2 * Windows Media Audio Voice decoder.
3 * Copyright (c) 2009 Ronald S. Bultje
5 * This file is part of Libav.
7 * Libav is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; either
10 * version 2.1 of the License, or (at your option) any later version.
12 * Libav is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with Libav; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
24 * @brief Windows Media Audio Voice compatible decoder
25 * @author Ronald S. Bultje <rsbultje@gmail.com>
28 #define UNCHECKED_BITSTREAM_READER 1
32 #include "libavutil/channel_layout.h"
33 #include "libavutil/float_dsp.h"
34 #include "libavutil/mem.h"
39 #include "wmavoice_data.h"
40 #include "celp_filters.h"
41 #include "acelp_vectors.h"
42 #include "acelp_filters.h"
48 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
49 #define MAX_LSPS 16 ///< maximum filter order
50 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
51 ///< of 16 for ASM input buffer alignment
52 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
53 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
54 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
55 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
56 ///< maximum number of samples per superframe
57 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
58 ///< was split over two packets
59 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
62 * Frame type VLC coding.
64 static VLC frame_type_vlc;
67 * Adaptive codebook types.
70 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
71 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
72 ///< we interpolate to get a per-sample pitch.
73 ///< Signal is generated using an asymmetric sinc
75 ///< @note see #wmavoice_ipol1_coeffs
76 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
77 ///< a Hamming sinc window function
78 ///< @note see #wmavoice_ipol2_coeffs
82 * Fixed codebook types.
85 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
86 ///< generated from a hardcoded (fixed) codebook
87 ///< with per-frame (low) gain values
88 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
90 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
91 ///< used in particular for low-bitrate streams
92 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
93 ///< combinations of either single pulses or
98 * Description of frame types.
100 static const struct frame_type_desc {
101 uint8_t n_blocks; ///< amount of blocks per frame (each block
102 ///< (contains 160/#n_blocks samples)
103 uint8_t log_n_blocks; ///< log2(#n_blocks)
104 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
105 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
106 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
107 ///< (rather than just one single pulse)
108 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
109 uint16_t frame_size; ///< the amount of bits that make up the block
110 ///< data (per frame)
111 } frame_descs[17] = {
112 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
113 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
114 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
115 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
116 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
117 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
118 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
119 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
120 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
121 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
122 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
123 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
124 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
125 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
126 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
127 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
128 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
132 * WMA Voice decoding context.
136 * @name Global values specified in the stream header / extradata or used all over.
139 GetBitContext gb; ///< packet bitreader. During decoder init,
140 ///< it contains the extradata from the
141 ///< demuxer. During decoding, it contains
143 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
145 int spillover_bitsize; ///< number of bits used to specify
146 ///< #spillover_nbits in the packet header
147 ///< = ceil(log2(ctx->block_align << 3))
148 int history_nsamples; ///< number of samples in history for signal
149 ///< prediction (through ACB)
151 /* postfilter specific values */
152 int do_apf; ///< whether to apply the averaged
153 ///< projection filter (APF)
154 int denoise_strength; ///< strength of denoising in Wiener filter
156 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
157 ///< Wiener filter coefficients (postfilter)
158 int dc_level; ///< Predicted amount of DC noise, based
159 ///< on which a DC removal filter is used
161 int lsps; ///< number of LSPs per frame [10 or 16]
162 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
163 int lsp_def_mode; ///< defines different sets of LSP defaults
165 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
166 ///< per-frame (independent coding)
167 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
168 ///< per superframe (residual coding)
170 int min_pitch_val; ///< base value for pitch parsing code
171 int max_pitch_val; ///< max value + 1 for pitch parsing
172 int pitch_nbits; ///< number of bits used to specify the
173 ///< pitch value in the frame header
174 int block_pitch_nbits; ///< number of bits used to specify the
175 ///< first block's pitch value
176 int block_pitch_range; ///< range of the block pitch
177 int block_delta_pitch_nbits; ///< number of bits used to specify the
178 ///< delta pitch between this and the last
179 ///< block's pitch value, used in all but
181 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
182 ///< from -this to +this-1)
183 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
189 * @name Packet values specified in the packet header or related to a packet.
191 * A packet is considered to be a single unit of data provided to this
192 * decoder by the demuxer.
195 int spillover_nbits; ///< number of bits of the previous packet's
196 ///< last superframe preceding this
197 ///< packet's first full superframe (useful
198 ///< for re-synchronization also)
199 int has_residual_lsps; ///< if set, superframes contain one set of
200 ///< LSPs that cover all frames, encoded as
201 ///< independent and residual LSPs; if not
202 ///< set, each frame contains its own, fully
203 ///< independent, LSPs
204 int skip_bits_next; ///< number of bits to skip at the next call
205 ///< to #wmavoice_decode_packet() (since
206 ///< they're part of the previous superframe)
208 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];
209 ///< cache for superframe data split over
210 ///< multiple packets
211 int sframe_cache_size; ///< set to >0 if we have data from an
212 ///< (incomplete) superframe from a previous
213 ///< packet that spilled over in the current
214 ///< packet; specifies the amount of bits in
216 PutBitContext pb; ///< bitstream writer for #sframe_cache
221 * @name Frame and superframe values
222 * Superframe and frame data - these can change from frame to frame,
223 * although some of them do in that case serve as a cache / history for
224 * the next frame or superframe.
227 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
229 int last_pitch_val; ///< pitch value of the previous frame
230 int last_acb_type; ///< frame type [0-2] of the previous frame
231 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
232 ///< << 16) / #MAX_FRAMESIZE
233 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
235 int aw_idx_is_ext; ///< whether the AW index was encoded in
236 ///< 8 bits (instead of 6)
237 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
238 ///< can apply the pulse, relative to the
239 ///< value in aw_first_pulse_off. The exact
240 ///< position of the first AW-pulse is within
241 ///< [pulse_off, pulse_off + this], and
242 ///< depends on bitstream values; [16 or 24]
243 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
244 ///< that this number can be negative (in
245 ///< which case it basically means "zero")
246 int aw_first_pulse_off[2]; ///< index of first sample to which to
247 ///< apply AW-pulses, or -0xff if unset
248 int aw_next_pulse_off_cache; ///< the position (relative to start of the
249 ///< second block) at which pulses should
250 ///< start to be positioned, serves as a
251 ///< cache for pitch-adaptive window pulses
254 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
255 ///< only used for comfort noise in #pRNG()
256 float gain_pred_err[6]; ///< cache for gain prediction
257 float excitation_history[MAX_SIGNAL_HISTORY];
258 ///< cache of the signal of previous
259 ///< superframes, used as a history for
260 ///< signal generation
261 float synth_history[MAX_LSPS]; ///< see #excitation_history
265 * @name Postfilter values
267 * Variables used for postfilter implementation, mostly history for
268 * smoothing and so on, and context variables for FFT/iFFT.
