_lsps, lsps[2], 1814 s->lsps * sizeof(*s->prev_lsps)); 1815 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE], 1816 s->lsps * sizeof(*synth)); 1817 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE], 1818 s->history_nsamples * sizeof(*excitation)); 1819 if (s->do_apf) 1820 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE], 1821 s->history_nsamples * sizeof(*s->zero_exc_pf)); 1822 1823 return 0; 1824} 1825 1826/** 1827 * Parse the packet header at the start of each packet (input data to this 1828 * decoder). 1829 * 1830 * @param s WMA Voice decoding context private data 1831 * @return 1 if not enough bits were available, or 0 on success. 1832 */ 1833static int parse_packet_header(WMAVoiceContext *s) 1834{ 1835 GetBitContext *gb = &s->gb; 1836 unsigned int res; 1837 1838 if (get_bits_left(gb) < 11) 1839 return 1; 1840 skip_bits(gb, 4); // packet sequence number 1841 s->has_residual_lsps = get_bits1(gb); 1842 do { 1843 res = get_bits(gb, 6); // number of superframes per packet 1844 // (minus first one if there is spillover) 1845 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize) 1846 return 1; 1847 } while (res == 0x3F); 1848 s->spillover_nbits = get_bits(gb, s->spillover_bitsize); 1849 1850 return 0; 1851} 1852 1853/** 1854 * Copy (unaligned) bits from gb/data/size to pb. 1855 * 1856 * @param pb target buffer to copy bits into 1857 * @param data source buffer to copy bits from 1858 * @param size size of the source data, in bytes 1859 * @param gb bit I/O context specifying the current position in the source. 1860 * data. This function might use this to align the bit position to 1861 * a whole-byte boundary before calling #ff_copy_bits() on aligned 1862 * source data 1863 * @param nbits the amount of bits to copy from source to target 1864 * 1865 * @note after calling this function, the current position in the input bit 1866 * I/O context is undefined. 1867 */ 1868static void copy_bits(PutBitContext *pb, 1869 const uint8_t *data, int size, 1870 GetBitContext *gb, int nbits) 1871{ 1872 int rmn_bytes, rmn_bits; 1873 1874 rmn_bits = rmn_bytes = get_bits_left(gb); 1875 if (rmn_bits < nbits) 1876 return; 1877 rmn_bits &= 7; rmn_bytes >>= 3; 1878 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0) 1879 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits)); 1880 ff_copy_bits(pb, data + size - rmn_bytes, 1881 FFMIN(nbits - rmn_bits, rmn_bytes << 3)); 1882} 1883 1884/** 1885 * Packet decoding: a packet is anything that the (ASF) demuxer contains, 1886 * and we expect that the demuxer / application provides it to us as such 1887 * (else you'll probably get garbage as output). Every packet has a size of 1888 * ctx->block_align bytes, starts with a packet header (see 1889 * #parse_packet_header()), and then a series of superframes. Superframe 1890 * boundaries may exceed packets, i.e. superframes can split data over 1891 * multiple (two) packets. 1892 * 1893 * For more information about frames, see #synth_superframe(). 1894 */ 1895static int wmavoice_decode_packet(AVCodecContext *ctx, void *data, 1896 int *data_size, AVPacket *avpkt) 1897{ 1898 WMAVoiceContext *s = ctx->priv_data; 1899 GetBitContext *gb = &s->gb; 1900 int size, res, pos; 1901 1902 if (*data_size < 480 * sizeof(float)) { 1903 av_log(ctx, AV_LOG_ERROR, 1904 "Output buffer too small (%d given - %lu needed)\n", 1905 *data_size, 480 * sizeof(float)); 1906 return -1; 1907 } 1908 *data_size = 0; 1909 1910 /* Packets are sometimes a multiple of ctx->block_align, with a packet 1911 * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer 1912 * feeds us ASF packets, which may concatenate multiple "codec" packets 1913 * in a single "muxer" packet, so we artificially emulate that by 1914 * capping the packet size at ctx->block_align. */ 1915 