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>
30 #include "libavutil/channel_layout.h"
31 #include "libavutil/float_dsp.h"
32 #include "libavutil/mem.h"
35 #include "bitstream.h"
38 #include "wmavoice_data.h"
39 #include "celp_filters.h"
40 #include "acelp_vectors.h"
41 #include "acelp_filters.h"
47 #define MAX_BLOCKS 8 ///< maximum number of blocks per frame
48 #define MAX_LSPS 16 ///< maximum filter order
49 #define MAX_LSPS_ALIGN16 16 ///< same as #MAX_LSPS; needs to be multiple
50 ///< of 16 for ASM input buffer alignment
51 #define MAX_FRAMES 3 ///< maximum number of frames per superframe
52 #define MAX_FRAMESIZE 160 ///< maximum number of samples per frame
53 #define MAX_SIGNAL_HISTORY 416 ///< maximum excitation signal history
54 #define MAX_SFRAMESIZE (MAX_FRAMESIZE * MAX_FRAMES)
55 ///< maximum number of samples per superframe
56 #define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
57 ///< was split over two packets
58 #define VLC_NBITS 6 ///< number of bits to read per VLC iteration
61 * Frame type VLC coding.
63 static VLC frame_type_vlc;
66 * Adaptive codebook types.
69 ACB_TYPE_NONE = 0, ///< no adaptive codebook (only hardcoded fixed)
70 ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
71 ///< we interpolate to get a per-sample pitch.
72 ///< Signal is generated using an asymmetric sinc
74 ///< @note see #wmavoice_ipol1_coeffs
75 ACB_TYPE_HAMMING = 2 ///< Per-block pitch with signal generation using
76 ///< a Hamming sinc window function
77 ///< @note see #wmavoice_ipol2_coeffs
81 * Fixed codebook types.
84 FCB_TYPE_SILENCE = 0, ///< comfort noise during silence
85 ///< generated from a hardcoded (fixed) codebook
86 ///< with per-frame (low) gain values
87 FCB_TYPE_HARDCODED = 1, ///< hardcoded (fixed) codebook with per-block
89 FCB_TYPE_AW_PULSES = 2, ///< Pitch-adaptive window (AW) pulse signals,
90 ///< used in particular for low-bitrate streams
91 FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
92 ///< combinations of either single pulses or
97 * Description of frame types.
99 static const struct frame_type_desc {
100 uint8_t n_blocks; ///< amount of blocks per frame (each block
101 ///< (contains 160/#n_blocks samples)
102 uint8_t log_n_blocks; ///< log2(#n_blocks)
103 uint8_t acb_type; ///< Adaptive codebook type (ACB_TYPE_*)
104 uint8_t fcb_type; ///< Fixed codebook type (FCB_TYPE_*)
105 uint8_t dbl_pulses; ///< how many pulse vectors have pulse pairs
106 ///< (rather than just one single pulse)
107 ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
108 uint16_t frame_size; ///< the amount of bits that make up the block
109 ///< data (per frame)
110 } frame_descs[17] = {
111 { 1, 0, ACB_TYPE_NONE, FCB_TYPE_SILENCE, 0, 0 },
112 { 2, 1, ACB_TYPE_NONE, FCB_TYPE_HARDCODED, 0, 28 },
113 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES, 0, 46 },
114 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 80 },
115 { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
116 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
117 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
118 { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
119 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 64 },
120 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 80 },
121 { 2, 1, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 104 },
122 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 108 },
123 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 132 },
124 { 4, 2, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 168 },
125 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 0, 176 },
126 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 2, 208 },
127 { 8, 3, ACB_TYPE_HAMMING, FCB_TYPE_EXC_PULSES, 5, 256 }
131 * WMA Voice decoding context.
133 typedef struct WMAVoiceContext {
135 * @name Global values specified in the stream header / extradata or used all over.
