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/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.
140 GetBitContext gb; ///< packet bitreader. During decoder init,
141 ///< it contains the extradata from the
142 ///< demuxer. During decoding, it contains
144 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
146 int spillover_bitsize; ///< number of bits used to specify
147 ///< #spillover_nbits in the packet header
148 ///< = ceil(log2(ctx->block_align << 3))
149 int history_nsamples; ///< number of samples in history for signal
150 ///< prediction (through ACB)
152 /* postfilter specific values */
153 int do_apf; ///< whether to apply the averaged
154 ///< projection filter (APF)
155 int denoise_strength; ///< strength of denoising in Wiener filter
157 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
158 ///< Wiener filter coefficients (postfilter)
159 int dc_level; ///< Predicted amount of DC noise, based
160 ///< on which a DC removal filter is used
162 int lsps; ///< number of LSPs per frame [10 or 16]
163 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
164 int lsp_def_mode; ///< defines different sets of LSP defaults
166 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
167 ///< per-frame (independent coding)
168 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
169 ///< per superframe (residual coding)
171 int min_pitch_val; ///< base value for pitch parsing code
172 int max_pitch_val; ///< max value + 1 for pitch parsing
173 int pitch_nbits; ///< number of bits used to specify the
174 ///< pitch value in the frame header
175 int block_pitch_nbits; ///< number of bits used to specify the
176 ///< first block's pitch value
177 int block_pitch_range; ///< range of the block pitch
178 int block_delta_pitch_nbits; ///< number of bits used to specify the
179 ///< delta pitch between this and the last
180 ///< block's pitch value, used in all but
182 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
183 ///< from -this to +this-1)
184 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
190 * @name Packet values specified in the packet header or related to a packet.
192 * A packet is considered to be a single unit of data provided to this
193 * decoder by the demuxer.
196 int spillover_nbits; ///< number of bits of the previous packet's
197 ///< last superframe preceding this
198 ///< packet's first full superframe (useful
199 ///< for re-synchronization also)
200 int has_residual_lsps; ///< if set, superframes contain one set of
201 ///< LSPs that cover all frames, encoded as
202 ///< independent and residual LSPs; if not
203 ///< set, each frame contains its own, fully
204 ///< independent, LSPs
205 int skip_bits_next; ///< number of bits to skip at the next call
206 ///< to #wmavoice_decode_packet() (since
207 ///< they're part of the previous superframe)
209 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];
210 ///< cache for superframe data split over
211 ///< multiple packets
212 int sframe_cache_size; ///< set to >0 if we have data from an
213 ///< (incomplete) superframe from a previous
214 ///< packet that spilled over in the current
215 ///< packet; specifies the amount of bits in
217 PutBitContext pb; ///< bitstream writer for #sframe_cache
222 * @name Frame and superframe values
223 * Superframe and frame data - these can change from frame to frame,
224 * although some of them do in that case serve as a cache / history for
225 * the next frame or superframe.
228 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
230 int last_pitch_val; ///< pitch value of the previous frame
231 int last_acb_type; ///< frame type [0-2] of the previous frame
232 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
233 ///< << 16) / #MAX_FRAMESIZE
234 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
236 int aw_idx_is_ext; ///< whether the AW index was encoded in
237 ///< 8 bits (instead of 6)
238 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
239 ///< can apply the pulse, relative to the
240 ///< value in aw_first_pulse_off. The exact
241 ///< position of the first AW-pulse is within
242 ///< [pulse_off, pulse_off + this], and
243 ///< depends on bitstream values; [16 or 24]
244 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
245 ///< that this number can be negative (in
246 ///< which case it basically means "zero")
247 int aw_first_pulse_off[2]; ///< index of first sample to which to
248 ///< apply AW-pulses, or -0xff if unset
249 int aw_next_pulse_off_cache; ///< the position (relative to start of the
250 ///< second block) at which pulses should
251 ///< start to be positioned, serves as a
252 ///< cache for pitch-adaptive window pulses
255 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
256 ///< only used for comfort noise in #pRNG()
257 float gain_pred_err[6]; ///< cache for gain prediction
258 float excitation_history[MAX_SIGNAL_HISTORY];
259 ///< cache of the signal of previous
260 ///< superframes, used as a history for
261 ///< signal generation
262 float synth_history[MAX_LSPS]; ///< see #excitation_history
266 * @name Postfilter values
268 * Variables used for postfilter implementation, mostly history for
269 * smoothing and so on, and context variables for FFT/iFFT.
272 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
273 ///< postfilter (for denoise filter)
274 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
275 ///< transform, part of postfilter)
276 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
278 float postfilter_agc; ///< gain control memory, used in
279 ///< #adaptive_gain_control()
280 float dcf_mem[2]; ///< DC filter history
281 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
282 ///< zero filter output (i.e. excitation)
284 float denoise_filter_cache[MAX_FRAMESIZE];
285 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
286 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
287 ///< aligned buffer for LPC tilting
288 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
289 ///< aligned buffer for denoise coefficients
290 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
291 ///< aligned buffer for postfilter speech
299 * Set up the variable bit mode (VBM) tree from container extradata.
300 * @param gb bit I/O context.
301 * The bit context (s->gb) should be loaded with byte 23-46 of the
302 * container extradata (i.e. the ones containing the VBM tree).
303 * @param vbm_tree pointer to array to which the decoded VBM tree will be
305 * @return 0 on success, <0 on error.
307 static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
309 static const uint8_t bits[] = {
312 10, 10, 10, 12, 12, 12,
315 static const uint16_t codes[] = {
316 0x0000, 0x0001, 0x0002, // 00/01/10
317 0x000c, 0x000d, 0x000e, // 11+00/01/10
318 0x003c, 0x003d, 0x003e, // 1111+00/01/10
319 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
320 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
321 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
322 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
324 int cntr[8] = { 0 }, n, res;
326 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
327 for (n = 0; n < 17; n++) {
328 res = get_bits(gb, 3);
329 if (cntr[res] > 3) // should be >= 3 + (res == 7))
331 vbm_tree[res * 3 + cntr[res]++] = n;
333 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
334 bits, 1, 1, codes, 2, 2, 132);
339 * Set up decoder with parameters from demuxer (extradata etc.).
