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 int cntr[8] = { 0 }, n, res;
310 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
311 for (n = 0; n < 17; n++) {
312 res = get_bits(gb, 3);
313 if (cntr[res] > 3) // should be >= 3 + (res == 7))
315 vbm_tree[res * 3 + cntr[res]++] = n;
320 static av_cold void wmavoice_init_static_data(AVCodec *codec)
322 static const uint8_t bits[] = {
325 10, 10, 10, 12, 12, 12,
328 static const uint16_t codes[] = {
329 0x0000, 0x0001, 0x0002, // 00/01/10
330 0x000c, 0x000d, 0x000e, // 11+00/01/10
331 0x003c, 0x003d, 0x003e, // 1111+00/01/10
332 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
333 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
334 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
335 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
338 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
339 bits, 1, 1, codes, 2, 2, 132);
343 * Set up decoder with parameters from demuxer (extradata etc.).
345 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
347 int n, flags, pitch_range, lsp16_flag;
348 WMAVoiceContext *s = ctx->priv_data;
352 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
353 * - byte 19-22: flags field (annoyingly in LE; see below for known
355 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
358 if (ctx->extradata_size != 46) {
359 av_log(ctx, AV_LOG_ERROR,
360 "Invalid extradata size %d (should be 46)\n",
361 ctx->extradata_size);
362 return AVERROR_INVALIDDATA;
364 flags = AV_RL32(ctx->extradata + 18);
365 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
366 s->do_apf = flags & 0x1;
368 ff_rdft_init(&s->rdft, 7, DFT_R2C);
369 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
370 ff_dct_init(&s->dct, 6, DCT_I);
371 ff_dct_init(&s->dst, 6, DST_I);
373 ff_sine_window_init(s->cos, 256);
374 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
375 for (n = 0; n < 255; n++) {
376 s->sin[n] = -s->sin[510 - n];
377 s->cos[510 - n] = s->cos[n];
380 s->denoise_strength = (flags >> 2) & 0xF;
381 if (s->denoise_strength >= 12) {
382 av_log(ctx, AV_LOG_ERROR,
383 "Invalid denoise filter strength %d (max=11)\n",
384 s->denoise_strength);
385 return AVERROR_INVALIDDATA;
387 s->denoise_tilt_corr = !!(flags & 0x40);
388 s->dc_level = (flags >> 7) & 0xF;
389 s->lsp_q_mode = !!(flags & 0x2000);
390 s->lsp_def_mode = !!(flags & 0x4000);
391 lsp16_flag = flags & 0x1000;
394 s->frame_lsp_bitsize = 34;
395 s->sframe_lsp_bitsize = 60;
398 s->frame_lsp_bitsize = 24;
399 s->sframe_lsp_bitsize = 48;
401 for (n = 0; n < s->lsps; n++)
402 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
404 init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
405 if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
406 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
407 return AVERROR_INVALIDDATA;
410 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
411 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
412 pitch_range = s->max_pitch_val - s->min_pitch_val;
413 if (pitch_range <= 0) {
414 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
415 return AVERROR_INVALIDDATA;
417 s->pitch_nbits = av_ceil_log2(pitch_range);
418 s->last_pitch_val = 40;
419 s->last_acb_type = ACB_TYPE_NONE;
420 s->history_nsamples = s->max_pitch_val + 8;
422 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
423 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
424 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
426 av_log(ctx, AV_LOG_ERROR,
427 "Unsupported samplerate %d (min=%d, max=%d)\n",
428 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
430 return AVERROR(ENOSYS);
433 s->block_conv_table[0] = s->min_pitch_val;
434 s->block_conv_table[1] = (pitch_range * 25) >> 6;
435 s->block_conv_table[2] = (pitch_range * 44) >> 6;
436 s->block_conv_table[3] = s->max_pitch_val - 1;
437 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
438 if (s->block_delta_pitch_hrange <= 0) {
439 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
440 return AVERROR_INVALIDDATA;
442 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
443 s->block_pitch_range = s->block_conv_table[2] +
444 s->block_conv_table[3] + 1 +
445 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
446 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
449 ctx->channel_layout = AV_CH_LAYOUT_MONO;
450 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
456 * @name Postfilter functions
457 * Postfilter functions (gain control, wiener denoise filter, DC filter,
458 * kalman smoothening, plus surrounding code to wrap it)
462 * Adaptive gain control (as used in postfilter).
464 * Identical to #ff_adaptive_gain_control() in acelp_vectors.c, except
465 * that the energy here is calculated using sum(abs(...)), whereas the
466 * other codecs (e.g. AMR-NB, SIPRO) use sqrt(dotproduct(...)).
468 * @param out output buffer for filtered samples
469 * @param in input buffer containing the samples as they are after the
470 * postfilter steps so far
471 * @param speech_synth input buffer containing speech synth before postfilter
472 * @param size input buffer size
473 * @param alpha exponential filter factor
474 * @param gain_mem pointer to filter memory (single float)
476 static void adaptive_gain_control(float *out, const float *in,
477 const float *speech_synth,
478 int size, float alpha, float *gain_mem)
481 float speech_energy = 0.0, postfilter_energy = 0.0, gain_scale_factor;
482 float mem = *gain_mem;
484 for (i = 0; i < size; i++) {
485 speech_energy += fabsf(speech_synth[i]);
486 postfilter_energy += fabsf(in[i]);
488 gain_scale_factor = (1.0 - alpha) * speech_energy / postfilter_energy;
490 for (i = 0; i < size; i++) {
491 mem = alpha * mem + gain_scale_factor;
492 out[i] = in[i] * mem;
499 * Kalman smoothing function.
501 * This function looks back pitch +/- 3 samples back into history to find
502 * the best fitting curve (that one giving the optimal gain of the two
503 * signals, i.e. the highest dot product between the two), and then
504 * uses that signal history to smoothen the output of the speech synthesis
507 * @param s WMA Voice decoding context
508 * @param pitch pitch of the speech signal
509 * @param in input speech signal
510 * @param out output pointer for smoothened signal
511 * @param size input/output buffer size
513 * @returns -1 if no smoothening took place, e.g. because no optimal
514 * fit could be found, or 0 on success.