138 BitstreamContext bc; ///< packet bitreader. During decoder init,
139 ///< it contains the extradata from the
140 ///< demuxer. During decoding, it contains
142 int8_t vbm_tree[25]; ///< converts VLC codes to frame type
144 int spillover_bitsize; ///< number of bits used to specify
145 ///< #spillover_nbits in the packet header
146 ///< = ceil(log2(ctx->block_align << 3))
147 int history_nsamples; ///< number of samples in history for signal
148 ///< prediction (through ACB)
150 /* postfilter specific values */
151 int do_apf; ///< whether to apply the averaged
152 ///< projection filter (APF)
153 int denoise_strength; ///< strength of denoising in Wiener filter
155 int denoise_tilt_corr; ///< Whether to apply tilt correction to the
156 ///< Wiener filter coefficients (postfilter)
157 int dc_level; ///< Predicted amount of DC noise, based
158 ///< on which a DC removal filter is used
160 int lsps; ///< number of LSPs per frame [10 or 16]
161 int lsp_q_mode; ///< defines quantizer defaults [0, 1]
162 int lsp_def_mode; ///< defines different sets of LSP defaults
164 int frame_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
165 ///< per-frame (independent coding)
166 int sframe_lsp_bitsize; ///< size (in bits) of LSPs, when encoded
167 ///< per superframe (residual coding)
169 int min_pitch_val; ///< base value for pitch parsing code
170 int max_pitch_val; ///< max value + 1 for pitch parsing
171 int pitch_nbits; ///< number of bits used to specify the
172 ///< pitch value in the frame header
173 int block_pitch_nbits; ///< number of bits used to specify the
174 ///< first block's pitch value
175 int block_pitch_range; ///< range of the block pitch
176 int block_delta_pitch_nbits; ///< number of bits used to specify the
177 ///< delta pitch between this and the last
178 ///< block's pitch value, used in all but
180 int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
181 ///< from -this to +this-1)
182 uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
188 * @name Packet values specified in the packet header or related to a packet.
190 * A packet is considered to be a single unit of data provided to this
191 * decoder by the demuxer.
194 int spillover_nbits; ///< number of bits of the previous packet's
195 ///< last superframe preceding this
196 ///< packet's first full superframe (useful
197 ///< for re-synchronization also)
198 int has_residual_lsps; ///< if set, superframes contain one set of
199 ///< LSPs that cover all frames, encoded as
200 ///< independent and residual LSPs; if not
201 ///< set, each frame contains its own, fully
202 ///< independent, LSPs
203 int skip_bits_next; ///< number of bits to skip at the next call
204 ///< to #wmavoice_decode_packet() (since
205 ///< they're part of the previous superframe)
207 uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + AV_INPUT_BUFFER_PADDING_SIZE];
208 ///< cache for superframe data split over
209 ///< multiple packets
210 int sframe_cache_size; ///< set to >0 if we have data from an
211 ///< (incomplete) superframe from a previous
212 ///< packet that spilled over in the current
213 ///< packet; specifies the amount of bits in
215 PutBitContext pb; ///< bitstream writer for #sframe_cache
220 * @name Frame and superframe values
221 * Superframe and frame data - these can change from frame to frame,
222 * although some of them do in that case serve as a cache / history for
223 * the next frame or superframe.