271 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
272 ///< postfilter (for denoise filter)
273 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
274 ///< transform, part of postfilter)
275 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
277 float postfilter_agc; ///< gain control memory, used in
278 ///< #adaptive_gain_control()
279 float dcf_mem[2]; ///< DC filter history
280 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
281 ///< zero filter output (i.e. excitation)
283 float denoise_filter_cache[MAX_FRAMESIZE];
284 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
285 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
286 ///< aligned buffer for LPC tilting
287 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
288 ///< aligned buffer for denoise coefficients
289 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
290 ///< aligned buffer for postfilter speech
298 * Set up the variable bit mode (VBM) tree from container extradata.
299 * @param gb bit I/O context.
300 * The bit context (s->gb) should be loaded with byte 23-46 of the
301 * container extradata (i.e. the ones containing the VBM tree).
302 * @param vbm_tree pointer to array to which the decoded VBM tree will be
304 * @return 0 on success, <0 on error.
306 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
308 static const uint8_t bits[] = {
311 10, 10, 10, 12, 12, 12,
314 static const uint16_t codes[] = {
315 0x0000, 0x0001, 0x0002, // 00/01/10
316 0x000c, 0x000d, 0x000e, // 11+00/01/10
317 0x003c, 0x003d, 0x003e, // 1111+00/01/10
318 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
319 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
320 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
321 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
323 int cntr[8] = { 0 }, n, res;
325 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
326 for (n = 0; n < 17; n++) {
327 res = get_bits(gb, 3);
328 if (cntr[res] > 3) // should be >= 3 + (res == 7))
330 vbm_tree[res * 3 + cntr[res]++] = n;
332 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
333 bits, 1, 1, codes, 2, 2, 132);
338 * Set up decoder with parameters from demuxer (extradata etc.).
340 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
342 int n, flags, pitch_range, lsp16_flag;
343 WMAVoiceContext *s = ctx->priv_data;
347 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
348 * - byte 19-22: flags field (annoyingly in LE; see below for known
350 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
353 if (ctx->extradata_size != 46) {
354 av_log(ctx, AV_LOG_ERROR,
355 "Invalid extradata size %d (should be 46)\n",
356 ctx->extradata_size);
359 flags = AV_RL32(ctx->extradata + 18);
360 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
361 s->do_apf = flags & 0x1;
363 ff_rdft_init(&s->rdft, 7, DFT_R2C);
364 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
365 ff_dct_init(&s->dct, 6, DCT_I);
366 ff_dct_init(&s->dst, 6, DST_I);
368 ff_sine_window_init(s->cos, 256);
369 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
370 for (n = 0; n < 255; n++) {
371 s->sin[n] = -s->sin[510 - n];
372 s->cos[510 - n] = s->cos[n];
375 s->denoise_strength = (flags >> 2) & 0xF;
376 if (s->denoise_strength >= 12) {
377 av_log(ctx, AV_LOG_ERROR,
378 "Invalid denoise filter strength %d (max=11)\n",
379 s->denoise_strength);
382 s->denoise_tilt_corr = !!(flags & 0x40);
383 s->dc_level = (flags >> 7) & 0xF;
384 s->lsp_q_mode = !!(flags & 0x2000);
385 s->lsp_def_mode = !!(flags & 0x4000);
386 lsp16_flag = flags & 0x1000;
389 s->frame_lsp_bitsize = 34;
390 s->sframe_lsp_bitsize = 60;
393 s->frame_lsp_bitsize = 24;
394 s->sframe_lsp_bitsize = 48;
396 for (n = 0; n < s->lsps; n++)
397 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
399 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
400 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
401 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
405 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
406 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
407 pitch_range = s->max_pitch_val - s->min_pitch_val;
408 if (pitch_range <= 0) {
409 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
412 s->pitch_nbits = av_ceil_log2(pitch_range);
413 s->last_pitch_val = 40;
414 s->last_acb_type = ACB_TYPE_NONE;
415 s->history_nsamples = s->max_pitch_val + 8;
417 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
418 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
419 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
421 av_log(ctx, AV_LOG_ERROR,
422 "Unsupported samplerate %d (min=%d, max=%d)\n",
423 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
428 s->block_conv_table[0] = s->min_pitch_val;
429 s->block_conv_table[1] = (pitch_range * 25) >> 6;
430 s->block_conv_table[2] = (pitch_range * 44) >> 6;
431 s->block_conv_table[3] = s->max_pitch_val - 1;
432 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
433 if (s->block_delta_pitch_hrange <= 0) {
434 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
437 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
438 s->block_pitch_range = s->block_conv_table[2] +
439 s->block_conv_table[3] + 1 +
440 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
441 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
444 ctx->channel_layout = AV_CH_LAYOUT_MONO;
445 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
451 * @name Postfilter functions
452 * Postfilter functions (gain control, wiener denoise filter, DC filter,
453 * kalman smoothening, plus surrounding code to wrap it)
457 * Adaptive gain control (as used in postfilter).
459 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
460 * that the energy here is calculated using sum(abs(...)), whereas the
461 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
463 * @param out output buffer for filtered samples
464 * @param in input buffer containing the samples as they are after the
465 * postfilter steps so far
466 * @param speech_synth input buffer containing speech synth before postfilter
467 * @param size input buffer size
468 * @param alpha exponential filter factor
469 * @param gain_mem pointer to filter memory (single float)
471 static void adaptive_gain_control(float *out, const float *in,
472 const float *speech_synth,
473 int size, float alpha, float *gain_mem)
476 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
477 float mem = *gain_mem;
479 for (i = 0; i < size; i++) {
480 speech_energy += fabsf(speech_synth[i]);
481 postfilter_energy += fabsf(in[i]);
483 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
485 for (i = 0; i < size; i++) {
486 mem = alpha * mem + gain_scale_factor;
487 out[i] = in[i] * mem;
494 * Kalman smoothing function.
496 * This function looks back pitch +/- 3 samples back into history to find
497 * the best fitting curve (that one giving the optimal gain of the two
498 * signals, i.e. the highest dot product between the two), and then
499 * uses that signal history to smoothen the output of the speech synthesis
502 * @param s WMA Voice decoding context
503 * @param pitch pitch of the speech signal
504 * @param in input speech signal
505 * @param out output pointer for smoothened signal
506 * @param size input/output buffer size
508 * @returns -1 if no smoothening took place, e.g. because no optimal
509 * fit could be found, or 0 on success.
511 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
512 const float *in, float *out, int size)
515 float optimal_gain = 0, dot;
516 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
517 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
520 /* find best fitting point in history */
522 dot = avpriv_scalarproduct_float_c(in, ptr, size);
523 if (dot > optimal_gain) {
527 } while (--ptr >= end);
529 if (optimal_gain <= 0)
531 dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
532 if (dot <= 0) // would be 1.0
535 if (optimal_gain <= dot) {
536 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
540 /* actual smoothing */
541 for (n = 0; n < size; n++)
542 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
548 * Get the tilt factor of a formant filter from its transfer function
549 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
550 * but somehow (??) it does a speech synthesis filter in the
551 * middle, which is missing here
553 * @param lpcs LPC coefficients
554 * @param n_lpcs Size of LPC buffer
555 * @returns the tilt factor
557 static float tilt_factor(const float *lpcs, int n_lpcs)
561 rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
562 rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
568 * Derive denoise filter coefficients (in real domain) from the LPCs.