516 static int kalman_smoothen(WMAVoiceContext *s, int pitch,
517 const float *in, float *out, int size)
520 float optimal_gain = 0, dot;
521 const float *ptr = &in[-FFMAX(s->min_pitch_val, pitch - 3)],
522 *end = &in[-FFMIN(s->max_pitch_val, pitch + 3)],
525 /* find best fitting point in history */
527 dot = avpriv_scalarproduct_float_c(in, ptr, size);
528 if (dot > optimal_gain) {
532 } while (--ptr >= end);
534 if (optimal_gain <= 0)
536 dot = avpriv_scalarproduct_float_c(best_hist_ptr, best_hist_ptr, size);
537 if (dot <= 0) // would be 1.0
540 if (optimal_gain <= dot) {
541 dot = dot / (dot + 0.6 * optimal_gain); // 0.625-1.000
545 /* actual smoothing */
546 for (n = 0; n < size; n++)
547 out[n] = best_hist_ptr[n] + dot * (in[n] - best_hist_ptr[n]);
553 * Get the tilt factor of a formant filter from its transfer function
554 * @see #tilt_factor() in amrnbdec.c, which does essentially the same,
555 * but somehow (??) it does a speech synthesis filter in the
556 * middle, which is missing here
558 * @param lpcs LPC coefficients
559 * @param n_lpcs Size of LPC buffer
560 * @returns the tilt factor
562 static float tilt_factor(const float *lpcs, int n_lpcs)
566 rh0 = 1.0 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
567 rh1 = lpcs[0] + avpriv_scalarproduct_float_c(lpcs, &lpcs[1], n_lpcs - 1);
573 * Derive denoise filter coefficients (in real domain) from the LPCs.
575 static void calc_input_response(WMAVoiceContext *s, float *lpcs,
576 int fcb_type, float *coeffs, int remainder)
578 float last_coeff, min = 15.0, max = -15.0;
579 float irange, angle_mul, gain_mul, range, sq;
582 /* Create frequency power spectrum of speech input (i.e. RDFT of LPCs) */
583 s->rdft.rdft_calc(&s->rdft, lpcs);
584 #define log_range(var, assign) do { \
585 float tmp = log10f(assign); var = tmp; \
586 max = FFMAX(max, tmp); min = FFMIN(min, tmp); \
588 log_range(last_coeff, lpcs[1] * lpcs[1]);
589 for (n = 1; n < 64; n++)
590 log_range(lpcs[n], lpcs[n * 2] * lpcs[n * 2] +
591 lpcs[n * 2 + 1] * lpcs[n * 2 + 1]);
592 log_range(lpcs[0], lpcs[0] * lpcs[0]);
595 lpcs[64] = last_coeff;
597 /* Now, use this spectrum to pick out these frequencies with higher
598 * (relative) power/energy (which we then take to be "not noise"),
599 * and set up a table (still in lpc[]) of (relative) gains per frequency.
600 * These frequencies will be maintained, while others ("noise") will be
601 * decreased in the filter output. */
602 irange = 64.0 / range; // so irange*(max-value) is in the range [0, 63]
603 gain_mul = range * (fcb_type == FCB_TYPE_HARDCODED ? (5.0 / 13.0) :
605 angle_mul = gain_mul * (8.0 * M_LN10 / M_PI);
606 for (n = 0; n <= 64; n++) {
609 idx = FFMAX(0, lrint((max - lpcs[n]) * irange) - 1);
610 pwr = wmavoice_denoise_power_table[s->denoise_strength][idx];
611 lpcs[n] = angle_mul * pwr;
613 /* 70.57 =~ 1/log10(1.0331663) */
614 idx = (pwr * gain_mul - 0.0295) * 70.570526123;
615 if (idx > 127) { // fall back if index falls outside table range
616 coeffs[n] = wmavoice_energy_table[127] *
617 powf(1.0331663, idx - 127);
619 coeffs[n] = wmavoice_energy_table[FFMAX(0, idx)];
622 /* calculate the Hilbert transform of the gains, which we do (since this
623 * is a sine input) by doing a phase shift (in theory, H(sin())=cos()).
624 * Hilbert_Transform(RDFT(x)) = Laplace_Transform(x), which calculates the
625 * "moment" of the LPCs in this filter. */
626 s->dct.dct_calc(&s->dct, lpcs);
627 s->dst.dct_calc(&s->dst, lpcs);
629 /* Split out the coefficient indexes into phase/magnitude pairs */
630 idx = 255 + av_clip(lpcs[64], -255, 255);
631 coeffs[0] = coeffs[0] * s->cos[idx];
632 idx = 255 + av_clip(lpcs[64] - 2 * lpcs[63], -255, 255);
633 last_coeff = coeffs[64] * s->cos[idx];
635 idx = 255 + av_clip(-lpcs[64] - 2 * lpcs[n - 1], -255, 255);
636 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
637 coeffs[n * 2] = coeffs[n] * s->cos[idx];
641 idx = 255 + av_clip( lpcs[64] - 2 * lpcs[n - 1], -255, 255);
642 coeffs[n * 2 + 1] = coeffs[n] * s->sin[idx];
643 coeffs[n * 2] = coeffs[n] * s->cos[idx];
645 coeffs[1] = last_coeff;
647 /* move into real domain */
648 s->irdft.rdft_calc(&s->irdft, coeffs);
650 /* tilt correction and normalize scale */
651 memset(&coeffs[remainder], 0, sizeof(coeffs[0]) * (128 - remainder));
652 if (s->denoise_tilt_corr) {
655 coeffs[remainder - 1] = 0;
656 ff_tilt_compensation(&tilt_mem,
657 -1.8 * tilt_factor(coeffs, remainder - 1),
660 sq = (1.0 / 64.0) * sqrtf(1 / avpriv_scalarproduct_float_c(coeffs, coeffs,
662 for (n = 0; n < remainder; n++)
667 * This function applies a Wiener filter on the (noisy) speech signal as
668 * a means to denoise it.
670 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
671 * - using this power spectrum, calculate (for each frequency) the Wiener
672 * filter gain, which depends on the frequency power and desired level
673 * of noise subtraction (when set too high, this leads to artifacts)
674 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
676 * - by doing a phase shift, calculate the Hilbert transform of this array
677 * of per-frequency filter-gains to get the filtering coefficients;
678 * - smoothen/normalize/de-tilt these filter coefficients as desired;
679 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
680 * to get the denoised speech signal;
681 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
682 * the frame boundary) are saved and applied to subsequent frames by an
683 * overlap-add method (otherwise you get clicking-artifacts).