226 double prev_lsps[MAX_LSPS]; ///< LSPs of the last frame of the previous
228 int last_pitch_val; ///< pitch value of the previous frame
229 int last_acb_type; ///< frame type [0-2] of the previous frame
230 int pitch_diff_sh16; ///< ((cur_pitch_val - #last_pitch_val)
231 ///< << 16) / #MAX_FRAMESIZE
232 float silence_gain; ///< set for use in blocks if #ACB_TYPE_NONE
234 int aw_idx_is_ext; ///< whether the AW index was encoded in
235 ///< 8 bits (instead of 6)
236 int aw_pulse_range; ///< the range over which #aw_pulse_set1()
237 ///< can apply the pulse, relative to the
238 ///< value in aw_first_pulse_off. The exact
239 ///< position of the first AW-pulse is within
240 ///< [pulse_off, pulse_off + this], and
241 ///< depends on bitstream values; [16 or 24]
242 int aw_n_pulses[2]; ///< number of AW-pulses in each block; note
243 ///< that this number can be negative (in
244 ///< which case it basically means "zero")
245 int aw_first_pulse_off[2]; ///< index of first sample to which to
246 ///< apply AW-pulses, or -0xff if unset
247 int aw_next_pulse_off_cache; ///< the position (relative to start of the
248 ///< second block) at which pulses should
249 ///< start to be positioned, serves as a
250 ///< cache for pitch-adaptive window pulses
253 int frame_cntr; ///< current frame index [0 - 0xFFFE]; is
254 ///< only used for comfort noise in #pRNG()
255 float gain_pred_err[6]; ///< cache for gain prediction
256 float excitation_history[MAX_SIGNAL_HISTORY];
257 ///< cache of the signal of previous
258 ///< superframes, used as a history for
259 ///< signal generation
260 float synth_history[MAX_LSPS]; ///< see #excitation_history
264 * @name Postfilter values
266 * Variables used for postfilter implementation, mostly history for
267 * smoothing and so on, and context variables for FFT/iFFT.
270 RDFTContext rdft, irdft; ///< contexts for FFT-calculation in the
271 ///< postfilter (for denoise filter)
272 DCTContext dct, dst; ///< contexts for phase shift (in Hilbert
273 ///< transform, part of postfilter)
274 float sin[511], cos[511]; ///< 8-bit cosine/sine windows over [-pi,pi]
276 float postfilter_agc; ///< gain control memory, used in
277 ///< #adaptive_gain_control()
278 float dcf_mem[2]; ///< DC filter history
279 float zero_exc_pf[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE];
280 ///< zero filter output (i.e. excitation)
282 float denoise_filter_cache[MAX_FRAMESIZE];
283 int denoise_filter_cache_size; ///< samples in #denoise_filter_cache
284 DECLARE_ALIGNED(32, float, tilted_lpcs_pf)[0x80];
285 ///< aligned buffer for LPC tilting
286 DECLARE_ALIGNED(32, float, denoise_coeffs_pf)[0x80];
287 ///< aligned buffer for denoise coefficients
288 DECLARE_ALIGNED(32, float, synth_filter_out_buf)[0x80 + MAX_LSPS_ALIGN16];
289 ///< aligned buffer for postfilter speech
297 * Set up the variable bit mode (VBM) tree from container extradata.
298 * @param bc bit I/O context.
299 * The bit context (s->bc) should be loaded with byte 23-46 of the
300 * container extradata (i.e. the ones containing the VBM tree).
301 * @param vbm_tree pointer to array to which the decoded VBM tree will be
303 * @return 0 on success, <0 on error.
305 static av_cold int decode_vbmtree(BitstreamContext *bc, int8_t vbm_tree[25])
307 int cntr[8] = { 0 }, n, res;
309 memset(vbm_tree, 0xff, sizeof(vbm_tree[0]) * 25);
310 for (n = 0; n < 17; n++) {
311 res = bitstream_read(bc, 3);
312 if (cntr[res] > 3) // should be >= 3 + (res == 7))
314 vbm_tree[res * 3 + cntr[res]++] = n;
319 static av_cold void wmavoice_init_static_data(AVCodec *codec)
321 static const uint8_t bits[] = {
324 10, 10, 10, 12, 12, 12,
327 static const uint16_t codes[] = {
328 0x0000, 0x0001, 0x0002, // 00/01/10
329 0x000c, 0x000d, 0x000e, // 11+00/01/10
330 0x003c, 0x003d, 0x003e, // 1111+00/01/10
331 0x00fc, 0x00fd, 0x00fe, // 111111+00/01/10
332 0x03fc, 0x03fd, 0x03fe, // 11111111+00/01/10
333 0x0ffc, 0x0ffd, 0x0ffe, // 1111111111+00/01/10
334 0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
337 INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
338 bits, 1, 1, codes, 2, 2, 132);
342 * Set up decoder with parameters from demuxer (extradata etc.).