570 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
571 int fcb_type, float *coeffs, int remainder)
573 float last_coeff, min = 15.0, max = -15.0;
574 float irange, angle_mul, gain_mul, range, sq;
577 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
578 s->rdft.rdft_calc(&s->rdft, lpcs);
579 #define log_range(var, assign) do { \
580 float tmp = log10f(assign); var = tmp; \
581 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
583 log_range(last_coeff, lpcs[1] * lpcs[1]);
584 for (n = 1; n < 64; n++)
585 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
586 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
587 log_range(lpcs[0], lpcs[0] * lpcs[0]);
590 lpcs[64] = last_coeff;
592 /* Now, use this spectrum to pick out these frequencies with higher
593 * (relative) power/energy (which we then take to be "not noise"),
594 * and set up a table (still in lpc[]) of (relative) gains per frequency.
595 * These frequencies will be maintained, while others ("noise") will be
596 * decreased in the filter output. */
597 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
598 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
600 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
601 for (n = 0; n <= 64; n++) {
604 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
605 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
606 lpcs[n] = angle_mul * pwr;
608 /* 70.57 =~ 1/log10(1.0331663) */
609 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
610 if (idx > 127) { // fallback if index falls outside table range
611 coeffs[n] = wmavoice_energy_table[127] *
612 powf(1.0331663, idx - 127);
614 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
617 /* calculate the Hilbert transform of the gains, which we do (since this
618 * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
619 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
620 * "moment" of the LPCs in this filter. */
621 s->dct.dct_calc(&s->dct, lpcs);
622 s->dst.dct_calc(&s->dst, lpcs);
624 /* Split out the coefficient indexes into phase/magnitude pairs */
625 idx = 255 + av_clip(lpcs[64], -255, 255);
626 coeffs[0] = coeffs[0] * s->cos[idx];
627 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
628 last_coeff = coeffs[64] * s->cos[idx];
630 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
631 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
632 coeffs[n * 2] = coeffs[n] * s->cos[idx];
636 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
637 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
638 coeffs[n * 2] = coeffs[n] * s->cos[idx];
640 coeffs[1] = last_coeff;
642 /* move into real domain */
643 s->irdft.rdft_calc(&s->irdft, coeffs);
645 /* tilt correction and normalize scale */
646 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
647 if (s->denoise_tilt_corr) {
650 coeffs[remainder - 1] = 0;
651 ff_tilt_compensation(&tilt_mem,
652 -1.8 * tilt_factor(coeffs, remainder - 1),
655 sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
657 for (n = 0; n < remainder; n++)
662 * This function applies a Wiener filter on the (noisy) speech signal as
663 * a means to denoise it.
665 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
666 * - using this power spectrum, calculate (for each frequency) the Wiener
667 * filter gain, which depends on the frequency power and desired level
668 * of noise subtraction (when set too high, this leads to artifacts)
669 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
671 * - by doing a phase shift, calculate the Hilbert transform of this array
672 * of per-frequency filter-gains to get the filtering coefficients;
673 * - smoothen/normalize/de-tilt these filter coefficients as desired;
674 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
675 * to get the denoised speech signal;
676 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
677 * the frame boundary) are saved and applied to subsequent frames by an
678 * overlap-add method (otherwise you get clicking-artifacts).
680 * @param s WMA Voice decoding context
681 * @param fcb_type Frame (codebook) type
682 * @param synth_pf input: the noisy speech signal, output: denoised speech
683 * data; should be 16-byte aligned (for ASM purposes)
684 * @param size size of the speech data
685 * @param lpcs LPCs used to synthesize this frame's speech data
687 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
688 float *synth_pf, int size,
691 int remainder, lim, n;
693 if (fcb_type != FCB_TYPE_SILENCE) {
694 float *tilted_lpcs = s->tilted_lpcs_pf,
695 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
697 tilted_lpcs[0] = 1.0;
698 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
699 memset(&tilted_lpcs[s->lsps + 1], 0,
700 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
701 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
702 tilted_lpcs, s->lsps + 2);
704 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
705 * size is applied to the next frame. All input beyond this is zero,
706 * and thus all output beyond this will go towards zero, hence we can
707 * limit to min(size-1, 127-size) as a performance consideration. */
708 remainder = FFMIN(127 - size, size - 1);
709 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
711 /* apply coefficients (in frequency spectrum domain), i.e. complex
712 * number multiplication */
713 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
714 s->rdft.rdft_calc(&s->rdft, synth_pf);
715 s->rdft.rdft_calc(&s->rdft, coeffs);
716 synth_pf[0] *= coeffs[0];
717 synth_pf[1] *= coeffs[1];
718 for (n = 1; n < 64; n++) {
719 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
720 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
721 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
723 s->irdft.rdft_calc(&s->irdft, synth_pf);
726 /* merge filter output with the history of previous runs */
727 if (s->denoise_filter_cache_size) {
728 lim = FFMIN(s->denoise_filter_cache_size, size);
729 for (n = 0; n < lim; n++)
730 synth_pf[n] += s->denoise_filter_cache[n];
731 s->denoise_filter_cache_size -= lim;
732 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
733 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
736 /* move remainder of filter output into a cache for future runs */
737 if (fcb_type != FCB_TYPE_SILENCE) {
738 lim = FFMIN(remainder, s->denoise_filter_cache_size);
739 for (n = 0; n < lim; n++)
740 s->denoise_filter_cache[n] += synth_pf[size + n];
741 if (lim < remainder) {
742 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
743 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
744 s->denoise_filter_cache_size = remainder;
750 * Averaging projection filter, the postfilter used in WMAVoice.