685 * @param s WMA Voice decoding context
686 * @param fcb_type Frame (codebook) type
687 * @param synth_pf input: the noisy speech signal, output: denoised speech
688 * data; should be 16-byte aligned (for ASM purposes)
689 * @param size size of the speech data
690 * @param lpcs LPCs used to synthesize this frame's speech data
692 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
693 float *synth_pf, int size,
696 int remainder, lim, n;
698 if (fcb_type != FCB_TYPE_SILENCE) {
699 float *tilted_lpcs = s->tilted_lpcs_pf,
700 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
702 tilted_lpcs[0] = 1.0;
703 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
704 memset(&tilted_lpcs[s->lsps + 1], 0,
705 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
706 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
707 tilted_lpcs, s->lsps + 2);
709 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
710 * size is applied to the next frame. All input beyond this is zero,
711 * and thus all output beyond this will go towards zero, hence we can
712 * limit to min(size-1, 127-size) as a performance consideration. */
713 remainder = FFMIN(127 - size, size - 1);
714 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
716 /* apply coefficients (in frequency spectrum domain), i.e. complex
717 * number multiplication */
718 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
719 s->rdft.rdft_calc(&s->rdft, synth_pf);
720 s->rdft.rdft_calc(&s->rdft, coeffs);
721 synth_pf[0] *= coeffs[0];
722 synth_pf[1] *= coeffs[1];
723 for (n = 1; n < 64; n++) {
724 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
725 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
726 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
728 s->irdft.rdft_calc(&s->irdft, synth_pf);
731 /* merge filter output with the history of previous runs */
732 if (s->denoise_filter_cache_size) {
733 lim = FFMIN(s->denoise_filter_cache_size, size);
734 for (n = 0; n < lim; n++)
735 synth_pf[n] += s->denoise_filter_cache[n];
736 s->denoise_filter_cache_size -= lim;
737 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
738 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
741 /* move remainder of filter output into a cache for future runs */
742 if (fcb_type != FCB_TYPE_SILENCE) {
743 lim = FFMIN(remainder, s->denoise_filter_cache_size);
744 for (n = 0; n < lim; n++)
745 s->denoise_filter_cache[n] += synth_pf[size + n];
746 if (lim < remainder) {
747 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
748 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
749 s->denoise_filter_cache_size = remainder;
755 * Averaging projection filter, the postfilter used in WMAVoice.
757 * This uses the following steps:
758 * - A zero-synthesis filter (generate excitation from synth signal)
759 * - Kalman smoothing on excitation, based on pitch
760 * - Re-synthesized smoothened output
761 * - Iterative Wiener denoise filter
762 * - Adaptive gain filter
765 * @param s WMAVoice decoding context
766 * @param synth Speech synthesis output (before postfilter)
767 * @param samples Output buffer for filtered samples
768 * @param size Buffer size of synth & samples
769 * @param lpcs Generated LPCs used for speech synthesis
770 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
771 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
772 * @param pitch Pitch of the input signal
774 static void postfilter(WMAVoiceContext *s, const float *synth,
775 float *samples, int size,
776 const float *lpcs, float *zero_exc_pf,
777 int fcb_type, int pitch)
779 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
780 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
781 *synth_filter_in = zero_exc_pf;
783 assert(size <= MAX_FRAMESIZE / 2);
785 /* generate excitation from input signal */
786 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
788 if (fcb_type >= FCB_TYPE_AW_PULSES &&
789 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
790 synth_filter_in = synth_filter_in_buf;
792 /* re-synthesize speech after smoothening, and keep history */
793 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
794 synth_filter_in, size, s->lsps);
795 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
796 sizeof(synth_pf[0]) * s->lsps);
798 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
800 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
803 if (s->dc_level > 8) {
804 /* remove ultra-low frequency DC noise / highpass filter;
805 * coefficients are identical to those used in SIPR decoding,
806 * and very closely resemble those used in AMR-NB decoding. */
807 ff_acelp_apply_order_2_transfer_function(samples, samples,
808 (const float[2]) { -1.99997, 1.0 },
809 (const float[2]) { -1.9330735188, 0.93589198496 },
810 0.93980580475, s->dcf_mem, size);
819 * @param lsps output pointer to the array that will hold the LSPs
820 * @param num number of LSPs to be dequantized
821 * @param values quantized values, contains n_stages values
822 * @param sizes range (i.e. max value) of each quantized value
823 * @param n_stages number of dequantization runs
824 * @param table dequantization table to be used
825 * @param mul_q LSF multiplier
826 * @param base_q base (lowest) LSF values
828 static void dequant_lsps(double *lsps, int num,
829 const uint16_t *values,
830 const uint16_t *sizes,
831 int n_stages, const uint8_t *table,
833 const double *base_q)
837 memset(lsps, 0, num * sizeof(*lsps));
838 for (n = 0; n < n_stages; n++) {
839 const uint8_t *t_off = &table[values[n] * num];
840 double base = base_q[n], mul = mul_q[n];
842 for (m = 0; m < num; m++)
843 lsps[m] += base + mul * t_off[m];
845 table += sizes[n] * num;
850 * @name LSP dequantization routines
851 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
852 * @note we assume enough bits are available, caller should check.
853 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
854 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
858 * Parse 10 independently-coded LSPs.
860 static void dequant_lsp10i(GetBitContext *gb, double *lsps)
862 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
863 static const double mul_lsf[4] = {
864 5.2187144800e-3, 1.4626986422e-3,
865 9.6179549166e-4, 1.1325736225e-3
867 static const double base_lsf[4] = {
868 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
869 M_PI * -3.3486e-2, M_PI * -5.7408e-2
873 v[0] = get_bits(gb, 8);
874 v[1] = get_bits(gb, 6);
875 v[2] = get_bits(gb, 5);
876 v[3] = get_bits(gb, 5);
878 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
883 * Parse 10 independently-coded LSPs, and then derive the tables to
884 * generate LSPs for the other frames from them (residual coding).
886 static void dequant_lsp10r(GetBitContext *gb,
887 double *i_lsps, const double *old,
888 double *a1, double *a2, int q_mode)
890 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
891 static const double mul_lsf[3] = {
892 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
894 static const double base_lsf[3] = {
895 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
897 const float (*ipol_tab)[2][10] = q_mode ?
898 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
899 uint16_t interpol, v[3];
902 dequant_lsp10i(gb, i_lsps);
904 interpol = get_bits(gb, 5);
905 v[0] = get_bits(gb, 7);
906 v[1] = get_bits(gb, 6);
907 v[2] = get_bits(gb, 6);
909 for (n = 0; n < 10; n++) {
910 double delta = old[n] - i_lsps[n];
911 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
912 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
915 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
920 * Parse 16 independently-coded LSPs.