344 static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
346 int n, flags, pitch_range, lsp16_flag;
347 WMAVoiceContext *s = ctx->priv_data;
351 * - byte 0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
352 * - byte 19-22: flags field (annoyingly in LE; see below for known
354 * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
357 if (ctx->extradata_size != 46) {
358 av_log(ctx, AV_LOG_ERROR,
359 "Invalid extradata size %d (should be 46)\n",
360 ctx->extradata_size);
361 return AVERROR_INVALIDDATA;
363 flags = AV_RL32(ctx->extradata + 18);
364 s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
365 s->do_apf = flags & 0x1;
367 ff_rdft_init(&s->rdft, 7, DFT_R2C);
368 ff_rdft_init(&s->irdft, 7, IDFT_C2R);
369 ff_dct_init(&s->dct, 6, DCT_I);
370 ff_dct_init(&s->dst, 6, DST_I);
372 ff_sine_window_init(s->cos, 256);
373 memcpy(&s->sin[255], s->cos, 256 * sizeof(s->cos[0]));
374 for (n = 0; n < 255; n++) {
375 s->sin[n] = -s->sin[510 - n];
376 s->cos[510 - n] = s->cos[n];
379 s->denoise_strength = (flags >> 2) & 0xF;
380 if (s->denoise_strength >= 12) {
381 av_log(ctx, AV_LOG_ERROR,
382 "Invalid denoise filter strength %d (max=11)\n",
383 s->denoise_strength);
384 return AVERROR_INVALIDDATA;
386 s->denoise_tilt_corr = !!(flags & 0x40);
387 s->dc_level = (flags >> 7) & 0xF;
388 s->lsp_q_mode = !!(flags & 0x2000);
389 s->lsp_def_mode = !!(flags & 0x4000);
390 lsp16_flag = flags & 0x1000;
393 s->frame_lsp_bitsize = 34;
394 s->sframe_lsp_bitsize = 60;
397 s->frame_lsp_bitsize = 24;
398 s->sframe_lsp_bitsize = 48;
400 for (n = 0; n < s->lsps; n++)
401 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
403 bitstream_init8(&s->bc, ctx->extradata + 22, ctx->extradata_size - 22);
404 if (decode_vbmtree(&s->bc, s->vbm_tree) < 0) {
405 av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
406 return AVERROR_INVALIDDATA;
409 s->min_pitch_val = ((ctx->sample_rate << 8) / 400 + 50) >> 8;
410 s->max_pitch_val = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
411 pitch_range = s->max_pitch_val - s->min_pitch_val;
412 if (pitch_range <= 0) {
413 av_log(ctx, AV_LOG_ERROR, "Invalid pitch range; broken extradata?\n");
414 return AVERROR_INVALIDDATA;
416 s->pitch_nbits = av_ceil_log2(pitch_range);
417 s->last_pitch_val = 40;
418 s->last_acb_type = ACB_TYPE_NONE;
419 s->history_nsamples = s->max_pitch_val + 8;
421 if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
422 int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
423 max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;
425 av_log(ctx, AV_LOG_ERROR,
426 "Unsupported samplerate %d (min=%d, max=%d)\n",
427 ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz
429 return AVERROR(ENOSYS);
432 s->block_conv_table[0] = s->min_pitch_val;
433 s->block_conv_table[1] = (pitch_range * 25) >> 6;
434 s->block_conv_table[2] = (pitch_range * 44) >> 6;
435 s->block_conv_table[3] = s->max_pitch_val - 1;
436 s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
437 if (s->block_delta_pitch_hrange <= 0) {
438 av_log(ctx, AV_LOG_ERROR, "Invalid delta pitch hrange; broken extradata?\n");
439 return AVERROR_INVALIDDATA;
441 s->block_delta_pitch_nbits = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
442 s->block_pitch_range = s->block_conv_table[2] +
443 s->block_conv_table[3] + 1 +
444 2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
445 s->block_pitch_nbits = av_ceil_log2(s->block_pitch_range);
448 ctx->channel_layout = AV_CH_LAYOUT_MONO;
449 ctx->sample_fmt = AV_SAMPLE_FMT_FLT;
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 = avpriv_scalarproduct_float_c(in, ptr, size);
527 if (dot > optimal_gain) {
531 } while (--ptr >= end);
533 if (optimal_gain <= 0)
535 dot = avpriv_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 + avpriv_scalarproduct_float_c(lpcs, lpcs, n_lpcs);
566 rh1 = lpcs[0] + avpriv_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) { // fall back 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 sine 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 / avpriv_scalarproduct_float_c(coeffs, coeffs,
661 for (n = 0; n < remainder; n++)
666 * This function applies a Wiener filter on the (noisy) speech signal as
667 * a means to denoise it.