341 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
343 int n, flags, pitch_range, lsp16_flag;
344 WMAVoiceContext *s = ctx->priv_data;
348 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
349 * - byte 19-22: flags field (annoyingly in LE; see below for known
351 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
354 if (ctx->extradata_size != 46) {
355 av_log(ctx, AV_LOG_ERROR,
356 "Invalid extradata size %d (should be 46)\n",
357 ctx->extradata_size);
360 flags = AV_RL32(ctx->extradata + 18);
361 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
362 s->do_apf = flags & 0x1;
364 ff_rdft_init(&s->rdft, 7, DFT_R2C);
365 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
366 ff_dct_init(&s->dct, 6, DCT_I);
367 ff_dct_init(&s->dst, 6, DST_I);
369 ff_sine_window_init(s->cos, 256);
370 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
371 for (n = 0; n < 255; n++) {
372 s->sin[n] = -s->sin[510 - n];
373 s->cos[510 - n] = s->cos[n];
376 s->denoise_strength = (flags >> 2) & 0xF;
377 if (s->denoise_strength >= 12) {
378 av_log(ctx, AV_LOG_ERROR,
379 "Invalid denoise filter strength %d (max=11)\n",
380 s->denoise_strength);
383 s->denoise_tilt_corr = !!(flags & 0x40);
384 s->dc_level = (flags >> 7) & 0xF;
385 s->lsp_q_mode = !!(flags & 0x2000);
386 s->lsp_def_mode = !!(flags & 0x4000);
387 lsp16_flag = flags & 0x1000;
390 s->frame_lsp_bitsize = 34;
391 s->sframe_lsp_bitsize = 60;
394 s->frame_lsp_bitsize = 24;
395 s->sframe_lsp_bitsize = 48;
397 for (n = 0; n < s->lsps; n++)
398 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
400 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
401 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
402 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
406 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
407 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
408 pitch_range = s->max_pitch_val - s->min_pitch_val;
409 if (pitch_range <= 0) {
410 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
413 s->pitch_nbits = av_ceil_log2(pitch_range);
414 s->last_pitch_val = 40;
415 s->last_acb_type = ACB_TYPE_NONE;
416 s->history_nsamples = s->max_pitch_val + 8;
418 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
419 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
420 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
422 av_log(ctx, AV_LOG_ERROR,
423 "Unsupported samplerate %d (min=%d, max=%d)\n",
424 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
429 s->block_conv_table[0] = s->min_pitch_val;
430 s->block_conv_table[1] = (pitch_range * 25) >> 6;
431 s->block_conv_table[2] = (pitch_range * 44) >> 6;
432 s->block_conv_table[3] = s->max_pitch_val - 1;
433 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
434 if (s->block_delta_pitch_hrange <= 0) {
435 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
438 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
439 s->block_pitch_range = s->block_conv_table[2] +
440 s->block_conv_table[3] + 1 +
441 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
442 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
445 ctx->channel_layout = AV_CH_LAYOUT_MONO;
446 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
448 avcodec_get_frame_defaults(&s->frame);
449 ctx->coded_frame = &s->frame;
455 * @name Postfilter functions
456 * Postfilter functions (gain control, wiener denoise filter, DC filter,
457 * kalman smoothening, plus surrounding code to wrap it)
461 * Adaptive gain control (as used in postfilter).
463 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
464 * that the energy here is calculated using sum(abs(...)), whereas the
465 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
467 * @param out output buffer for filtered samples
468 * @param in input buffer containing the samples as they are after the
469 * postfilter steps so far
470 * @param speech_synth input buffer containing speech synth before postfilter
471 * @param size input buffer size
472 * @param alpha exponential filter factor
473 * @param gain_mem pointer to filter memory (single float)
475 static void adaptive_gain_control(float *out, const float *in,
476 const float *speech_synth,
477 int size, float alpha, float *gain_mem)
480 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
481 float mem = *gain_mem;
483 for (i = 0; i < size; i++) {
484 speech_energy += fabsf(speech_synth[i]);
485 postfilter_energy += fabsf(in[i]);
487 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
489 for (i = 0; i < size; i++) {
490 mem = alpha * mem + gain_scale_factor;
491 out[i] = in[i] * mem;
498 * Kalman smoothing function.
500 * This function looks back pitch +/- 3 samples back into history to find
501 * the best fitting curve (that one giving the optimal gain of the two
502 * signals, i.e. the highest dot product between the two), and then
503 * uses that signal history to smoothen the output of the speech synthesis
506 * @param s WMA Voice decoding context
507 * @param pitch pitch of the speech signal
508 * @param in input speech signal
509 * @param out output pointer for smoothened signal
510 * @param size input/output buffer size
512 * @returns -1 if no smoothening took place, e.g. because no optimal
513 * fit could be found, or 0 on success.
515 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
516 const float *in, float *out, int size)
519 float optimal_gain = 0, dot;
520 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
521 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
524 /* find best fitting point in history */
526 dot = ff_scalarproduct_float_c(in, ptr, size);
527 if (dot > optimal_gain) {
531 } while (--ptr >= end);
533 if (optimal_gain <= 0)
535 dot = ff_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
536 if (dot <= 0) // would be 1.0
539 if (optimal_gain <= dot) {
540 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
544 /* actual smoothing */
545 for (n = 0; n < size; n++)
546 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
552 * Get the tilt factor of a formant filter from its transfer function
553 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
554 * but somehow (??) it does a speech synthesis filter in the
555 * middle, which is missing here
557 * @param lpcs LPC coefficients
558 * @param n_lpcs Size of LPC buffer
559 * @returns the tilt factor
561 static float tilt_factor(const float *lpcs, int n_lpcs)
565 rh0 = 1.0 + ff_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
566 rh1 = lpcs[0] + ff_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
572 * Derive denoise filter coefficients (in real domain) from the LPCs.
574 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
575 int fcb_type, float *coeffs, int remainder)
577 float last_coeff, min = 15.0, max = -15.0;
578 float irange, angle_mul, gain_mul, range, sq;
581 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
582 s->rdft.rdft_calc(&s->rdft, lpcs);
583 #define log_range(var, assign) do { \
584 float tmp = log10f(assign); var = tmp; \
585 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
587 log_range(last_coeff, lpcs[1] * lpcs[1]);
588 for (n = 1; n < 64; n++)
589 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
590 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
591 log_range(lpcs[0], lpcs[0] * lpcs[0]);
594 lpcs[64] = last_coeff;
596 /* Now, use this spectrum to pick out these frequencies with higher
597 * (relative) power/energy (which we then take to be "not noise"),
598 * and set up a table (still in lpc[]) of (relative) gains per frequency.
599 * These frequencies will be maintained, while others ("noise") will be
600 * decreased in the filter output. */
601 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
602 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
604 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
605 for (n = 0; n <= 64; n++) {
608 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
609 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
610 lpcs[n] = angle_mul * pwr;
612 /* 70.57 =~ 1/log10(1.0331663) */
613 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
614 if (idx > 127) { // fallback if index falls outside table range
615 coeffs[n] = wmavoice_energy_table[127] *
616 powf(1.0331663, idx - 127);
618 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
621 /* calculate the Hilbert transform of the gains, which we do (since this
622 * is a sinus input) by doing a phase shift (in theory, H(sin())=cos()).