922 static void dequant_lsp16i(GetBitContext *gb, double *lsps)
924 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
925 static const double mul_lsf[5] = {
926 3.3439586280e-3, 6.9908173703e-4,
927 3.3216608306e-3, 1.0334960326e-3,
930 static const double base_lsf[5] = {
931 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
932 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
937 v[0] = get_bits(gb, 8);
938 v[1] = get_bits(gb, 6);
939 v[2] = get_bits(gb, 7);
940 v[3] = get_bits(gb, 6);
941 v[4] = get_bits(gb, 7);
943 dequant_lsps( lsps, 5, v, vec_sizes, 2,
944 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
945 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
946 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
947 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
948 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
952 * Parse 16 independently-coded LSPs, and then derive the tables to
953 * generate LSPs for the other frames from them (residual coding).
955 static void dequant_lsp16r(GetBitContext *gb,
956 double *i_lsps, const double *old,
957 double *a1, double *a2, int q_mode)
959 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
960 static const double mul_lsf[3] = {
961 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
963 static const double base_lsf[3] = {
964 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
966 const float (*ipol_tab)[2][16] = q_mode ?
967 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
968 uint16_t interpol, v[3];
971 dequant_lsp16i(gb, i_lsps);
973 interpol = get_bits(gb, 5);
974 v[0] = get_bits(gb, 7);
975 v[1] = get_bits(gb, 7);
976 v[2] = get_bits(gb, 7);
978 for (n = 0; n < 16; n++) {
979 double delta = old[n] - i_lsps[n];
980 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
981 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
984 dequant_lsps( a2, 10, v, vec_sizes, 1,
985 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
986 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
987 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
988 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
989 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
994 * @name Pitch-adaptive window coding functions
995 * The next few functions are for pitch-adaptive window coding.
999 * Parse the offset of the first pitch-adaptive window pulses, and
1000 * the distribution of pulses between the two blocks in this frame.
1001 * @param s WMA Voice decoding context private data
1002 * @param gb bit I/O context
1003 * @param pitch pitch for each block in this frame
1005 static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
1008 static const int16_t start_offset[94] = {
1009 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1010 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1011 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1012 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1013 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1014 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1015 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1016 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1020 /* position of pulse */
1021 s->aw_idx_is_ext = 0;
1022 if ((bits = get_bits(gb, 6)) >= 54) {
1023 s->aw_idx_is_ext = 1;
1024 bits += (bits - 54) * 3 + get_bits(gb, 2);
1027 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1028 * the distribution of the pulses in each block contained in this frame. */
1029 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1030 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1031 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1032 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1033 offset += s->aw_n_pulses[0] * pitch[0];
1034 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1035 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1037 /* if continuing from a position before the block, reset position to
1038 * start of block (when corrected for the range over which it can be
1039 * spread in aw_pulse_set1()). */
1040 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1041 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1042 s->aw_first_pulse_off[1] -= pitch[1];
1043 if (start_offset[bits] < 0)
1044 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1045 s->aw_first_pulse_off[0] -= pitch[0];
1050 * Apply second set of pitch-adaptive window pulses.
1051 * @param s WMA Voice decoding context private data
1052 * @param gb bit I/O context
1053 * @param block_idx block index in frame [0, 1]
1054 * @param fcb structure containing fixed codebook vector info
1055 * @return -1 on error, 0 otherwise
1057 static int aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
1058 int block_idx, AMRFixed *fcb)
1060 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1061 uint16_t *use_mask = use_mask_mem + 2;
1062 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1063 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1064 * of idx are the position of the bit within a particular item in the
1065 * array (0 being the most significant bit, and 15 being the least
1066 * significant bit), and the remainder (>> 4) is the index in the
1067 * use_mask[]-array. This is faster and uses less memory than using a
1068 * 80-byte/80-int array. */
1069 int pulse_off = s->aw_first_pulse_off[block_idx],
1070 pulse_start, n, idx, range, aidx, start_off = 0;
1072 /* set offset of first pulse to within this block */
1073 if (s->aw_n_pulses[block_idx] > 0)
1074 while (pulse_off + s->aw_pulse_range < 1)
1075 pulse_off += fcb->pitch_lag;
1077 /* find range per pulse */
1078 if (s->aw_n_pulses[0] > 0) {
1079 if (block_idx == 0) {
1081 } else /* block_idx = 1 */ {
1083 if (s->aw_n_pulses[block_idx] > 0)
1084 pulse_off = s->aw_next_pulse_off_cache;
1088 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1090 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1091 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1092 * we exclude that range from being pulsed again in this function. */
1093 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1094 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1095 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1096 if (s->aw_n_pulses[block_idx] > 0)
1097 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1098 int excl_range = s->aw_pulse_range; // always 16 or 24
1099 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1100 int first_sh = 16 - (idx & 15);
1101 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1102 excl_range -= first_sh;
1103 if (excl_range >= 16) {
1104 *use_mask_ptr++ = 0;
1105 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1107 *use_mask_ptr &= 0xFFFF >> excl_range;
1110 /* find the 'aidx'th offset that is not excluded */
1111 aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1112 for (n = 0; n <= aidx; pulse_start++) {
1113 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1114 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1115 if (use_mask[0]) idx = 0x0F;
1116 else if (use_mask[1]) idx = 0x1F;
1117 else if (use_mask[2]) idx = 0x2F;
1118 else if (use_mask[3]) idx = 0x3F;
1119 else if (use_mask[4]) idx = 0x4F;
1121 idx -= av_log2_16bit(use_mask[idx >> 4]);
1123 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1124 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1130 fcb->x[fcb->n] = start_off;
1131 fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
1134 /* set offset for next block, relative to start of that block */
1135 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1136 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1141 * Apply first set of pitch-adaptive window pulses.
1142 * @param s WMA Voice decoding context private data
1143 * @param gb bit I/O context
1144 * @param block_idx block index in frame [0, 1]
1145 * @param fcb storage location for fixed codebook pulse info
1147 static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
1148 int block_idx, AMRFixed *fcb)
1150 int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1153 if (s->aw_n_pulses[block_idx] > 0) {
1154 int n, v_mask, i_mask, sh, n_pulses;
1156 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1161 } else { // 4 pulses, 1:sign + 2:index each
1168 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1169 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1170 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1171 s->aw_first_pulse_off[block_idx];
1172 while (fcb->x[fcb->n] < 0)
1173 fcb->x[fcb->n] += fcb->pitch_lag;
1174 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1178 int num2 = (val & 0x1FF) >> 1, delta, idx;
1180 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1181 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1182 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1183 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1184 v = (val & 0x200) ? -1.0 : 1.0;
1186 fcb->no_repeat_mask |= 3 << fcb->n;
1187 fcb->x[fcb->n] = idx - delta;
1189 fcb->x[fcb->n + 1] = idx;
1190 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1198 * Generate a random number from frame_cntr and block_idx, which will lief
1199 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1200 * table of size 1000 of which you want to read block_size entries).