669 * - take RDFT of LPCs to get the power spectrum of the noise + speech;
670 * - using this power spectrum, calculate (for each frequency) the Wiener
671 * filter gain, which depends on the frequency power and desired level
672 * of noise subtraction (when set too high, this leads to artifacts)
673 * We can do this symmetrically over the X-axis (so 0-4kHz is the inverse
675 * - by doing a phase shift, calculate the Hilbert transform of this array
676 * of per-frequency filter-gains to get the filtering coefficients;
677 * - smoothen/normalize/de-tilt these filter coefficients as desired;
678 * - take RDFT of noisy sound, apply the coefficients and take its IRDFT
679 * to get the denoised speech signal;
680 * - the leftover (i.e. output of the IRDFT on denoised speech data beyond
681 * the frame boundary) are saved and applied to subsequent frames by an
682 * overlap-add method (otherwise you get clicking-artifacts).
684 * @param s WMA Voice decoding context
685 * @param fcb_type Frame (codebook) type
686 * @param synth_pf input: the noisy speech signal, output: denoised speech
687 * data; should be 16-byte aligned (for ASM purposes)
688 * @param size size of the speech data
689 * @param lpcs LPCs used to synthesize this frame's speech data
691 static void wiener_denoise(WMAVoiceContext *s, int fcb_type,
692 float *synth_pf, int size,
695 int remainder, lim, n;
697 if (fcb_type != FCB_TYPE_SILENCE) {
698 float *tilted_lpcs = s->tilted_lpcs_pf,
699 *coeffs = s->denoise_coeffs_pf, tilt_mem = 0;
701 tilted_lpcs[0] = 1.0;
702 memcpy(&tilted_lpcs[1], lpcs, sizeof(lpcs[0]) * s->lsps);
703 memset(&tilted_lpcs[s->lsps + 1], 0,
704 sizeof(tilted_lpcs[0]) * (128 - s->lsps - 1));
705 ff_tilt_compensation(&tilt_mem, 0.7 * tilt_factor(lpcs, s->lsps),
706 tilted_lpcs, s->lsps + 2);
708 /* The IRDFT output (127 samples for 7-bit filter) beyond the frame
709 * size is applied to the next frame. All input beyond this is zero,
710 * and thus all output beyond this will go towards zero, hence we can
711 * limit to min(size-1, 127-size) as a performance consideration. */
712 remainder = FFMIN(127 - size, size - 1);
713 calc_input_response(s, tilted_lpcs, fcb_type, coeffs, remainder);
715 /* apply coefficients (in frequency spectrum domain), i.e. complex
716 * number multiplication */
717 memset(&synth_pf[size], 0, sizeof(synth_pf[0]) * (128 - size));
718 s->rdft.rdft_calc(&s->rdft, synth_pf);
719 s->rdft.rdft_calc(&s->rdft, coeffs);
720 synth_pf[0] *= coeffs[0];
721 synth_pf[1] *= coeffs[1];
722 for (n = 1; n < 64; n++) {
723 float v1 = synth_pf[n * 2], v2 = synth_pf[n * 2 + 1];
724 synth_pf[n * 2] = v1 * coeffs[n * 2] - v2 * coeffs[n * 2 + 1];
725 synth_pf[n * 2 + 1] = v2 * coeffs[n * 2] + v1 * coeffs[n * 2 + 1];
727 s->irdft.rdft_calc(&s->irdft, synth_pf);
730 /* merge filter output with the history of previous runs */
731 if (s->denoise_filter_cache_size) {
732 lim = FFMIN(s->denoise_filter_cache_size, size);
733 for (n = 0; n < lim; n++)
734 synth_pf[n] += s->denoise_filter_cache[n];
735 s->denoise_filter_cache_size -= lim;
736 memmove(s->denoise_filter_cache, &s->denoise_filter_cache[size],
737 sizeof(s->denoise_filter_cache[0]) * s->denoise_filter_cache_size);
740 /* move remainder of filter output into a cache for future runs */
741 if (fcb_type != FCB_TYPE_SILENCE) {
742 lim = FFMIN(remainder, s->denoise_filter_cache_size);
743 for (n = 0; n < lim; n++)
744 s->denoise_filter_cache[n] += synth_pf[size + n];
745 if (lim < remainder) {
746 memcpy(&s->denoise_filter_cache[lim], &synth_pf[size + lim],
747 sizeof(s->denoise_filter_cache[0]) * (remainder - lim));
748 s->denoise_filter_cache_size = remainder;
754 * Averaging projection filter, the postfilter used in WMAVoice.