623 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
624 * "moment" of the LPCs in this filter. */
625 s->dct.dct_calc(&s->dct, lpcs);
626 s->dst.dct_calc(&s->dst, lpcs);
628 /* Split out the coefficient indexes into phase/magnitude pairs */
629 idx = 255 + av_clip(lpcs[64], -255, 255);
630 coeffs[0] = coeffs[0] * s->cos[idx];
631 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
632 last_coeff = coeffs[64] * s->cos[idx];
634 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
635 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
636 coeffs[n * 2] = coeffs[n] * s->cos[idx];
640 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
641 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
642 coeffs[n * 2] = coeffs[n] * s->cos[idx];
644 coeffs[1] = last_coeff;
646 /* move into real domain */
647 s->irdft.rdft_calc(&s->irdft, coeffs);
649 /* tilt correction and normalize scale */
650 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
651 if (s->denoise_tilt_corr) {
654 coeffs[remainder - 1] = 0;
655 ff_tilt_compensation(&tilt_mem,
656 -1.8 * tilt_factor(coeffs, remainder - 1),
659 sq = (1.0 / 64.0) * sqrtf(1 / ff_scalarproduct_float_c(coeffs, coeffs, remainder));
660 for (n = 0; n < remainder; n++)
665 * This function applies a Wiener filter on the (noisy) speech signal as
666 * a means to denoise it.
668 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
669 * - using this power spectrum, calculate (for each frequency) the Wiener
670 * filter gain, which depends on the frequency power and desired level
671 * of noise subtraction (when set too high, this leads to artifacts)
672 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
674 * - by doing a phase shift, calculate the Hilbert transform of this array
675 * of per-frequency filter-gains to get the filtering coefficients;
676 * - smoothen/normalize/de-tilt these filter coefficients as desired;
677 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
678 * to get the denoised speech signal;
679 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
680 * the frame boundary) are saved and applied to subsequent frames by an
681 * overlap-add method (otherwise you get clicking-artifacts).
683 * @param s WMA Voice decoding context
684 * @param fcb_type Frame (codebook) type
685 * @param synth_pf input: the noisy speech signal, output: denoised speech
686 * data; should be 16-byte aligned (for ASM purposes)
687 * @param size size of the speech data
688 * @param lpcs LPCs used to synthesize this frame's speech data
690 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
691 float *synth_pf, int size,
694 int remainder, lim, n;
696 if (fcb_type != FCB_TYPE_SILENCE) {
697 float *tilted_lpcs = s->tilted_lpcs_pf,
698 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
700 tilted_lpcs[0] = 1.0;
701 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
702 memset(&tilted_lpcs[s->lsps + 1], 0,
703 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
704 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
705 tilted_lpcs, s->lsps + 2);
707 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
708 * size is applied to the next frame. All input beyond this is zero,
709 * and thus all output beyond this will go towards zero, hence we can
710 * limit to min(size-1, 127-size) as a performance consideration. */
711 remainder = FFMIN(127 - size, size - 1);
712 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
714 /* apply coefficients (in frequency spectrum domain), i.e. complex
715 * number multiplication */
716 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
717 s->rdft.rdft_calc(&s->rdft, synth_pf);
718 s->rdft.rdft_calc(&s->rdft, coeffs);
719 synth_pf[0] *= coeffs[0];
720 synth_pf[1] *= coeffs[1];
721 for (n = 1; n < 64; n++) {
722 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
723 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
724 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
726 s->irdft.rdft_calc(&s->irdft, synth_pf);
729 /* merge filter output with the history of previous runs */
730 if (s->denoise_filter_cache_size) {
731 lim = FFMIN(s->denoise_filter_cache_size, size);
732 for (n = 0; n < lim; n++)
733 synth_pf[n] += s->denoise_filter_cache[n];
734 s->denoise_filter_cache_size -= lim;
735 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
736 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
739 /* move remainder of filter output into a cache for future runs */
740 if (fcb_type != FCB_TYPE_SILENCE) {
741 lim = FFMIN(remainder, s->denoise_filter_cache_size);
742 for (n = 0; n < lim; n++)
743 s->denoise_filter_cache[n] += synth_pf[size + n];
744 if (lim < remainder) {
745 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
746 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
747 s->denoise_filter_cache_size = remainder;
753 * Averaging projection filter, the postfilter used in WMAVoice.
755 * This uses the following steps:
756 * - A zero-synthesis filter (generate excitation from synth signal)
757 * - Kalman smoothing on excitation, based on pitch
758 * - Re-synthesized smoothened output
759 * - Iterative Wiener denoise filter
760 * - Adaptive gain filter
763 * @param s WMAVoice decoding context
764 * @param synth Speech synthesis output (before postfilter)
765 * @param samples Output buffer for filtered samples
766 * @param size Buffer size of synth & samples
767 * @param lpcs Generated LPCs used for speech synthesis
768 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
769 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
770 * @param pitch Pitch of the input signal
772 static void postfilter(WMAVoiceContext *s, const float *synth,
773 float *samples, int size,
774 const float *lpcs, float *zero_exc_pf,
775 int fcb_type, int pitch)
777 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
778 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
779 *synth_filter_in = zero_exc_pf;
781 assert(size <= MAX_FRAMESIZE / 2);
783 /* generate excitation from input signal */
784 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
786 if (fcb_type >= FCB_TYPE_AW_PULSES &&
787 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
788 synth_filter_in = synth_filter_in_buf;
790 /* re-synthesize speech after smoothening, and keep history */
791 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
792 synth_filter_in, size, s->lsps);
793 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
794 sizeof(synth_pf[0]) * s->lsps);
796 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
798 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
801 if (s->dc_level > 8) {
802 /* remove ultra-low frequency DC noise / highpass filter;
803 * coefficients are identical to those used in SIPR decoding,
804 * and very closely resemble those used in AMR-NB decoding. */
805 ff_acelp_apply_order_2_transfer_function(samples, samples,
806 (const float[2]) { -1.99997, 1.0 },
807 (const float[2]) { -1.9330735188, 0.93589198496 },
808 0.93980580475, s->dcf_mem, size);
817 * @param lsps output pointer to the array that will hold the LSPs
818 * @param num number of LSPs to be dequantized
819 * @param values quantized values, contains n_stages values
820 * @param sizes range (i.e. max value) of each quantized value
821 * @param n_stages number of dequantization runs
822 * @param table dequantization table to be used
823 * @param mul_q LSF multiplier
824 * @param base_q base (lowest) LSF values
826 static void dequant_lsps(double *lsps, int num,
827 const uint16_t *values,
828 const uint16_t *sizes,
829 int n_stages, const uint8_t *table,
831 const double *base_q)
835 memset(lsps, 0, num * sizeof(*lsps));
836 for (n = 0; n < n_stages; n++) {
837 const uint8_t *t_off = &table[values[n] * num];
838 double base = base_q[n], mul = mul_q[n];
840 for (m = 0; m < num; m++)
841 lsps[m] += base + mul * t_off[m];
843 table += sizes[n] * num;
848 * @name LSP dequantization routines
849 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
850 * @note we assume enough bits are available, caller should check.