1202 * @param frame_cntr current frame number
1203 * @param block_num current block index
1204 * @param block_size amount of entries we want to read from a table
1205 * that has 1000 entries
1206 * @return a (non-)random number in the [0, 1000 - block_size] range.
1208 static int pRNG(int frame_cntr, int block_num, int block_size)
1210 /* array to simplify the calculation of z:
1211 * y = (x % 9) * 5 + 6;
1212 * z = (49995 * x) / y;
1213 * Since y only has 9 values, we can remove the division by using a
1214 * LUT and using FASTDIV-style divisions. For each of the 9 values
1215 * of y, we can rewrite z as:
1216 * z = x * (49995 / y) + x * ((49995 % y) / y)
1217 * In this table, each col represents one possible value of y, the
1218 * first number is 49995 / y, and the second is the FASTDIV variant
1219 * of 49995 % y / y. */
1220 static const unsigned int div_tbl[9][2] = {
1221 { 8332, 3 * 715827883U }, // y = 6
1222 { 4545, 0 * 390451573U }, // y = 11
1223 { 3124, 11 * 268435456U }, // y = 16
1224 { 2380, 15 * 204522253U }, // y = 21
1225 { 1922, 23 * 165191050U }, // y = 26
1226 { 1612, 23 * 138547333U }, // y = 31
1227 { 1388, 27 * 119304648U }, // y = 36
1228 { 1219, 16 * 104755300U }, // y = 41
1229 { 1086, 39 * 93368855U } // y = 46
1231 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1232 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1233 // so this is effectively a modulo (%)
1234 y = x - 9 * MULH(477218589, x); // x % 9
1235 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1236 // z = x * 49995 / (y * 5 + 6)
1237 return z % (1000 - block_size);
1241 * Parse hardcoded signal for a single block.
1242 * @note see #synth_block().
1244 static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
1245 int block_idx, int size,
1246 const struct frame_type_desc *frame_desc,
1252 assert(size <= MAX_FRAMESIZE);
1254 /* Set the offset from which we start reading wmavoice_std_codebook */
1255 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1256 r_idx = pRNG(s->frame_cntr, block_idx, size);
1257 gain = s->silence_gain;
1258 } else /* FCB_TYPE_HARDCODED */ {
1259 r_idx = get_bits(gb, 8);
1260 gain = wmavoice_gain_universal[get_bits(gb, 6)];
1263 /* Clear gain prediction parameters */
1264 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1266 /* Apply gain to hardcoded codebook and use that as excitation signal */
1267 for (n = 0; n < size; n++)
1268 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1272 * Parse FCB/ACB signal for a single block.
1273 * @note see #synth_block().
1275 static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
1276 int block_idx, int size,
1277 int block_pitch_sh2,
1278 const struct frame_type_desc *frame_desc,
1281 static const float gain_coeff[6] = {
1282 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1284 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1285 int n, idx, gain_weight;
1288 assert(size <= MAX_FRAMESIZE / 2);
1289 memset(pulses, 0, sizeof(*pulses) * size);
1291 fcb.pitch_lag = block_pitch_sh2 >> 2;
1292 fcb.pitch_fac = 1.0;
1293 fcb.no_repeat_mask = 0;
1296 /* For the other frame types, this is where we apply the innovation
1297 * (fixed) codebook pulses of the speech signal. */
1298 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1299 aw_pulse_set1(s, gb, block_idx, &fcb);
1300 if (aw_pulse_set2(s, gb, block_idx, &fcb)) {
1301 /* Conceal the block with silence and return.
1302 * Skip the correct amount of bits to read the next
1303 * block from the correct offset. */
1304 int r_idx = pRNG(s->frame_cntr, block_idx, size);
1306 for (n = 0; n < size; n++)
1308 wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1309 skip_bits(gb, 7 + 1);
1312 } else /* FCB_TYPE_EXC_PULSES */ {
1313 int offset_nbits = 5 - frame_desc->log_n_blocks;
1315 fcb.no_repeat_mask = -1;
1316 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1317 * (instead of double) for a subset of pulses */
1318 for (n = 0; n < 5; n++) {
1322 sign = get_bits1(gb) ? 1.0 : -1.0;
1323 pos1 = get_bits(gb, offset_nbits);
1324 fcb.x[fcb.n] = n + 5 * pos1;
1325 fcb.y[fcb.n++] = sign;
1326 if (n < frame_desc->dbl_pulses) {
1327 pos2 = get_bits(gb, offset_nbits);
1328 fcb.x[fcb.n] = n + 5 * pos2;
1329 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1333 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1335 /* Calculate gain for adaptive & fixed codebook signal.
1336 * see ff_amr_set_fixed_gain(). */
1337 idx = get_bits(gb, 7);
1338 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1340 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1341 acb_gain = wmavoice_gain_codebook_acb[idx];
1342 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1343 -2.9957322736 /* log(0.05) */,
1344 1.6094379124 /* log(5.0) */);
1346 gain_weight = 8 >> frame_desc->log_n_blocks;
1347 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1348 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1349 for (n = 0; n < gain_weight; n++)
1350 s->gain_pred_err[n] = pred_err;
1352 /* Calculation of adaptive codebook */
1353 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1355 for (n = 0; n < size; n += len) {
1357 int abs_idx = block_idx * size + n;
1358 int pitch_sh16 = (s->last_pitch_val << 16) +
1359 s->pitch_diff_sh16 * abs_idx;
1360 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1361 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1362 idx = idx_sh16 >> 16;
1363 if (s->pitch_diff_sh16) {
1364 if (s->pitch_diff_sh16 > 0) {
1365 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1367 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1368 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1373 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1374 wmavoice_ipol1_coeffs, 17,
1377 } else /* ACB_TYPE_HAMMING */ {
1378 int block_pitch = block_pitch_sh2 >> 2;
1379 idx = block_pitch_sh2 & 3;
1381 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1382 wmavoice_ipol2_coeffs, 4,
1385 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1386 sizeof(float) * size);
1389 /* Interpolate ACB/FCB and use as excitation signal */
1390 ff_weighted_vector_sumf(excitation, excitation, pulses,
1391 acb_gain, fcb_gain, size);
1395 * Parse data in a single block.