756 * This uses the following steps:
757 * - A zero-synthesis filter (generate excitation from synth signal)
758 * - Kalman smoothing on excitation, based on pitch
759 * - Re-synthesized smoothened output
760 * - Iterative Wiener denoise filter
761 * - Adaptive gain filter
764 * @param s WMAVoice decoding context
765 * @param synth Speech synthesis output (before postfilter)
766 * @param samples Output buffer for filtered samples
767 * @param size Buffer size of synth & samples
768 * @param lpcs Generated LPCs used for speech synthesis
769 * @param zero_exc_pf destination for zero synthesis filter (16-byte aligned)
770 * @param fcb_type Frame type (silence, hardcoded, AW-pulses or FCB-pulses)
771 * @param pitch Pitch of the input signal
773 static void postfilter(WMAVoiceContext *s, const float *synth,
774 float *samples, int size,
775 const float *lpcs, float *zero_exc_pf,
776 int fcb_type, int pitch)
778 float synth_filter_in_buf[MAX_FRAMESIZE / 2],
779 *synth_pf = &s->synth_filter_out_buf[MAX_LSPS_ALIGN16],
780 *synth_filter_in = zero_exc_pf;
782 assert(size <= MAX_FRAMESIZE / 2);
784 /* generate excitation from input signal */
785 ff_celp_lp_zero_synthesis_filterf(zero_exc_pf, lpcs, synth, size, s->lsps);
787 if (fcb_type >= FCB_TYPE_AW_PULSES &&
788 !kalman_smoothen(s, pitch, zero_exc_pf, synth_filter_in_buf, size))
789 synth_filter_in = synth_filter_in_buf;
791 /* re-synthesize speech after smoothening, and keep history */
792 ff_celp_lp_synthesis_filterf(synth_pf, lpcs,
793 synth_filter_in, size, s->lsps);
794 memcpy(&synth_pf[-s->lsps], &synth_pf[size - s->lsps],
795 sizeof(synth_pf[0]) * s->lsps);
797 wiener_denoise(s, fcb_type, synth_pf, size, lpcs);
799 adaptive_gain_control(samples, synth_pf, synth, size, 0.99,
802 if (s->dc_level > 8) {
803 /* remove ultra-low frequency DC noise / highpass filter;
804 * coefficients are identical to those used in SIPR decoding,
805 * and very closely resemble those used in AMR-NB decoding. */
806 ff_acelp_apply_order_2_transfer_function(samples, samples,
807 (const float[2]) { -1.99997, 1.0 },
808 (const float[2]) { -1.9330735188, 0.93589198496 },
809 0.93980580475, s->dcf_mem, size);
818 * @param lsps output pointer to the array that will hold the LSPs
819 * @param num number of LSPs to be dequantized
820 * @param values quantized values, contains n_stages values
821 * @param sizes range (i.e. max value) of each quantized value
822 * @param n_stages number of dequantization runs
823 * @param table dequantization table to be used
824 * @param mul_q LSF multiplier
825 * @param base_q base (lowest) LSF values
827 static void dequant_lsps(double *lsps, int num,
828 const uint16_t *values,
829 const uint16_t *sizes,
830 int n_stages, const uint8_t *table,
832 const double *base_q)
836 memset(lsps, 0, num * sizeof(*lsps));
837 for (n = 0; n < n_stages; n++) {
838 const uint8_t *t_off = &table[values[n] * num];
839 double base = base_q[n], mul = mul_q[n];
841 for (m = 0; m < num; m++)
842 lsps[m] += base + mul * t_off[m];
844 table += sizes[n] * num;
849 * @name LSP dequantization routines
850 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
851 * @note we assume enough bits are available, caller should check.