752 * This uses the following steps:
753 * - A zero-synthesis filter (generate excitation from synth signal)
754 * - Kalman smoothing on excitation, based on pitch
755 * - Re-synthesized smoothened output
756 * - Iterative Wiener denoise filter
757 * - Adaptive gain filter
760 * @param s WMAVoice decoding context
761 * @param synth Speech synthesis output (before postfilter)
762 * @param samples Output buffer for filtered samples
763 * @param size Buffer size of synth & samples
764 * @param lpcs Generated LPCs used for speech synthesis
765 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
766 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
767 * @param pitch Pitch of the input signal
769 static void postfilter(WMAVoiceContext *s, const float *synth,
770 float *samples, int size,
771 const float *lpcs, float *zero_exc_pf,
772 int fcb_type, int pitch)
774 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
775 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
776 *synth_filter_in = zero_exc_pf;
778 assert(size <= MAX_FRAMESIZE / 2);
780 /* generate excitation from input signal */
781 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
783 if (fcb_type >= FCB_TYPE_AW_PULSES &&
784 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
785 synth_filter_in = synth_filter_in_buf;
787 /* re-synthesize speech after smoothening, and keep history */
788 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
789 synth_filter_in, size, s->lsps);
790 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
791 sizeof(synth_pf[0]) * s->lsps);
793 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
795 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
798 if (s->dc_level > 8) {
799 /* remove ultra-low frequency DC noise / highpass filter;
800 * coefficients are identical to those used in SIPR decoding,
801 * and very closely resemble those used in AMR-NB decoding. */
802 ff_acelp_apply_order_2_transfer_function(samples, samples,
803 (const float[2]) { -1.99997, 1.0 },
804 (const float[2]) { -1.9330735188, 0.93589198496 },
805 0.93980580475, s->dcf_mem, size);
814 * @param lsps output pointer to the array that will hold the LSPs
815 * @param num number of LSPs to be dequantized
816 * @param values quantized values, contains n_stages values
817 * @param sizes range (i.e. max value) of each quantized value
818 * @param n_stages number of dequantization runs
819 * @param table dequantization table to be used
820 * @param mul_q LSF multiplier
821 * @param base_q base (lowest) LSF values
823 static void dequant_lsps(double *lsps, int num,
824 const uint16_t *values,
825 const uint16_t *sizes,
826 int n_stages, const uint8_t *table,
828 const double *base_q)
832 memset(lsps, 0, num * sizeof(*lsps));
833 for (n = 0; n < n_stages; n++) {
834 const uint8_t *t_off = &table[values[n] * num];
835 double base = base_q[n], mul = mul_q[n];
837 for (m = 0; m < num; m++)
838 lsps[m] += base + mul * t_off[m];
840 table += sizes[n] * num;
845 * @name LSP dequantization routines
846 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
847 * @note we assume enough bits are available, caller should check.
848 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
849 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
853 * Parse 10 independently-coded LSPs.
855 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
857 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
858 static const double mul_lsf[4] = {
859 5.2187144800e-3, 1.4626986422e-3,
860 9.6179549166e-4, 1.1325736225e-3
862 static const double base_lsf[4] = {
863 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
864 M_PI * -3.3486e-2, M_PI * -5.7408e-2
868 v[0] = get_bits(gb, 8);
869 v[1] = get_bits(gb, 6);
870 v[2] = get_bits(gb, 5);
871 v[3] = get_bits(gb, 5);
873 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
878 * Parse 10 independently-coded LSPs, and then derive the tables to
879 * generate LSPs for the other frames from them (residual coding).
881 static void dequant_lsp10r(GetBitContext *gb,
882 double *i_lsps, const double *old,
883 double *a1, double *a2, int q_mode)
885 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
886 static const double mul_lsf[3] = {
887 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
889 static const double base_lsf[3] = {
890 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
892 const float (*ipol_tab)[2][10] = q_mode ?
893 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
894 uint16_t interpol, v[3];
897 dequant_lsp10i(gb, i_lsps);
899 interpol = get_bits(gb, 5);
900 v[0] = get_bits(gb, 7);
901 v[1] = get_bits(gb, 6);
902 v[2] = get_bits(gb, 6);
904 for (n = 0; n < 10; n++) {
905 double delta = old[n] - i_lsps[n];
906 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
907 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
910 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
915 * Parse 16 independently-coded LSPs.
917 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
919 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
920 static const double mul_lsf[5] = {
921 3.3439586280e-3, 6.9908173703e-4,
922 3.3216608306e-3, 1.0334960326e-3,
925 static const double base_lsf[5] = {
926 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
927 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
932 v[0] = get_bits(gb, 8);
933 v[1] = get_bits(gb, 6);
934 v[2] = get_bits(gb, 7);
935 v[3] = get_bits(gb, 6);
936 v[4] = get_bits(gb, 7);
938 dequant_lsps( lsps, 5, v, vec_sizes, 2,
939 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
940 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
941 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
942 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
943 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
947 * Parse 16 independently-coded LSPs, and then derive the tables to
948 * generate LSPs for the other frames from them (residual coding).
950 static void dequant_lsp16r(GetBitContext *gb,
951 double *i_lsps, const double *old,
952 double *a1, double *a2, int q_mode)
954 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
955 static const double mul_lsf[3] = {
956 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
958 static const double base_lsf[3] = {
959 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
961 const float (*ipol_tab)[2][16] = q_mode ?
962 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
963 uint16_t interpol, v[3];
966 dequant_lsp16i(gb, i_lsps);
968 interpol = get_bits(gb, 5);
969 v[0] = get_bits(gb, 7);
970 v[1] = get_bits(gb, 7);
971 v[2] = get_bits(gb, 7);
973 for (n = 0; n < 16; n++) {
974 double delta = old[n] - i_lsps[n];
975 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
976 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
979 dequant_lsps( a2, 10, v, vec_sizes, 1,
980 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
981 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
982 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
983 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
984 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
989 * @name Pitch-adaptive window coding functions
990 * The next few functions are for pitch-adaptive window coding.
994 * Parse the offset of the first pitch-adaptive window pulses, and
995 * the distribution of pulses between the two blocks in this frame.
996 * @param s WMA Voice decoding context private data
997 * @param gb bit I/O context
998 * @param pitch pitch for each block in this frame
1000 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1003 static const int16_t start_offset[94] = {
1004 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1005 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1006 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1007 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1008 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1009 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1010 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1011 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1015 /* position of pulse */
1016 s->aw_idx_is_ext = 0;
1017 if ((bits = get_bits(gb, 6)) >= 54) {
1018 s->aw_idx_is_ext = 1;
1019 bits += (bits - 54) * 3 + get_bits(gb, 2);
1022 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1023 * the distribution of the pulses in each block contained in this frame. */
1024 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1025 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1026 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1027 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1028 offset += s->aw_n_pulses[0] * pitch[0];
1029 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1030 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1032 /* if continuing from a position before the block, reset position to
1033 * start of block (when corrected for the range over which it can be
1034 * spread in aw_pulse_set1()). */
1035 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1036 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1037 s->aw_first_pulse_off[1] -= pitch[1];
1038 if (start_offset[bits] < 0)
1039 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1040 s->aw_first_pulse_off[0] -= pitch[0];
1045 * Apply second set of pitch-adaptive window pulses.