1396 * @note we assume enough bits are available, caller should check.
1398 * @param s WMA Voice decoding context private data
1399 * @param gb bit I/O context
1400 * @param block_idx index of the to-be-read block
1401 * @param size amount of samples to be read in this block
1402 * @param block_pitch_sh2 pitch for this block << 2
1403 * @param lsps LSPs for (the end of) this frame
1404 * @param prev_lsps LSPs for the last frame
1405 * @param frame_desc frame type descriptor
1406 * @param excitation target memory for the ACB+FCB interpolated signal
1407 * @param synth target memory for the speech synthesis filter output
1408 * @return 0 on success, <0 on error.
1410 static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
1411 int block_idx, int size,
1412 int block_pitch_sh2,
1413 const double *lsps, const double *prev_lsps,
1414 const struct frame_type_desc *frame_desc,
1415 float *excitation, float *synth)
1417 double i_lsps[MAX_LSPS];
1418 float lpcs[MAX_LSPS];
1422 if (frame_desc->acb_type == ACB_TYPE_NONE)
1423 synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
1425 synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
1426 frame_desc, excitation);
1428 /* convert interpolated LSPs to LPCs */
1429 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1430 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1431 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1432 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1434 /* Speech synthesis */
1435 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1439 * Synthesize output samples for a single frame.
1440 * @note we assume enough bits are available, caller should check.
1442 * @param ctx WMA Voice decoder context
1443 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
1444 * @param frame_idx Frame number within superframe [0-2]
1445 * @param samples pointer to output sample buffer, has space for at least 160
1447 * @param lsps LSP array
1448 * @param prev_lsps array of previous frame's LSPs
1449 * @param excitation target buffer for excitation signal
1450 * @param synth target buffer for synthesized speech data
1451 * @return 0 on success, <0 on error.
1453 static int synth_frame(AVCodecContext *ctx, GetBitContext *gb, int frame_idx,
1455 const double *lsps, const double *prev_lsps,
1456 float *excitation, float *synth)
1458 WMAVoiceContext *s = ctx->priv_data;
1459 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1460 int pitch[MAX_BLOCKS], last_block_pitch;
1462 /* Parse frame type ("frame header"), see frame_descs */
1463 int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)], block_nsamples;
1466 av_log(ctx, AV_LOG_ERROR,
1467 "Invalid frame type VLC code, skipping\n");
1468 return AVERROR_INVALIDDATA;
1471 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1473 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1474 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1475 /* Pitch is provided per frame, which is interpreted as the pitch of
1476 * the last sample of the last block of this frame. We can interpolate
1477 * the pitch of other blocks (and even pitch-per-sample) by gradually
1478 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1479 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1480 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1481 cur_pitch_val = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
1482 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1483 if (s->last_acb_type == ACB_TYPE_NONE ||
1484 20 * abs(cur_pitch_val - s->last_pitch_val) >
1485 (cur_pitch_val + s->last_pitch_val))
1486 s->last_pitch_val = cur_pitch_val;
1488 /* pitch per block */
1489 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1490 int fac = n * 2 + 1;
1492 pitch[n] = (MUL16(fac, cur_pitch_val) +
1493 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1494 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1497 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1498 s->pitch_diff_sh16 =
1499 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1502 /* Global gain (if silence) and pitch-adaptive window coordinates */
1503 switch (frame_descs[bd_idx].fcb_type) {
1504 case FCB_TYPE_SILENCE:
1505 s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
1507 case FCB_TYPE_AW_PULSES:
1508 aw_parse_coords(s, gb, pitch);
1512 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1515 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1516 switch (frame_descs[bd_idx].acb_type) {
1517 case ACB_TYPE_HAMMING: {
1518 /* Pitch is given per block. Per-block pitches are encoded as an
1519 * absolute value for the first block, and then delta values
1520 * relative to this value) for all subsequent blocks. The scale of
1521 * this pitch value is semi-logaritmic compared to its use in the
1522 * decoder, so we convert it to normal scale also. */
1524 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1525 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1526 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1529 block_pitch = get_bits(gb, s->block_pitch_nbits);
1531 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1532 get_bits(gb, s->block_delta_pitch_nbits);
1533 /* Convert last_ so that any next delta is within _range */
1534 last_block_pitch = av_clip(block_pitch,
1535 s->block_delta_pitch_hrange,
1536 s->block_pitch_range -
1537 s->block_delta_pitch_hrange);
1539 /* Convert semi-log-style scale back to normal scale */
1540 if (block_pitch < t1) {
1541 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1544 if (block_pitch < t2) {
1546 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1549 if (block_pitch < t3) {
1551 (s->block_conv_table[2] + block_pitch) << 2;
1553 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1556 pitch[n] = bl_pitch_sh2 >> 2;
1560 case ACB_TYPE_ASYMMETRIC: {
1561 bl_pitch_sh2 = pitch[n] << 2;
1565 default: // ACB_TYPE_NONE has no pitch
1570 synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
1571 lsps, prev_lsps, &frame_descs[bd_idx],
1572 &excitation[n * block_nsamples],
1573 &synth[n * block_nsamples]);
1576 /* Averaging projection filter, if applicable. Else, just copy samples
1577 * from synthesis buffer */
1579 double i_lsps[MAX_LSPS];
1580 float lpcs[MAX_LSPS];
1582 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1583 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1584 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1585 postfilter(s, synth, samples, 80, lpcs,
1586 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1587 frame_descs[bd_idx].fcb_type, pitch[0]);
1589 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1590 i_lsps[n] = cos(lsps[n]);
1591 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1592 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1593 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1594 frame_descs[bd_idx].fcb_type, pitch[0]);
1596 memcpy(samples, synth, 160 * sizeof(synth[0]));
1598 /* Cache values for next frame */
1600 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1601 s->last_acb_type = frame_descs[bd_idx].acb_type;
1602 switch (frame_descs[bd_idx].acb_type) {
1604 s->last_pitch_val = 0;
1606 case ACB_TYPE_ASYMMETRIC:
1607 s->last_pitch_val = cur_pitch_val;
1609 case ACB_TYPE_HAMMING:
1610 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1618 * Ensure minimum value for first item, maximum value for last value,
1619 * proper spacing between each value and proper ordering.
1621 * @param lsps array of LSPs
1622 * @param num size of LSP array
1624 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1625 * useful to put in a generic location later on. Parts are also
1626 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1627 * which is in float.