851 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
852 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
856 * Parse 10 independently-coded LSPs.
858 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
860 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
861 static const double mul_lsf[4] = {
862 5.2187144800e-3, 1.4626986422e-3,
863 9.6179549166e-4, 1.1325736225e-3
865 static const double base_lsf[4] = {
866 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
867 M_PI * -3.3486e-2, M_PI * -5.7408e-2
871 v[0] = get_bits(gb, 8);
872 v[1] = get_bits(gb, 6);
873 v[2] = get_bits(gb, 5);
874 v[3] = get_bits(gb, 5);
876 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
881 * Parse 10 independently-coded LSPs, and then derive the tables to
882 * generate LSPs for the other frames from them (residual coding).
884 static void dequant_lsp10r(GetBitContext *gb,
885 double *i_lsps, const double *old,
886 double *a1, double *a2, int q_mode)
888 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
889 static const double mul_lsf[3] = {
890 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
892 static const double base_lsf[3] = {
893 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
895 const float (*ipol_tab)[2][10] = q_mode ?
896 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
897 uint16_t interpol, v[3];
900 dequant_lsp10i(gb, i_lsps);
902 interpol = get_bits(gb, 5);
903 v[0] = get_bits(gb, 7);
904 v[1] = get_bits(gb, 6);
905 v[2] = get_bits(gb, 6);
907 for (n = 0; n < 10; n++) {
908 double delta = old[n] - i_lsps[n];
909 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
910 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
913 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
918 * Parse 16 independently-coded LSPs.
920 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
922 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
923 static const double mul_lsf[5] = {
924 3.3439586280e-3, 6.9908173703e-4,
925 3.3216608306e-3, 1.0334960326e-3,
928 static const double base_lsf[5] = {
929 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
930 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
935 v[0] = get_bits(gb, 8);
936 v[1] = get_bits(gb, 6);
937 v[2] = get_bits(gb, 7);
938 v[3] = get_bits(gb, 6);
939 v[4] = get_bits(gb, 7);
941 dequant_lsps( lsps, 5, v, vec_sizes, 2,
942 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
943 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
944 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
945 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
946 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
950 * Parse 16 independently-coded LSPs, and then derive the tables to
951 * generate LSPs for the other frames from them (residual coding).
953 static void dequant_lsp16r(GetBitContext *gb,
954 double *i_lsps, const double *old,
955 double *a1, double *a2, int q_mode)
957 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
958 static const double mul_lsf[3] = {
959 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
961 static const double base_lsf[3] = {
962 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
964 const float (*ipol_tab)[2][16] = q_mode ?
965 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
966 uint16_t interpol, v[3];
969 dequant_lsp16i(gb, i_lsps);
971 interpol = get_bits(gb, 5);
972 v[0] = get_bits(gb, 7);
973 v[1] = get_bits(gb, 7);
974 v[2] = get_bits(gb, 7);
976 for (n = 0; n < 16; n++) {
977 double delta = old[n] - i_lsps[n];
978 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
979 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
982 dequant_lsps( a2, 10, v, vec_sizes, 1,
983 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
984 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
985 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
986 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
987 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
992 * @name Pitch-adaptive window coding functions
993 * The next few functions are for pitch-adaptive window coding.
997 * Parse the offset of the first pitch-adaptive window pulses, and
998 * the distribution of pulses between the two blocks in this frame.
999 * @param s WMA Voice decoding context private data
1000 * @param gb bit I/O context
1001 * @param pitch pitch for each block in this frame
1003 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1006 static const int16_t start_offset[94] = {
1007 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1008 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1009 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1010 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1011 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1012 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1013 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1014 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1018 /* position of pulse */
1019 s->aw_idx_is_ext = 0;
1020 if ((bits = get_bits(gb, 6)) >= 54) {
1021 s->aw_idx_is_ext = 1;
1022 bits += (bits - 54) * 3 + get_bits(gb, 2);
1025 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1026 * the distribution of the pulses in each block contained in this frame. */
1027 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1028 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1029 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1030 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1031 offset += s->aw_n_pulses[0] * pitch[0];
1032 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1033 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1035 /* if continuing from a position before the block, reset position to
1036 * start of block (when corrected for the range over which it can be
1037 * spread in aw_pulse_set1()). */
1038 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1039 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1040 s->aw_first_pulse_off[1] -= pitch[1];
1041 if (start_offset[bits] < 0)
1042 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1043 s->aw_first_pulse_off[0] -= pitch[0];
1048 * Apply second set of pitch-adaptive window pulses.
1049 * @param s WMA Voice decoding context private data
1050 * @param gb bit I/O context
1051 * @param block_idx block index in frame [0, 1]
1052 * @param fcb structure containing fixed codebook vector info
1054 static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1055 int block_idx, AMRFixed *fcb)
1057 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1058 uint16_t *use_mask = use_mask_mem + 2;
1059 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1060 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1061 * of idx are the position of the bit within a particular item in the
1062 * array (0 being the most significant bit, and 15 being the least
1063 * significant bit), and the remainder (>> 4) is the index in the
1064 * use_mask[]-array. This is faster and uses less memory than using a
1065 * 80-byte/80-int array. */
1066 int pulse_off = s->aw_first_pulse_off[block_idx],
1067 pulse_start, n, idx, range, aidx, start_off = 0;
1069 /* set offset of first pulse to within this block */
1070 if (s->aw_n_pulses[block_idx] > 0)
1071 while (pulse_off + s->aw_pulse_range < 1)
1072 pulse_off += fcb->pitch_lag;
1074 /* find range per pulse */
1075 if (s->aw_n_pulses[0] > 0) {
1076 if (block_idx == 0) {
1078 } else /* block_idx = 1 */ {
1080 if (s->aw_n_pulses[block_idx] > 0)
1081 pulse_off = s->aw_next_pulse_off_cache;
1085 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1087 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1088 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1089 * we exclude that range from being pulsed again in this function. */
1090 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1091 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1092 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1093 if (s->aw_n_pulses[block_idx] > 0)
1094 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1095 int excl_range = s->aw_pulse_range; // always 16 or 24
1096 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1097 int first_sh = 16 - (idx & 15);
1098 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1099 excl_range -= first_sh;
1100 if (excl_range >= 16) {
1101 *use_mask_ptr++ = 0;
1102 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1104 *use_mask_ptr &= 0xFFFF >> excl_range;
1107 /* find the 'aidx'th offset that is not excluded */
1108 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1109 for (n = 0; n <= aidx; pulse_start++) {
1110 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1111 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1112 if (use_mask[0]) idx = 0x0F;
1113 else if (use_mask[1]) idx = 0x1F;
1114 else if (use_mask[2]) idx = 0x2F;
1115 else if (use_mask[3]) idx = 0x3F;
1116 else if (use_mask[4]) idx = 0x4F;
1118 idx -= av_log2_16bit(use_mask[idx >> 4]);
1120 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1121 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1127 fcb->x[fcb->n] = start_off;
1128 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1131 /* set offset for next block, relative to start of that block */
1132 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1133 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1137 * Apply first set of pitch-adaptive window pulses.