1046 * @param s WMA Voice decoding context private data
1047 * @param gb bit I/O context
1048 * @param block_idx block index in frame [0, 1]
1049 * @param fcb structure containing fixed codebook vector info
1051 static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1052 int block_idx, AMRFixed *fcb)
1054 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1055 uint16_t *use_mask = use_mask_mem + 2;
1056 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1057 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1058 * of idx are the position of the bit within a particular item in the
1059 * array (0 being the most significant bit, and 15 being the least
1060 * significant bit), and the remainder (>> 4) is the index in the
1061 * use_mask[]-array. This is faster and uses less memory than using a
1062 * 80-byte/80-int array. */
1063 int pulse_off = s->aw_first_pulse_off[block_idx],
1064 pulse_start, n, idx, range, aidx, start_off = 0;
1066 /* set offset of first pulse to within this block */
1067 if (s->aw_n_pulses[block_idx] > 0)
1068 while (pulse_off + s->aw_pulse_range < 1)
1069 pulse_off += fcb->pitch_lag;
1071 /* find range per pulse */
1072 if (s->aw_n_pulses[0] > 0) {
1073 if (block_idx == 0) {
1075 } else /* block_idx = 1 */ {
1077 if (s->aw_n_pulses[block_idx] > 0)
1078 pulse_off = s->aw_next_pulse_off_cache;
1082 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1084 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1085 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1086 * we exclude that range from being pulsed again in this function. */
1087 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1088 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1089 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1090 if (s->aw_n_pulses[block_idx] > 0)
1091 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1092 int excl_range = s->aw_pulse_range; // always 16 or 24
1093 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1094 int first_sh = 16 - (idx & 15);
1095 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1096 excl_range -= first_sh;
1097 if (excl_range >= 16) {
1098 *use_mask_ptr++ = 0;
1099 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1101 *use_mask_ptr &= 0xFFFF >> excl_range;
1104 /* find the 'aidx'th offset that is not excluded */
1105 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1106 for (n = 0; n <= aidx; pulse_start++) {
1107 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1108 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1109 if (use_mask[0]) idx = 0x0F;
1110 else if (use_mask[1]) idx = 0x1F;
1111 else if (use_mask[2]) idx = 0x2F;
1112 else if (use_mask[3]) idx = 0x3F;
1113 else if (use_mask[4]) idx = 0x4F;
1115 idx -= av_log2_16bit(use_mask[idx >> 4]);
1117 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1118 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1124 fcb->x[fcb->n] = start_off;
1125 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1128 /* set offset for next block, relative to start of that block */
1129 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1130 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1134 * Apply first set of pitch-adaptive window pulses.
1135 * @param s WMA Voice decoding context private data
1136 * @param gb bit I/O context
1137 * @param block_idx block index in frame [0, 1]
1138 * @param fcb storage location for fixed codebook pulse info
1140 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1141 int block_idx, AMRFixed *fcb)
1143 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1146 if (s->aw_n_pulses[block_idx] > 0) {
1147 int n, v_mask, i_mask, sh, n_pulses;
1149 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1154 } else { // 4 pulses, 1:sign + 2:index each
1161 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1162 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1163 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1164 s->aw_first_pulse_off[block_idx];
1165 while (fcb->x[fcb->n] < 0)
1166 fcb->x[fcb->n] += fcb->pitch_lag;
1167 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1171 int num2 = (val & 0x1FF) >> 1, delta, idx;
1173 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1174 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1175 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1176 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1177 v = (val & 0x200) ? -1.0 : 1.0;
1179 fcb->no_repeat_mask |= 3 << fcb->n;
1180 fcb->x[fcb->n] = idx - delta;
1182 fcb->x[fcb->n + 1] = idx;
1183 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1191 * Generate a random number from frame_cntr and block_idx, which will lief
1192 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1193 * table of size 1000 of which you want to read block_size entries).
1195 * @param frame_cntr current frame number
1196 * @param block_num current block index
1197 * @param block_size amount of entries we want to read from a table
1198 * that has 1000 entries
1199 * @return a (non-)random number in the [0, 1000 - block_size] range.
1201 static int pRNG(int frame_cntr, int block_num, int block_size)
1203 /* array to simplify the calculation of z:
1204 * y = (x % 9) * 5 + 6;
1205 * z = (49995 * x) / y;
1206 * Since y only has 9 values, we can remove the division by using a
1207 * LUT and using FASTDIV-style divisions. For each of the 9 values
1208 * of y, we can rewrite z as:
1209 * z = x * (49995 / y) + x * ((49995 % y) / y)
1210 * In this table, each col represents one possible value of y, the
1211 * first number is 49995 / y, and the second is the FASTDIV variant
1212 * of 49995 % y / y. */
1213 static const unsigned int div_tbl[9][2] = {
1214 { 8332, 3 * 715827883U }, // y = 6
1215 { 4545, 0 * 390451573U }, // y = 11
1216 { 3124, 11 * 268435456U }, // y = 16
1217 { 2380, 15 * 204522253U }, // y = 21
1218 { 1922, 23 * 165191050U }, // y = 26
1219 { 1612, 23 * 138547333U }, // y = 31
1220 { 1388, 27 * 119304648U }, // y = 36
1221 { 1219, 16 * 104755300U }, // y = 41
1222 { 1086, 39 * 93368855U } // y = 46
1224 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1225 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1226 // so this is effectively a modulo (%)
1227 y = x - 9 * MULH(477218589, x); // x % 9
1228 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1229 // z = x * 49995 / (y * 5 + 6)
1230 return z % (1000 - block_size);
1234 * Parse hardcoded signal for a single block.
1235 * @note see #synth_block().
1237 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1238 int block_idx, int size,
1239 const struct frame_type_desc *frame_desc,
1245 assert(size <= MAX_FRAMESIZE);
1247 /* Set the offset from which we start reading wmavoice_std_codebook */
1248 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1249 r_idx = pRNG(s->frame_cntr, block_idx, size);
1250 gain = s->silence_gain;
1251 } else /* FCB_TYPE_HARDCODED */ {
1252 r_idx = get_bits(gb, 8);
1253 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1256 /* Clear gain prediction parameters */
1257 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1259 /* Apply gain to hardcoded codebook and use that as excitation signal */
1260 for (n = 0; n < size; n++)
1261 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1265 * Parse FCB/ACB signal for a single block.
1266 * @note see #synth_block().
1268 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1269 int block_idx, int size,
1270 int block_pitch_sh2,
1271 const struct frame_type_desc *frame_desc,
1274 static const float gain_coeff[6] = {
1275 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1277 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1278 int n, idx, gain_weight;
1281 assert(size <= MAX_FRAMESIZE / 2);
1282 memset(pulses, 0, sizeof(*pulses) * size);
1284 fcb.pitch_lag = block_pitch_sh2 >> 2;
1285 fcb.pitch_fac = 1.0;
1286 fcb.no_repeat_mask = 0;
1289 /* For the other frame types, this is where we apply the innovation
1290 * (fixed) codebook pulses of the speech signal. */
1291 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1292 aw_pulse_set1(s, gb, block_idx, &fcb);
1293 aw_pulse_set2(s, gb, block_idx, &fcb);
1294 } else /* FCB_TYPE_EXC_PULSES */ {
1295 int offset_nbits = 5 - frame_desc->log_n_blocks;
1297 fcb.no_repeat_mask = -1;
1298 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1299 * (instead of double) for a subset of pulses */
1300 for (n = 0; n < 5; n++) {
1304 sign = get_bits1(gb) ? 1.0 : -1.0;
1305 pos1 = get_bits(gb, offset_nbits);
1306 fcb.x[fcb.n] = n + 5 * pos1;
1307 fcb.y[fcb.n++] = sign;
1308 if (n < frame_desc->dbl_pulses) {
1309 pos2 = get_bits(gb, offset_nbits);
1310 fcb.x[fcb.n] = n + 5 * pos2;
1311 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1315 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1317 /* Calculate gain for adaptive & fixed codebook signal.