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align); 1916 if (!size) 1917 return 0; 1918 init_get_bits(&s->gb, avpkt->data, size << 3); 1919 1920 /* size == ctx->block_align is used to indicate whether we are dealing with 1921 * a new packet or a packet of which we already read the packet header 1922 * previously. */ 1923 if (size == ctx->block_align) { // new packet header 1924 if ((res = parse_packet_header(s)) < 0) 1925 return res; 1926 1927 /* If the packet header specifies a s->spillover_nbits, then we want 1928 * to push out all data of the previous packet (+ spillover) before 1929 * continuing to parse new superframes in the current packet. */ 1930 if (s->spillover_nbits > 0) { 1931 if (s->sframe_cache_size > 0) { 1932 int cnt = get_bits_count(gb); 1933 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits); 1934 flush_put_bits(&s->pb); 1935 s->sframe_cache_size += s->spillover_nbits; 1936 if ((res = synth_superframe(ctx, data, data_size)) == 0 && 1937 *data_size > 0) { 1938 cnt += s->spillover_nbits; 1939 s->skip_bits_next = cnt & 7; 1940 return cnt >> 3; 1941 } else 1942 skip_bits_long (gb, s->spillover_nbits - cnt + 1943 get_bits_count(gb)); // resync 1944 } else 1945 skip_bits_long(gb, s->spillover_nbits); // resync 1946 } 1947 } else if (s->skip_bits_next) 1948 skip_bits(gb, s->skip_bits_next); 1949 1950 /* Try parsing superframes in current packet */ 1951 s->sframe_cache_size = 0; 1952 s->skip_bits_next = 0; 1953 pos = get_bits_left(gb); 1954 if ((res = synth_superframe(ctx, data, data_size)) < 0) { 1955 return res; 1956 } else if (*data_size > 0) { 1957 int cnt = get_bits_count(gb); 1958 s->skip_bits_next = cnt & 7; 1959 return cnt >> 3; 1960 } else if ((s->sframe_cache_size = pos) > 0) { 1961 /* rewind bit reader to start of last (incomplete) superframe... */ 1962 init_get_bits(gb, avpkt->data, size << 3); 1963 skip_bits_long(gb, (size << 3) - pos); 1964 assert(get_bits_left(gb) == pos); 1965 1966 /* ...and cache it for spillover in next packet */ 1967 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE); 1968 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size); 1969 // FIXME bad - just copy bytes as whole and add use the 1970 // skip_bits_next field 1971 } 1972 1973 return size; 1974} 1975 1976static av_cold int wmavoice_decode_end(AVCodecContext *ctx) 1977{ 1978 WMAVoiceContext *s = ctx->priv_data; 1979 1980 if (s->do_apf) { 1981 ff_rdft_end(&s->rdft); 1982 ff_rdft_end(&s->irdft); 1983 ff_dct_end(&s->dct); 1984 ff_dct_end(&s->dst); 1985 } 1986 1987 return 0; 1988} 1989 1990static av_cold void wmavoice_flush(AVCodecContext *ctx) 1991{ 1992 WMAVoiceContext *s = ctx->priv_data; 1993 int n; 1994 1995 s->postfilter_agc = 0; 1996 s->sframe_cache_size = 0; 1997 s->skip_bits_next = 0; 1998 for (n = 0; n < s->lsps; n++) 1999 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0); 2000 memset(s->excitation_history, 0, 2001 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY); 2002 memset(s->synth_history, 0, 2003 sizeof(*s->synth_history) * MAX_LSPS); 2004 memset(s->gain_pred_err, 0, 2005 sizeof(s->gain_pred_err)); 2006 2007 if (s->do_apf) { 2008 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0, 2009 sizeof(*s->synth_filter_out_buf) * s->lsps); 2010 memset(s->dcf_mem, 0, 2011 sizeof(*s->dcf_mem) * 2); 2012 memset(s->zero_exc_pf, 0, 2013 sizeof(*s->zero_exc_pf) * s->history_nsamples); 2014 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache)); 2015 } 2016} 2017 2018AVCodec wmavoice_decoder = { 2019 "wmavoice", 2020 AVMEDIA_TYPE_AUDIO, 2021 CODEC_ID_WMAVOICE, 2022 sizeof(WMAVoiceContext), 2023 wmavoice_decode_init, 2024 NULL, 2025 wmavoice_decode_end, 2026 wmavoice_decode_packet, 2027 CODEC_CAP_SUBFRAMES, 2028 .flush = wmavoice_flush, 2029 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"), 2030}; 2031