1138 * @param s WMA Voice decoding context private data
1139 * @param gb bit I/O context
1140 * @param block_idx block index in frame [0, 1]
1141 * @param fcb storage location for fixed codebook pulse info
1143 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1144 int block_idx, AMRFixed *fcb)
1146 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1149 if (s->aw_n_pulses[block_idx] > 0) {
1150 int n, v_mask, i_mask, sh, n_pulses;
1152 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1157 } else { // 4 pulses, 1:sign + 2:index each
1164 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1165 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1166 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1167 s->aw_first_pulse_off[block_idx];
1168 while (fcb->x[fcb->n] < 0)
1169 fcb->x[fcb->n] += fcb->pitch_lag;
1170 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1174 int num2 = (val & 0x1FF) >> 1, delta, idx;
1176 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1177 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1178 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1179 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1180 v = (val & 0x200) ? -1.0 : 1.0;
1182 fcb->no_repeat_mask |= 3 << fcb->n;
1183 fcb->x[fcb->n] = idx - delta;
1185 fcb->x[fcb->n + 1] = idx;
1186 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1194 * Generate a random number from frame_cntr and block_idx, which will lief
1195 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1196 * table of size 1000 of which you want to read block_size entries).
1198 * @param frame_cntr current frame number
1199 * @param block_num current block index
1200 * @param block_size amount of entries we want to read from a table
1201 * that has 1000 entries
1202 * @return a (non-)random number in the [0, 1000 - block_size] range.
1204 static int pRNG(int frame_cntr, int block_num, int block_size)
1206 /* array to simplify the calculation of z:
1207 * y = (x % 9) * 5 + 6;
1208 * z = (49995 * x) / y;
1209 * Since y only has 9 values, we can remove the division by using a
1210 * LUT and using FASTDIV-style divisions. For each of the 9 values
1211 * of y, we can rewrite z as:
1212 * z = x * (49995 / y) + x * ((49995 % y) / y)
1213 * In this table, each col represents one possible value of y, the
1214 * first number is 49995 / y, and the second is the FASTDIV variant
1215 * of 49995 % y / y. */
1216 static const unsigned int div_tbl[9][2] = {
1217 { 8332, 3 * 715827883U }, // y = 6
1218 { 4545, 0 * 390451573U }, // y = 11
1219 { 3124, 11 * 268435456U }, // y = 16
1220 { 2380, 15 * 204522253U }, // y = 21
1221 { 1922, 23 * 165191050U }, // y = 26
1222 { 1612, 23 * 138547333U }, // y = 31
1223 { 1388, 27 * 119304648U }, // y = 36
1224 { 1219, 16 * 104755300U }, // y = 41
1225 { 1086, 39 * 93368855U } // y = 46
1227 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1228 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1229 // so this is effectively a modulo (%)
1230 y = x - 9 * MULH(477218589, x); // x % 9
1231 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1232 // z = x * 49995 / (y * 5 + 6)
1233 return z % (1000 - block_size);
1237 * Parse hardcoded signal for a single block.
1238 * @note see #synth_block().
1240 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1241 int block_idx, int size,
1242 const struct frame_type_desc *frame_desc,
1248 assert(size <= MAX_FRAMESIZE);
1250 /* Set the offset from which we start reading wmavoice_std_codebook */
1251 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1252 r_idx = pRNG(s->frame_cntr, block_idx, size);
1253 gain = s->silence_gain;
1254 } else /* FCB_TYPE_HARDCODED */ {
1255 r_idx = get_bits(gb, 8);
1256 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1259 /* Clear gain prediction parameters */
1260 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1262 /* Apply gain to hardcoded codebook and use that as excitation signal */
1263 for (n = 0; n < size; n++)
1264 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1268 * Parse FCB/ACB signal for a single block.
1269 * @note see #synth_block().
1271 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1272 int block_idx, int size,
1273 int block_pitch_sh2,
1274 const struct frame_type_desc *frame_desc,
1277 static const float gain_coeff[6] = {
1278 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1280 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1281 int n, idx, gain_weight;
1284 assert(size <= MAX_FRAMESIZE / 2);
1285 memset(pulses, 0, sizeof(*pulses) * size);
1287 fcb.pitch_lag = block_pitch_sh2 >> 2;
1288 fcb.pitch_fac = 1.0;
1289 fcb.no_repeat_mask = 0;
1292 /* For the other frame types, this is where we apply the innovation
1293 * (fixed) codebook pulses of the speech signal. */
1294 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1295 aw_pulse_set1(s, gb, block_idx, &fcb);
1296 aw_pulse_set2(s, gb, block_idx, &fcb);
1297 } else /* FCB_TYPE_EXC_PULSES */ {
1298 int offset_nbits = 5 - frame_desc->log_n_blocks;
1300 fcb.no_repeat_mask = -1;
1301 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1302 * (instead of double) for a subset of pulses */
1303 for (n = 0; n < 5; n++) {
1307 sign = get_bits1(gb) ? 1.0 : -1.0;
1308 pos1 = get_bits(gb, offset_nbits);
1309 fcb.x[fcb.n] = n + 5 * pos1;
1310 fcb.y[fcb.n++] = sign;
1311 if (n < frame_desc->dbl_pulses) {
1312 pos2 = get_bits(gb, offset_nbits);
1313 fcb.x[fcb.n] = n + 5 * pos2;
1314 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1318 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1320 /* Calculate gain for adaptive & fixed codebook signal.