1629 static void stabilize_lsps(double *lsps, int num)
1633 /* set minimum value for first, maximum value for last and minimum
1634 * spacing between LSF values.
1635 * Very similar to ff_set_min_dist_lsf(), but in double. */
1636 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1637 for (n = 1; n < num; n++)
1638 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1639 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1641 /* reorder (looks like one-time / non-recursed bubblesort).
1642 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1643 for (n = 1; n < num; n++) {
1644 if (lsps[n] < lsps[n - 1]) {
1645 for (m = 1; m < num; m++) {
1646 double tmp = lsps[m];
1647 for (l = m - 1; l >= 0; l--) {
1648 if (lsps[l] <= tmp) break;
1649 lsps[l + 1] = lsps[l];
1659 * Test if there's enough bits to read 1 superframe.
1661 * @param orig_gb bit I/O context used for reading. This function
1662 * does not modify the state of the bitreader; it
1663 * only uses it to copy the current stream position
1664 * @param s WMA Voice decoding context private data
1665 * @return < 0 on error, 1 on not enough bits or 0 if OK.
1667 static int check_bits_for_superframe(GetBitContext *orig_gb,
1670 GetBitContext s_gb, *gb = &s_gb;
1671 int n, need_bits, bd_idx;
1672 const struct frame_type_desc *frame_desc;
1674 /* initialize a copy */
1675 init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
1676 skip_bits_long(gb, get_bits_count(orig_gb));
1677 assert(get_bits_left(gb) == get_bits_left(orig_gb));
1679 /* superframe header */
1680 if (get_bits_left(gb) < 14)
1683 return AVERROR(ENOSYS); // WMAPro-in-WMAVoice superframe
1684 if (get_bits1(gb)) skip_bits(gb, 12); // number of samples in superframe
1685 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1686 if (get_bits_left(gb) < s->sframe_lsp_bitsize)
1688 skip_bits_long(gb, s->sframe_lsp_bitsize);
1692 for (n = 0; n < MAX_FRAMES; n++) {
1693 int aw_idx_is_ext = 0;
1695 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1696 if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
1697 skip_bits_long(gb, s->frame_lsp_bitsize);
1699 bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
1701 return AVERROR_INVALIDDATA; // invalid frame type VLC code
1702 frame_desc = &frame_descs[bd_idx];
1703 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1704 if (get_bits_left(gb) < s->pitch_nbits)
1706 skip_bits_long(gb, s->pitch_nbits);
1708 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1710 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1711 int tmp = get_bits(gb, 6);
1719 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1720 need_bits = s->block_pitch_nbits +
1721 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1722 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1723 need_bits = 2 * !aw_idx_is_ext;
1726 need_bits += frame_desc->frame_size;
1727 if (get_bits_left(gb) < need_bits)
1729 skip_bits_long(gb, need_bits);
1736 * Synthesize output samples for a single superframe. If we have any data
1737 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1740 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1741 * to give a total of 480 samples per frame. See #synth_frame() for frame
1742 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1743 * (if these are globally specified for all frames (residually); they can
1744 * also be specified individually per-frame. See the s->has_residual_lsps
1745 * option), and can specify the number of samples encoded in this superframe
1746 * (if less than 480), usually used to prevent blanks at track boundaries.
1748 * @param ctx WMA Voice decoder context
1749 * @return 0 on success, <0 on error or 1 if there was not enough data to
1750 * fully parse the superframe
1752 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1755 WMAVoiceContext *s = ctx->priv_data;
1756 GetBitContext *gb = &s->gb, s_gb;
1757 int n, res, n_samples = 480;
1758 double lsps[MAX_FRAMES][MAX_LSPS];
1759 const double *mean_lsf = s->lsps == 16 ?
1760 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1761 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1762 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1765 memcpy(synth, s->synth_history,
1766 s->lsps * sizeof(*synth));
1767 memcpy(excitation, s->excitation_history,
1768 s->history_nsamples * sizeof(*excitation));
1770 if (s->sframe_cache_size > 0) {
1772 init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
1773 s->sframe_cache_size = 0;
1776 if ((res = check_bits_for_superframe(gb, s)) == 1) {
1782 /* First bit is speech/music bit, it differentiates between WMAVoice
1783 * speech samples (the actual codec) and WMAVoice music samples, which
1784 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1786 if (!get_bits1(gb)) {
1787 avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1788 return AVERROR_PATCHWELCOME;
1791 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1792 if (get_bits1(gb)) {
1793 if ((n_samples = get_bits(gb, 12)) > 480) {
1794 av_log(ctx, AV_LOG_ERROR,
1795 "Superframe encodes >480 samples (%d), not allowed\n",
1797 return AVERROR_INVALIDDATA;
1800 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1801 if (s->has_residual_lsps) {
1802 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1804 for (n = 0; n < s->lsps; n++)
1805 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1807 if (s->lsps == 10) {
1808 dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1809 } else /* s->lsps == 16 */
1810 dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1812 for (n = 0; n < s->lsps; n++) {
1813 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1814 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1815 lsps[2][n] += mean_lsf[n];
1817 for (n = 0; n < 3; n++)
1818 stabilize_lsps(lsps[n], s->lsps);
1821 /* get output buffer */
1822 frame->nb_samples = 480;
1823 if ((res = ff_get_buffer(ctx, frame, 0)) < 0) {
1824 av_log(ctx, AV_LOG_ERROR, "get_buffer() failed\n");
1827 frame->nb_samples = n_samples;
1828 samples = (float *)frame->data[0];
1830 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1831 for (n = 0; n < 3; n++) {
1832 if (!s->has_residual_lsps) {
1835 if (s->lsps == 10) {
1836 dequant_lsp10i(gb, lsps[n]);
1837 } else /* s->lsps == 16 */
1838 dequant_lsp16i(gb, lsps[n]);
1840 for (m = 0; m < s->lsps; m++)
1841 lsps[n][m] += mean_lsf[m];
1842 stabilize_lsps(lsps[n], s->lsps);
1845 if ((res = synth_frame(ctx, gb, n,
1846 &samples[n * MAX_FRAMESIZE],
1847 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1848 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1849 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1855 /* Statistics? FIXME - we don't check for length, a slight overrun
1856 * will be caught by internal buffer padding, and anything else
1857 * will be skipped, not read. */
1858 if (get_bits1(gb)) {
1859 res = get_bits(gb, 4);
1860 skip_bits(gb, 10 * (res + 1));
1865 /* Update history */
1866 memcpy(s->prev_lsps, lsps[2],
1867 s->lsps * sizeof(*s->prev_lsps));
1868 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1869 s->lsps * sizeof(*synth));
1870 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1871 s->history_nsamples * sizeof(*excitation));
1873 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1874 s->history_nsamples * sizeof(*s->zero_exc_pf));
1880 * Parse the packet header at the start of each packet (input data to this
1883 * @param s WMA Voice decoding context private data
1884 * @return 1 if not enough bits were available, or 0 on success.