1318 * see ff_amr_set_fixed_gain(). */
1319 idx = get_bits(gb, 7);
1320 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1322 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1323 acb_gain = wmavoice_gain_codebook_acb[idx];
1324 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1325 -2.9957322736 /* log(0.05) */,
1326 1.6094379124 /* log(5.0) */);
1328 gain_weight = 8 >> frame_desc->log_n_blocks;
1329 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1330 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1331 for (n = 0; n < gain_weight; n++)
1332 s->gain_pred_err[n] = pred_err;
1334 /* Calculation of adaptive codebook */
1335 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1337 for (n = 0; n < size; n += len) {
1339 int abs_idx = block_idx * size + n;
1340 int pitch_sh16 = (s->last_pitch_val << 16) +
1341 s->pitch_diff_sh16 * abs_idx;
1342 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1343 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1344 idx = idx_sh16 >> 16;
1345 if (s->pitch_diff_sh16) {
1346 if (s->pitch_diff_sh16 > 0) {
1347 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1349 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1350 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1355 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1356 wmavoice_ipol1_coeffs, 17,
1359 } else /* ACB_TYPE_HAMMING */ {
1360 int block_pitch = block_pitch_sh2 >> 2;
1361 idx = block_pitch_sh2 & 3;
1363 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1364 wmavoice_ipol2_coeffs, 4,
1367 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1368 sizeof(float) * size);
1371 /* Interpolate ACB/FCB and use as excitation signal */
1372 ff_weighted_vector_sumf(excitation, excitation, pulses,
1373 acb_gain, fcb_gain, size);
1377 * Parse data in a single block.
1378 * @note we assume enough bits are available, caller should check.
1380 * @param s WMA Voice decoding context private data
1381 * @param gb bit I/O context
1382 * @param block_idx index of the to-be-read block
1383 * @param size amount of samples to be read in this block
1384 * @param block_pitch_sh2 pitch for this block << 2
1385 * @param lsps LSPs for (the end of) this frame
1386 * @param prev_lsps LSPs for the last frame
1387 * @param frame_desc frame type descriptor
1388 * @param excitation target memory for the ACB+FCB interpolated signal
1389 * @param synth target memory for the speech synthesis filter output
1390 * @return 0 on success, <0 on error.
1392 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1393 int block_idx, int size,
1394 int block_pitch_sh2,
1395 const double *lsps, const double *prev_lsps,
1396 const struct frame_type_desc *frame_desc,
1397 float *excitation, float *synth)
1399 double i_lsps[MAX_LSPS];
1400 float lpcs[MAX_LSPS];
1404 if (frame_desc->acb_type == ACB_TYPE_NONE)
1405 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1407 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1408 frame_desc, excitation);
1410 /* convert interpolated LSPs to LPCs */
1411 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1412 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1413 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1414 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1416 /* Speech synthesis */
1417 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1421 * Synthesize output samples for a single frame.
1422 * @note we assume enough bits are available, caller should check.
1424 * @param ctx WMA Voice decoder context
1425 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1426 * @param frame_idx Frame number within superframe [0-2]
1427 * @param samples pointer to output sample buffer, has space for at least 160
1429 * @param lsps LSP array
1430 * @param prev_lsps array of previous frame's LSPs
1431 * @param excitation target buffer for excitation signal
1432 * @param synth target buffer for synthesized speech data
1433 * @return 0 on success, <0 on error.
1435 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1437 const double *lsps, const double *prev_lsps,
1438 float *excitation, float *synth)
1440 WMAVoiceContext *s = ctx->priv_data;
1441 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1442 int pitch[MAX_BLOCKS], last_block_pitch;
1444 /* Parse frame type ("frame header"), see frame_descs */
1445 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1448 av_log(ctx, AV_LOG_ERROR,
1449 "Invalid frame type VLC code, skipping\n");
1453 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1455 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1456 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1457 /* Pitch is provided per frame, which is interpreted as the pitch of
1458 * the last sample of the last block of this frame. We can interpolate
1459 * the pitch of other blocks (and even pitch-per-sample) by gradually
1460 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1461 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1462 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1463 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1464 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1465 if (s->last_acb_type == ACB_TYPE_NONE ||
1466 20 * abs(cur_pitch_val - s->last_pitch_val) >
1467 (cur_pitch_val + s->last_pitch_val))
1468 s->last_pitch_val = cur_pitch_val;
1470 /* pitch per block */
1471 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1472 int fac = n * 2 + 1;
1474 pitch[n] = (MUL16(fac, cur_pitch_val) +
1475 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1476 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1479 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1480 s->pitch_diff_sh16 =
1481 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1484 /* Global gain (if silence) and pitch-adaptive window coordinates */
1485 switch (frame_descs[bd_idx].fcb_type) {
1486 case FCB_TYPE_SILENCE:
1487 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1489 case FCB_TYPE_AW_PULSES:
1490 aw_parse_coords(s, gb, pitch);
1494 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1497 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1498 switch (frame_descs[bd_idx].acb_type) {
1499 case ACB_TYPE_HAMMING: {
1500 /* Pitch is given per block. Per-block pitches are encoded as an
1501 * absolute value for the first block, and then delta values
1502 * relative to this value) for all subsequent blocks. The scale of
1503 * this pitch value is semi-logaritmic compared to its use in the
1504 * decoder, so we convert it to normal scale also. */
1506 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1507 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1508 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1511 block_pitch = get_bits(gb, s->block_pitch_nbits);
1513 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1514 get_bits(gb, s->block_delta_pitch_nbits);
1515 /* Convert last_ so that any next delta is within _range */
1516 last_block_pitch = av_clip(block_pitch,
1517 s->block_delta_pitch_hrange,
1518 s->block_pitch_range -
1519 s->block_delta_pitch_hrange);
1521 /* Convert semi-log-style scale back to normal scale */
1522 if (block_pitch < t1) {
1523 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1526 if (block_pitch < t2) {
1528 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1531 if (block_pitch < t3) {
1533 (s->block_conv_table[2] + block_pitch) << 2;
1535 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1538 pitch[n] = bl_pitch_sh2 >> 2;
1542 case ACB_TYPE_ASYMMETRIC: {
1543 bl_pitch_sh2 = pitch[n] << 2;
1547 default: // ACB_TYPE_NONE has no pitch
1552 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1553 lsps, prev_lsps, &frame_descs[bd_idx],
1554 &excitation[n * block_nsamples],
1555 &synth[n * block_nsamples]);
1558 /* Averaging projection filter, if applicable. Else, just copy samples
1559 * from synthesis buffer */
1561 double i_lsps[MAX_LSPS];
1562 float lpcs[MAX_LSPS];
1564 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1565 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1566 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1567 postfilter(s, synth, samples, 80, lpcs,
1568 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1569 frame_descs[bd_idx].fcb_type, pitch[0]);
1571 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1572 i_lsps[n] = cos(lsps[n]);
1573 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1574 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1575 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1576 frame_descs[bd_idx].fcb_type, pitch[0]);
1578 memcpy(samples, synth, 160 * sizeof(synth[0]));
1580 /* Cache values for next frame */
1582 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1583 s->last_acb_type = frame_descs[bd_idx].acb_type;
1584 switch (frame_descs[bd_idx].acb_type) {
1586 s->last_pitch_val = 0;
1588 case ACB_TYPE_ASYMMETRIC:
1589 s->last_pitch_val = cur_pitch_val;
1591 case ACB_TYPE_HAMMING:
1592 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1600 * Ensure minimum value for first item, maximum value for last value,
1601 * proper spacing between each value and proper ordering.