1321 * see ff_amr_set_fixed_gain(). */
1322 idx = get_bits(gb, 7);
1323 fcb_gain = expf(ff_scalarproduct_float_c(s->gain_pred_err, gain_coeff, 6) -
1324 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1325 acb_gain = wmavoice_gain_codebook_acb[idx];
1326 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1327 -2.9957322736 /* log(0.05) */,
1328 1.6094379124 /* log(5.0) */);
1330 gain_weight = 8 >> frame_desc->log_n_blocks;
1331 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1332 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1333 for (n = 0; n < gain_weight; n++)
1334 s->gain_pred_err[n] = pred_err;
1336 /* Calculation of adaptive codebook */
1337 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1339 for (n = 0; n < size; n += len) {
1341 int abs_idx = block_idx * size + n;
1342 int pitch_sh16 = (s->last_pitch_val << 16) +
1343 s->pitch_diff_sh16 * abs_idx;
1344 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1345 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1346 idx = idx_sh16 >> 16;
1347 if (s->pitch_diff_sh16) {
1348 if (s->pitch_diff_sh16 > 0) {
1349 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1351 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1352 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1357 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1358 wmavoice_ipol1_coeffs, 17,
1361 } else /* ACB_TYPE_HAMMING */ {
1362 int block_pitch = block_pitch_sh2 >> 2;
1363 idx = block_pitch_sh2 & 3;
1365 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1366 wmavoice_ipol2_coeffs, 4,
1369 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1370 sizeof(float) * size);
1373 /* Interpolate ACB/FCB and use as excitation signal */
1374 ff_weighted_vector_sumf(excitation, excitation, pulses,
1375 acb_gain, fcb_gain, size);
1379 * Parse data in a single block.
1380 * @note we assume enough bits are available, caller should check.
1382 * @param s WMA Voice decoding context private data
1383 * @param gb bit I/O context
1384 * @param block_idx index of the to-be-read block
1385 * @param size amount of samples to be read in this block
1386 * @param block_pitch_sh2 pitch for this block << 2
1387 * @param lsps LSPs for (the end of) this frame
1388 * @param prev_lsps LSPs for the last frame
1389 * @param frame_desc frame type descriptor
1390 * @param excitation target memory for the ACB+FCB interpolated signal
1391 * @param synth target memory for the speech synthesis filter output
1392 * @return 0 on success, <0 on error.
1394 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1395 int block_idx, int size,
1396 int block_pitch_sh2,
1397 const double *lsps, const double *prev_lsps,
1398 const struct frame_type_desc *frame_desc,
1399 float *excitation, float *synth)
1401 double i_lsps[MAX_LSPS];
1402 float lpcs[MAX_LSPS];
1406 if (frame_desc->acb_type == ACB_TYPE_NONE)
1407 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1409 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1410 frame_desc, excitation);
1412 /* convert interpolated LSPs to LPCs */
1413 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1414 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1415 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1416 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1418 /* Speech synthesis */
1419 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1423 * Synthesize output samples for a single frame.
1424 * @note we assume enough bits are available, caller should check.
1426 * @param ctx WMA Voice decoder context
1427 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1428 * @param frame_idx Frame number within superframe [0-2]
1429 * @param samples pointer to output sample buffer, has space for at least 160
1431 * @param lsps LSP array
1432 * @param prev_lsps array of previous frame's LSPs
1433 * @param excitation target buffer for excitation signal
1434 * @param synth target buffer for synthesized speech data
1435 * @return 0 on success, <0 on error.
1437 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1439 const double *lsps, const double *prev_lsps,
1440 float *excitation, float *synth)
1442 WMAVoiceContext *s = ctx->priv_data;
1443 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1444 int pitch[MAX_BLOCKS], last_block_pitch;
1446 /* Parse frame type ("frame header"), see frame_descs */
1447 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1450 av_log(ctx, AV_LOG_ERROR,
1451 "Invalid frame type VLC code, skipping\n");
1455 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1457 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1458 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1459 /* Pitch is provided per frame, which is interpreted as the pitch of
1460 * the last sample of the last block of this frame. We can interpolate
1461 * the pitch of other blocks (and even pitch-per-sample) by gradually
1462 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1463 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1464 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1465 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1466 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1467 if (s->last_acb_type == ACB_TYPE_NONE ||
1468 20 * abs(cur_pitch_val - s->last_pitch_val) >
1469 (cur_pitch_val + s->last_pitch_val))
1470 s->last_pitch_val = cur_pitch_val;
1472 /* pitch per block */
1473 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1474 int fac = n * 2 + 1;
1476 pitch[n] = (MUL16(fac, cur_pitch_val) +
1477 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1478 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1481 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1482 s->pitch_diff_sh16 =
1483 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1486 /* Global gain (if silence) and pitch-adaptive window coordinates */
1487 switch (frame_descs[bd_idx].fcb_type) {
1488 case FCB_TYPE_SILENCE:
1489 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1491 case FCB_TYPE_AW_PULSES:
1492 aw_parse_coords(s, gb, pitch);
1496 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1499 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1500 switch (frame_descs[bd_idx].acb_type) {
1501 case ACB_TYPE_HAMMING: {
1502 /* Pitch is given per block. Per-block pitches are encoded as an
1503 * absolute value for the first block, and then delta values
1504 * relative to this value) for all subsequent blocks. The scale of
1505 * this pitch value is semi-logaritmic compared to its use in the
1506 * decoder, so we convert it to normal scale also. */
1508 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1509 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1510 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1513 block_pitch = get_bits(gb, s->block_pitch_nbits);
1515 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1516 get_bits(gb, s->block_delta_pitch_nbits);
1517 /* Convert last_ so that any next delta is within _range */
1518 last_block_pitch = av_clip(block_pitch,
1519 s->block_delta_pitch_hrange,
1520 s->block_pitch_range -
1521 s->block_delta_pitch_hrange);
1523 /* Convert semi-log-style scale back to normal scale */
1524 if (block_pitch < t1) {
1525 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1528 if (block_pitch < t2) {
1530 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1533 if (block_pitch < t3) {
1535 (s->block_conv_table[2] + block_pitch) << 2;
1537 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1540 pitch[n] = bl_pitch_sh2 >> 2;
1544 case ACB_TYPE_ASYMMETRIC: {
1545 bl_pitch_sh2 = pitch[n] << 2;
1549 default: // ACB_TYPE_NONE has no pitch
1554 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1555 lsps, prev_lsps, &frame_descs[bd_idx],
1556 &excitation[n * block_nsamples],
1557 &synth[n * block_nsamples]);
1560 /* Averaging projection filter, if applicable. Else, just copy samples
1561 * from synthesis buffer */
1563 double i_lsps[MAX_LSPS];
1564 float lpcs[MAX_LSPS];
1566 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1567 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1568 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1569 postfilter(s, synth, samples, 80, lpcs,
1570 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1571 frame_descs[bd_idx].fcb_type, pitch[0]);
1573 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1574 i_lsps[n] = cos(lsps[n]);
1575 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1576 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1577 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1578 frame_descs[bd_idx].fcb_type, pitch[0]);
1580 memcpy(samples, synth, 160 * sizeof(synth[0]));
1582 /* Cache values for next frame */
1584 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1585 s->last_acb_type = frame_descs[bd_idx].acb_type;
1586 switch (frame_descs[bd_idx].acb_type) {
1588 s->last_pitch_val = 0;
1590 case ACB_TYPE_ASYMMETRIC:
1591 s->last_pitch_val = cur_pitch_val;
1593 case ACB_TYPE_HAMMING:
1594 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1602 * Ensure minimum value for first item, maximum value for last value,
1603 * proper spacing between each value and proper ordering.