1886 static int parse_packet_header(WMAVoiceContext *s)
1888 GetBitContext *gb = &s->gb;
1891 if (get_bits_left(gb) < 11)
1893 skip_bits(gb, 4); // packet sequence number
1894 s->has_residual_lsps = get_bits1(gb);
1896 res = get_bits(gb, 6); // number of superframes per packet
1897 // (minus first one if there is spillover)
1898 if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
1900 } while (res == 0x3F);
1901 s->spillover_nbits = get_bits(gb, s->spillover_bitsize);
1907 * Copy (unaligned) bits from gb/data/size to pb.
1909 * @param pb target buffer to copy bits into
1910 * @param data source buffer to copy bits from
1911 * @param size size of the source data, in bytes
1912 * @param gb bit I/O context specifying the current position in the source.
1913 * data. This function might use this to align the bit position to
1914 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1916 * @param nbits the amount of bits to copy from source to target
1918 * @note after calling this function, the current position in the input bit
1919 * I/O context is undefined.
1921 static void copy_bits(PutBitContext *pb,
1922 const uint8_t *data, int size,
1923 GetBitContext *gb, int nbits)
1925 int rmn_bytes, rmn_bits;
1927 rmn_bits = rmn_bytes = get_bits_left(gb);
1928 if (rmn_bits < nbits)
1930 if (nbits > pb->size_in_bits - put_bits_count(pb))
1932 rmn_bits &= 7; rmn_bytes >>= 3;
1933 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1934 put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
1935 avpriv_copy_bits(pb, data + size - rmn_bytes,
1936 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1940 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1941 * and we expect that the demuxer / application provides it to us as such
1942 * (else you'll probably get garbage as output). Every packet has a size of
1943 * ctx->block_align bytes, starts with a packet header (see
1944 * #parse_packet_header()), and then a series of superframes. Superframe
1945 * boundaries may exceed packets, i.e. superframes can split data over
1946 * multiple (two) packets.
1948 * For more information about frames, see #synth_superframe().
1950 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1951 int *got_frame_ptr, AVPacket *avpkt)
1953 WMAVoiceContext *s = ctx->priv_data;
1954 GetBitContext *gb = &s->gb;
1957 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1958 * header at each ctx->block_align bytes. However, Libav's ASF demuxer
1959 * feeds us ASF packets, which may concatenate multiple "codec" packets
1960 * in a single "muxer" packet, so we artificially emulate that by
1961 * capping the packet size at ctx->block_align. */
1962 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1967 init_get_bits(&s->gb, avpkt->data, size << 3);
1969 /* size == ctx->block_align is used to indicate whether we are dealing with
1970 * a new packet or a packet of which we already read the packet header
1972 if (size == ctx->block_align) { // new packet header
1973 if ((res = parse_packet_header(s)) < 0)
1976 /* If the packet header specifies a s->spillover_nbits, then we want
1977 * to push out all data of the previous packet (+ spillover) before
1978 * continuing to parse new superframes in the current packet. */
1979 if (s->spillover_nbits > 0) {
1980 if (s->sframe_cache_size > 0) {
1981 int cnt = get_bits_count(gb);
1982 copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
1983 flush_put_bits(&s->pb);
1984 s->sframe_cache_size += s->spillover_nbits;
1985 if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
1987 cnt += s->spillover_nbits;
1988 s->skip_bits_next = cnt & 7;
1991 skip_bits_long (gb, s->spillover_nbits - cnt +
1992 get_bits_count(gb)); // resync
1994 skip_bits_long(gb, s->spillover_nbits); // resync
1996 } else if (s->skip_bits_next)
1997 skip_bits(gb, s->skip_bits_next);
1999 /* Try parsing superframes in current packet */
2000 s->sframe_cache_size = 0;
2001 s->skip_bits_next = 0;
2002 pos = get_bits_left(gb);
2003 if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
2005 } else if (*got_frame_ptr) {
2006 int cnt = get_bits_count(gb);
2007 s->skip_bits_next = cnt & 7;
2009 } else if ((s->sframe_cache_size = pos) > 0) {
2010 /* rewind bit reader to start of last (incomplete) superframe... */
2011 init_get_bits(gb, avpkt->data, size << 3);
2012 skip_bits_long(gb, (size << 3) - pos);
2013 assert(get_bits_left(gb) == pos);
2015 /* ...and cache it for spillover in next packet */
2016 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
2017 copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
2018 // FIXME bad - just copy bytes as whole and add use the
2019 // skip_bits_next field
2025 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2027 WMAVoiceContext *s = ctx->priv_data;
2030 ff_rdft_end(&s->rdft);
2031 ff_rdft_end(&s->irdft);
2032 ff_dct_end(&s->dct);
2033 ff_dct_end(&s->dst);
2039 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2041 WMAVoiceContext *s = ctx->priv_data;
2044 s->postfilter_agc = 0;
2045 s->sframe_cache_size = 0;
2046 s->skip_bits_next = 0;
2047 for (n = 0; n < s->lsps; n++)
2048 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2049 memset(s->excitation_history, 0,
2050 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2051 memset(s->synth_history, 0,
2052 sizeof(*s->synth_history) * MAX_LSPS);
2053 memset(s->gain_pred_err, 0,
2054 sizeof(s->gain_pred_err));
2057 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2058 sizeof(*s->synth_filter_out_buf) * s->lsps);
2059 memset(s->dcf_mem, 0,
2060 sizeof(*s->dcf_mem) * 2);
2061 memset(s->zero_exc_pf, 0,
2062 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2063 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2067 AVCodec ff_wmavoice_decoder = {
2069 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
2070 .type = AVMEDIA_TYPE_AUDIO,
2071 .id = AV_CODEC_ID_WMAVOICE,
2072 .priv_data_size = sizeof(WMAVoiceContext),
2073 .init = wmavoice_decode_init,
2074 .init_static_data = wmavoice_init_static_data,
2075 .close = wmavoice_decode_end,
2076 .decode = wmavoice_decode_packet,
2077 .capabilities = CODEC_CAP_SUBFRAMES | CODEC_CAP_DR1,
2078 .flush = wmavoice_flush,