1603 * @param lsps array of LSPs
1604 * @param num size of LSP array
1606 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1607 * useful to put in a generic location later on. Parts are also
1608 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1609 * which is in float.
1611 static void stabilize_lsps(double *lsps, int num)
1615 /* set minimum value for first, maximum value for last and minimum
1616 * spacing between LSF values.
1617 * Very similar to ff_set_min_dist_lsf(), but in double. */
1618 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1619 for (n = 1; n < num; n++)
1620 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1621 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1623 /* reorder (looks like one-time / non-recursed bubblesort).
1624 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1625 for (n = 1; n < num; n++) {
1626 if (lsps[n] < lsps[n - 1]) {
1627 for (m = 1; m < num; m++) {
1628 double tmp = lsps[m];
1629 for (l = m - 1; l >= 0; l--) {
1630 if (lsps[l] <= tmp) break;
1631 lsps[l + 1] = lsps[l];
1641 * Test if there's enough bits to read 1 superframe.
1643 * @param orig_gb bit I/O context used for reading. This function
1644 * does not modify the state of the bitreader; it
1645 * only uses it to copy the current stream position
1646 * @param s WMA Voice decoding context private data
1647 * @return -1 if unsupported, 1 on not enough bits or 0 if OK.
1649 static int check_bits_for_superframe(GetBitContext *orig_gb,
1652 GetBitContext s_gb, *gb = &s_gb;
1653 int n, need_bits, bd_idx;
1654 const struct frame_type_desc *frame_desc;
1656 /* initialize a copy */
1657 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1658 skip_bits_long(gb, get_bits_count(orig_gb));
1659 assert(get_bits_left(gb) == get_bits_left(orig_gb));
1661 /* superframe header */
1662 if (get_bits_left(gb) < 14)
1665 return -1; // WMAPro-in-WMAVoice superframe
1666 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1667 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1668 if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1670 skip_bits_long(gb, s->sframe_lsp_bitsize);
1674 for (n = 0; n < MAX_FRAMES; n++) {
1675 int aw_idx_is_ext = 0;
1677 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1678 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1679 skip_bits_long(gb, s->frame_lsp_bitsize);
1681 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1683 return -1; // invalid frame type VLC code
1684 frame_desc = &frame_descs[bd_idx];
1685 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1686 if (get_bits_left(gb) < s->pitch_nbits)
1688 skip_bits_long(gb, s->pitch_nbits);
1690 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1692 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1693 int tmp = get_bits(gb, 6);
1701 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1702 need_bits = s->block_pitch_nbits +
1703 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1704 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1705 need_bits = 2 * !aw_idx_is_ext;
1708 need_bits += frame_desc->frame_size;
1709 if (get_bits_left(gb) < need_bits)
1711 skip_bits_long(gb, need_bits);
1718 * Synthesize output samples for a single superframe. If we have any data
1719 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1722 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1723 * to give a total of 480 samples per frame. See #synth_frame() for frame
1724 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1725 * (if these are globally specified for all frames (residually); they can
1726 * also be specified individually per-frame. See the s->has_residual_lsps
1727 * option), and can specify the number of samples encoded in this superframe
1728 * (if less than 480), usually used to prevent blanks at track boundaries.
1730 * @param ctx WMA Voice decoder context
1731 * @return 0 on success, <0 on error or 1 if there was not enough data to
1732 * fully parse the superframe
1734 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1737 WMAVoiceContext *s = ctx->priv_data;
1738 GetBitContext *gb = &s->gb, s_gb;
1739 int n, res, n_samples = 480;
1740 double lsps[MAX_FRAMES][MAX_LSPS];
1741 const double *mean_lsf = s->lsps == 16 ?
1742 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1743 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1744 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1747 memcpy(synth, s->synth_history,
1748 s->lsps * sizeof(*synth));
1749 memcpy(excitation, s->excitation_history,
1750 s->history_nsamples * sizeof(*excitation));
1752 if (s->sframe_cache_size > 0) {
1754 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1755 s->sframe_cache_size = 0;
1758 if ((res = check_bits_for_superframe(gb, s)) == 1) {
1763 /* First bit is speech/music bit, it differentiates between WMAVoice
1764 * speech samples (the actual codec) and WMAVoice music samples, which
1765 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1767 if (!get_bits1(gb)) {
1768 av_log_missing_feature(ctx, "WMAPro-in-WMAVoice", 1);
1769 return AVERROR_PATCHWELCOME;
1772 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1773 if (get_bits1(gb)) {
1774 if ((n_samples = get_bits(gb, 12)) > 480) {
1775 av_log(ctx, AV_LOG_ERROR,
1776 "Superframe encodes >480 samples (%d), not allowed\n",
1781 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1782 if (s->has_residual_lsps) {
1783 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1785 for (n = 0; n < s->lsps; n++)
1786 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1788 if (s->lsps == 10) {
1789 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1790 } else /* s->lsps == 16 */
1791 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1793 for (n = 0; n < s->lsps; n++) {
1794 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1795 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1796 lsps[2][n] += mean_lsf[n];
1798 for (n = 0; n < 3; n++)
1799 stabilize_lsps(lsps[n], s->lsps);
1802 /* get output buffer */
1803 frame->nb_samples = 480;
1804 if ((res = ff_get_buffer(ctx, frame, 0)) < 0) {
1805 av_log(ctx, AV_LOG_ERROR, "get_buffer() failed\n");
1808 frame->nb_samples = n_samples;
1809 samples = (float *)frame->data[0];
1811 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1812 for (n = 0; n < 3; n++) {
1813 if (!s->has_residual_lsps) {
1816 if (s->lsps == 10) {
1817 dequant_lsp10i(gb, lsps[n]);
1818 } else /* s->lsps == 16 */
1819 dequant_lsp16i(gb, lsps[n]);
1821 for (m = 0; m < s->lsps; m++)
1822 lsps[n][m] += mean_lsf[m];
1823 stabilize_lsps(lsps[n], s->lsps);
1826 if ((res = synth_frame(ctx, gb, n,
1827 &samples[n * MAX_FRAMESIZE],
1828 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1829 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1830 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1836 /* Statistics? FIXME - we don't check for length, a slight overrun
1837 * will be caught by internal buffer padding, and anything else
1838 * will be skipped, not read. */
1839 if (get_bits1(gb)) {
1840 res = get_bits(gb, 4);
1841 skip_bits(gb, 10 * (res + 1));
1846 /* Update history */
1847 memcpy(s->prev_lsps, lsps[2],
1848 s->lsps * sizeof(*s->prev_lsps));
1849 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1850 s->lsps * sizeof(*synth));
1851 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1852 s->history_nsamples * sizeof(*excitation));
1854 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1855 s->history_nsamples * sizeof(*s->zero_exc_pf));
1861 * Parse the packet header at the start of each packet (input data to this
1864 * @param s WMA Voice decoding context private data
1865 * @return 1 if not enough bits were available, or 0 on success.