1605 * @param lsps array of LSPs
1606 * @param num size of LSP array
1608 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1609 * useful to put in a generic location later on. Parts are also
1610 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1611 * which is in float.
1613 static void stabilize_lsps(double *lsps, int num)
1617 /* set minimum value for first, maximum value for last and minimum
1618 * spacing between LSF values.
1619 * Very similar to ff_set_min_dist_lsf(), but in double. */
1620 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1621 for (n = 1; n < num; n++)
1622 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1623 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1625 /* reorder (looks like one-time / non-recursed bubblesort).
1626 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1627 for (n = 1; n < num; n++) {
1628 if (lsps[n] < lsps[n - 1]) {
1629 for (m = 1; m < num; m++) {
1630 double tmp = lsps[m];
1631 for (l = m - 1; l >= 0; l--) {
1632 if (lsps[l] <= tmp) break;
1633 lsps[l + 1] = lsps[l];
1643 * Test if there's enough bits to read 1 superframe.
1645 * @param orig_gb bit I/O context used for reading. This function
1646 * does not modify the state of the bitreader; it
1647 * only uses it to copy the current stream position
1648 * @param s WMA Voice decoding context private data
1649 * @return -1 if unsupported, 1 on not enough bits or 0 if OK.
1651 static int check_bits_for_superframe(GetBitContext *orig_gb,
1654 GetBitContext s_gb, *gb = &s_gb;
1655 int n, need_bits, bd_idx;
1656 const struct frame_type_desc *frame_desc;
1658 /* initialize a copy */
1659 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1660 skip_bits_long(gb, get_bits_count(orig_gb));
1661 assert(get_bits_left(gb) == get_bits_left(orig_gb));
1663 /* superframe header */
1664 if (get_bits_left(gb) < 14)
1667 return -1; // WMAPro-in-WMAVoice superframe
1668 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1669 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1670 if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1672 skip_bits_long(gb, s->sframe_lsp_bitsize);
1676 for (n = 0; n < MAX_FRAMES; n++) {
1677 int aw_idx_is_ext = 0;
1679 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1680 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1681 skip_bits_long(gb, s->frame_lsp_bitsize);
1683 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1685 return -1; // invalid frame type VLC code
1686 frame_desc = &frame_descs[bd_idx];
1687 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1688 if (get_bits_left(gb) < s->pitch_nbits)
1690 skip_bits_long(gb, s->pitch_nbits);
1692 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1694 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1695 int tmp = get_bits(gb, 6);
1703 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1704 need_bits = s->block_pitch_nbits +
1705 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1706 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1707 need_bits = 2 * !aw_idx_is_ext;
1710 need_bits += frame_desc->frame_size;
1711 if (get_bits_left(gb) < need_bits)
1713 skip_bits_long(gb, need_bits);
1720 * Synthesize output samples for a single superframe. If we have any data
1721 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1724 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1725 * to give a total of 480 samples per frame. See #synth_frame() for frame
1726 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1727 * (if these are globally specified for all frames (residually); they can
1728 * also be specified individually per-frame. See the s->has_residual_lsps
1729 * option), and can specify the number of samples encoded in this superframe
1730 * (if less than 480), usually used to prevent blanks at track boundaries.
1732 * @param ctx WMA Voice decoder context
1733 * @return 0 on success, <0 on error or 1 if there was not enough data to
1734 * fully parse the superframe
1736 static int synth_superframe(AVCodecContext *ctx, int *got_frame_ptr)
1738 WMAVoiceContext *s = ctx->priv_data;
1739 GetBitContext *gb = &s->gb, s_gb;
1740 int n, res, n_samples = 480;
1741 double lsps[MAX_FRAMES][MAX_LSPS];
1742 const double *mean_lsf = s->lsps == 16 ?
1743 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1744 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1745 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1748 memcpy(synth, s->synth_history,
1749 s->lsps * sizeof(*synth));
1750 memcpy(excitation, s->excitation_history,
1751 s->history_nsamples * sizeof(*excitation));
1753 if (s->sframe_cache_size > 0) {
1755 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1756 s->sframe_cache_size = 0;
1759 if ((res = check_bits_for_superframe(gb, s)) == 1) {
1764 /* First bit is speech/music bit, it differentiates between WMAVoice
1765 * speech samples (the actual codec) and WMAVoice music samples, which
1766 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1768 if (!get_bits1(gb)) {
1769 av_log_missing_feature(ctx, "WMAPro-in-WMAVoice", 1);
1770 return AVERROR_PATCHWELCOME;
1773 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1774 if (get_bits1(gb)) {
1775 if ((n_samples = get_bits(gb, 12)) > 480) {
1776 av_log(ctx, AV_LOG_ERROR,
1777 "Superframe encodes >480 samples (%d), not allowed\n",
1782 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1783 if (s->has_residual_lsps) {
1784 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1786 for (n = 0; n < s->lsps; n++)
1787 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1789 if (s->lsps == 10) {
1790 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1791 } else /* s->lsps == 16 */
1792 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1794 for (n = 0; n < s->lsps; n++) {
1795 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1796 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1797 lsps[2][n] += mean_lsf[n];
1799 for (n = 0; n < 3; n++)
1800 stabilize_lsps(lsps[n], s->lsps);
1803 /* get output buffer */
1804 s->frame.nb_samples = 480;
1805 if ((res = ff_get_buffer(ctx, &s->frame)) < 0) {
1806 av_log(ctx, AV_LOG_ERROR, "get_buffer() failed\n");
1809 s->frame.nb_samples = n_samples;
1810 samples = (float *)s->frame.data[0];
1812 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1813 for (n = 0; n < 3; n++) {
1814 if (!s->has_residual_lsps) {
1817 if (s->lsps == 10) {
1818 dequant_lsp10i(gb, lsps[n]);
1819 } else /* s->lsps == 16 */
1820 dequant_lsp16i(gb, lsps[n]);
1822 for (m = 0; m < s->lsps; m++)
1823 lsps[n][m] += mean_lsf[m];
1824 stabilize_lsps(lsps[n], s->lsps);
1827 if ((res = synth_frame(ctx, gb, n,
1828 &samples[n * MAX_FRAMESIZE],
1829 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1830 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1831 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1837 /* Statistics? FIXME - we don't check for length, a slight overrun
1838 * will be caught by internal buffer padding, and anything else
1839 * will be skipped, not read. */
1840 if (get_bits1(gb)) {
1841 res = get_bits(gb, 4);
1842 skip_bits(gb, 10 * (res + 1));
1847 /* Update history */
1848 memcpy(s->prev_lsps, lsps[2],
1849 s->lsps * sizeof(*s->prev_lsps));
1850 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1851 s->lsps * sizeof(*synth));
1852 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1853 s->history_nsamples * sizeof(*excitation));
1855 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1856 s->history_nsamples * sizeof(*s->zero_exc_pf));
1862 * Parse the packet header at the start of each packet (input data to this
1865 * @param s WMA Voice decoding context private data
1866 * @return 1 if not enough bits were available, or 0 on success.
1868 static int parse_packet_header(WMAVoiceContext *s)
1870 GetBitContext *gb = &s->gb;
1873 if (get_bits_left(gb) < 11)
1875 skip_bits(gb, 4); // packet sequence number
1876 s->has_residual_lsps = get_bits1(gb);
1878 res = get_bits(gb, 6); // number of superframes per packet
1879 // (minus first one if there is spillover)
1880 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1882 } while (res == 0x3F);
1883 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1889 * Copy (unaligned) bits from gb/data/size to pb.
1891 * @param pb target buffer to copy bits into
1892 * @param data source buffer to copy bits from
1893 * @param size size of the source data, in bytes
1894 * @param gb bit I/O context specifying the current position in the source.
1895 * data. This function might use this to align the bit position to
1896 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1898 * @param nbits the amount of bits to copy from source to target
1900 * @note after calling this function, the current position in the input bit
1901 * I/O context is undefined.
1903 static void copy_bits(PutBitContext *pb,
1904 const uint8_t *data, int size,
1905 GetBitContext *gb, int nbits)
1907 int rmn_bytes, rmn_bits;
1909 rmn_bits = rmn_bytes = get_bits_left(gb);
1910 if (rmn_bits < nbits)
1912 if (nbits > pb->size_in_bits - put_bits_count(pb))
1914 rmn_bits &= 7; rmn_bytes >>= 3;
1915 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1916 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1917 avpriv_copy_bits(pb, data + size - rmn_bytes,
1918 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1922 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1923 * and we expect that the demuxer / application provides it to us as such
1924 * (else you'll probably get garbage as output). Every packet has a size of
1925 * ctx->block_align bytes, starts with a packet header (see
1926 * #parse_packet_header()), and then a series of superframes. Superframe
1927 * boundaries may exceed packets, i.e. superframes can split data over
1928 * multiple (two) packets.
1930 * For more information about frames, see #synth_superframe().
1932 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1933 int *got_frame_ptr, AVPacket *avpkt)
1935 WMAVoiceContext *s = ctx->priv_data;
1936 GetBitContext *gb = &s->gb;
1939 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1940 * header at each ctx->block_align bytes. However, Libav's ASF demuxer
1941 * feeds us ASF packets, which may concatenate multiple "codec" packets
1942 * in a single "muxer" packet, so we artificially emulate that by
1943 * capping the packet size at ctx->block_align. */
1944 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1949 init_get_bits(&s->gb, avpkt->data, size << 3);
1951 /* size == ctx->block_align is used to indicate whether we are dealing with
1952 * a new packet or a packet of which we already read the packet header
1954 if (size == ctx->block_align) { // new packet header
1955 if ((res = parse_packet_header(s)) < 0)
1958 /* If the packet header specifies a s->spillover_nbits, then we want
1959 * to push out all data of the previous packet (+ spillover) before
1960 * continuing to parse new superframes in the current packet. */
1961 if (s->spillover_nbits > 0) {
1962 if (s->sframe_cache_size > 0) {
1963 int cnt = get_bits_count(gb);
1964 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1965 flush_put_bits(&s->pb);
1966 s->sframe_cache_size += s->spillover_nbits;
1967 if ((res = synth_superframe(ctx, got_frame_ptr)) == 0 &&
1969 cnt += s->spillover_nbits;
1970 s->skip_bits_next = cnt & 7;
1971 *(AVFrame *)data = s->frame;
1974 skip_bits_long (gb, s->spillover_nbits - cnt +
1975 get_bits_count(gb)); // resync
1977 skip_bits_long(gb, s->spillover_nbits); // resync
1979 } else if (s->skip_bits_next)
1980 skip_bits(gb, s->skip_bits_next);
1982 /* Try parsing superframes in current packet */
1983 s->sframe_cache_size = 0;
1984 s->skip_bits_next = 0;
1985 pos = get_bits_left(gb);
1986 if ((res = synth_superframe(ctx, got_frame_ptr)) < 0) {
1988 } else if (*got_frame_ptr) {
1989 int cnt = get_bits_count(gb);
1990 s->skip_bits_next = cnt & 7;
1991 *(AVFrame *)data = s->frame;
1993 } else if ((s->sframe_cache_size = pos) > 0) {
1994 /* rewind bit reader to start of last (incomplete) superframe... */
1995 init_get_bits(gb, avpkt->data, size << 3);
1996 skip_bits_long(gb, (size << 3) - pos);
1997 assert(get_bits_left(gb) == pos);
1999 /* ...and cache it for spillover in next packet */
2000 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
2001 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
2002 // FIXME bad - just copy bytes as whole and add use the
2003 // skip_bits_next field
2009 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2011 WMAVoiceContext *s = ctx->priv_data;
2014 ff_rdft_end(&s->rdft);
2015 ff_rdft_end(&s->irdft);
2016 ff_dct_end(&s->dct);
2017 ff_dct_end(&s->dst);
2023 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2025 WMAVoiceContext *s = ctx->priv_data;
2028 s->postfilter_agc = 0;
2029 s->sframe_cache_size = 0;
2030 s->skip_bits_next = 0;
2031 for (n = 0; n < s->lsps; n++)
2032 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2033 memset(s->excitation_history, 0,
2034 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2035 memset(s->synth_history, 0,
2036 sizeof(*s->synth_history) * MAX_LSPS);
2037 memset(s->gain_pred_err, 0,
2038 sizeof(s->gain_pred_err));
2041 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2042 sizeof(*s->synth_filter_out_buf) * s->lsps);
2043 memset(s->dcf_mem, 0,
2044 sizeof(*s->dcf_mem) * 2);
2045 memset(s->zero_exc_pf, 0,
2046 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2047 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2051 AVCodec ff_wmavoice_decoder = {
2053 .type = AVMEDIA_TYPE_AUDIO,
2054 .id = AV_CODEC_ID_WMAVOICE,
2055 .priv_data_size = sizeof(WMAVoiceContext),
2056 .init = wmavoice_decode_init,
2057 .close = wmavoice_decode_end,
2058 .decode = wmavoice_decode_packet,
2059 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
2060 .flush = wmavoice_flush,
2061 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),