1867 static int parse_packet_header(WMAVoiceContext *s)
1869 GetBitContext *gb = &s->gb;
1872 if (get_bits_left(gb) < 11)
1874 skip_bits(gb, 4); // packet sequence number
1875 s->has_residual_lsps = get_bits1(gb);
1877 res = get_bits(gb, 6); // number of superframes per packet
1878 // (minus first one if there is spillover)
1879 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1881 } while (res == 0x3F);
1882 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1888 * Copy (unaligned) bits from gb/data/size to pb.
1890 * @param pb target buffer to copy bits into
1891 * @param data source buffer to copy bits from
1892 * @param size size of the source data, in bytes
1893 * @param gb bit I/O context specifying the current position in the source.
1894 * data. This function might use this to align the bit position to
1895 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1897 * @param nbits the amount of bits to copy from source to target
1899 * @note after calling this function, the current position in the input bit
1900 * I/O context is undefined.
1902 static void copy_bits(PutBitContext *pb,
1903 const uint8_t *data, int size,
1904 GetBitContext *gb, int nbits)
1906 int rmn_bytes, rmn_bits;
1908 rmn_bits = rmn_bytes = get_bits_left(gb);
1909 if (rmn_bits < nbits)
1911 if (nbits > pb->size_in_bits - put_bits_count(pb))
1913 rmn_bits &= 7; rmn_bytes >>= 3;
1914 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1915 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1916 avpriv_copy_bits(pb, data + size - rmn_bytes,
1917 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1921 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1922 * and we expect that the demuxer / application provides it to us as such
1923 * (else you'll probably get garbage as output). Every packet has a size of
1924 * ctx->block_align bytes, starts with a packet header (see
1925 * #parse_packet_header()), and then a series of superframes. Superframe
1926 * boundaries may exceed packets, i.e. superframes can split data over
1927 * multiple (two) packets.
1929 * For more information about frames, see #synth_superframe().
1931 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1932 int *got_frame_ptr, AVPacket *avpkt)
1934 WMAVoiceContext *s = ctx->priv_data;
1935 GetBitContext *gb = &s->gb;
1938 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1939 * header at each ctx->block_align bytes. However, Libav's ASF demuxer
1940 * feeds us ASF packets, which may concatenate multiple "codec" packets
1941 * in a single "muxer" packet, so we artificially emulate that by
1942 * capping the packet size at ctx->block_align. */
1943 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1948 init_get_bits(&s->gb, avpkt->data, size << 3);
1950 /* size == ctx->block_align is used to indicate whether we are dealing with
1951 * a new packet or a packet of which we already read the packet header
1953 if (size == ctx->block_align) { // new packet header
1954 if ((res = parse_packet_header(s)) < 0)
1957 /* If the packet header specifies a s->spillover_nbits, then we want
1958 * to push out all data of the previous packet (+ spillover) before
1959 * continuing to parse new superframes in the current packet. */
1960 if (s->spillover_nbits > 0) {
1961 if (s->sframe_cache_size > 0) {
1962 int cnt = get_bits_count(gb);
1963 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1964 flush_put_bits(&s->pb);
1965 s->sframe_cache_size += s->spillover_nbits;
1966 if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
1968 cnt += s->spillover_nbits;
1969 s->skip_bits_next = cnt & 7;
1972 skip_bits_long (gb, s->spillover_nbits - cnt +
1973 get_bits_count(gb)); // resync
1975 skip_bits_long(gb, s->spillover_nbits); // resync
1977 } else if (s->skip_bits_next)
1978 skip_bits(gb, s->skip_bits_next);
1980 /* Try parsing superframes in current packet */
1981 s->sframe_cache_size = 0;
1982 s->skip_bits_next = 0;
1983 pos = get_bits_left(gb);
1984 if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
1986 } else if (*got_frame_ptr) {
1987 int cnt = get_bits_count(gb);
1988 s->skip_bits_next = cnt & 7;
1990 } else if ((s->sframe_cache_size = pos) > 0) {
1991 /* rewind bit reader to start of last (incomplete) superframe... */
1992 init_get_bits(gb, avpkt->data, size << 3);
1993 skip_bits_long(gb, (size << 3) - pos);
1994 assert(get_bits_left(gb) == pos);
1996 /* ...and cache it for spillover in next packet */
1997 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
1998 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
1999 // FIXME bad - just copy bytes as whole and add use the
2000 // skip_bits_next field
2006 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2008 WMAVoiceContext *s = ctx->priv_data;
2011 ff_rdft_end(&s->rdft);
2012 ff_rdft_end(&s->irdft);
2013 ff_dct_end(&s->dct);
2014 ff_dct_end(&s->dst);
2020 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2022 WMAVoiceContext *s = ctx->priv_data;
2025 s->postfilter_agc = 0;
2026 s->sframe_cache_size = 0;
2027 s->skip_bits_next = 0;
2028 for (n = 0; n < s->lsps; n++)
2029 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2030 memset(s->excitation_history, 0,
2031 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2032 memset(s->synth_history, 0,
2033 sizeof(*s->synth_history) * MAX_LSPS);
2034 memset(s->gain_pred_err, 0,
2035 sizeof(s->gain_pred_err));
2038 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2039 sizeof(*s->synth_filter_out_buf) * s->lsps);
2040 memset(s->dcf_mem, 0,
2041 sizeof(*s->dcf_mem) * 2);
2042 memset(s->zero_exc_pf, 0,
2043 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2044 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2048 AVCodec ff_wmavoice_decoder = {
2050 .type = AVMEDIA_TYPE_AUDIO,
2051 .id = AV_CODEC_ID_WMAVOICE,
2052 .priv_data_size = sizeof(WMAVoiceContext),
2053 .init = wmavoice_decode_init,
2054 .close = wmavoice_decode_end,
2055 .decode = wmavoice_decode_packet,
2056 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
2057 .flush = wmavoice_flush,
2058 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),