852 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
853 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
857 * Parse 10 independently-coded LSPs.
859 static void dequant_lsp10i(BitstreamContext *bc, double *lsps)
861 static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
862 static const double mul_lsf[4] = {
863 5.2187144800e-3, 1.4626986422e-3,
864 9.6179549166e-4, 1.1325736225e-3
866 static const double base_lsf[4] = {
867 M_PI * -2.15522e-1, M_PI * -6.1646e-2,
868 M_PI * -3.3486e-2, M_PI * -5.7408e-2
872 v[0] = bitstream_read(bc, 8);
873 v[1] = bitstream_read(bc, 6);
874 v[2] = bitstream_read(bc, 5);
875 v[3] = bitstream_read(bc, 5);
877 dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
882 * Parse 10 independently-coded LSPs, and then derive the tables to
883 * generate LSPs for the other frames from them (residual coding).
885 static void dequant_lsp10r(BitstreamContext *bc,
886 double *i_lsps, const double *old,
887 double *a1, double *a2, int q_mode)
889 static const uint16_t vec_sizes[3] = { 128, 64, 64 };
890 static const double mul_lsf[3] = {
891 2.5807601174e-3, 1.2354460219e-3, 1.1763821673e-3
893 static const double base_lsf[3] = {
894 M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
896 const float (*ipol_tab)[2][10] = q_mode ?
897 wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
898 uint16_t interpol, v[3];
901 dequant_lsp10i(bc, i_lsps);
903 interpol = bitstream_read(bc, 5);
904 v[0] = bitstream_read(bc, 7);
905 v[1] = bitstream_read(bc, 6);
906 v[2] = bitstream_read(bc, 6);
908 for (n = 0; n < 10; n++) {
909 double delta = old[n] - i_lsps[n];
910 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
911 a1[10 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
914 dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
919 * Parse 16 independently-coded LSPs.
921 static void dequant_lsp16i(BitstreamContext *bc, double *lsps)
923 static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
924 static const double mul_lsf[5] = {
925 3.3439586280e-3, 6.9908173703e-4,
926 3.3216608306e-3, 1.0334960326e-3,
929 static const double base_lsf[5] = {
930 M_PI * -1.27576e-1, M_PI * -2.4292e-2,
931 M_PI * -1.28094e-1, M_PI * -3.2128e-2,
936 v[0] = bitstream_read(bc, 8);
937 v[1] = bitstream_read(bc, 6);
938 v[2] = bitstream_read(bc, 7);
939 v[3] = bitstream_read(bc, 6);
940 v[4] = bitstream_read(bc, 7);
942 dequant_lsps( lsps, 5, v, vec_sizes, 2,
943 wmavoice_dq_lsp16i1, mul_lsf, base_lsf);
944 dequant_lsps(&lsps[5], 5, &v[2], &vec_sizes[2], 2,
945 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
946 dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
947 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
951 * Parse 16 independently-coded LSPs, and then derive the tables to
952 * generate LSPs for the other frames from them (residual coding).
954 static void dequant_lsp16r(BitstreamContext *bc,
955 double *i_lsps, const double *old,
956 double *a1, double *a2, int q_mode)
958 static const uint16_t vec_sizes[3] = { 128, 128, 128 };
959 static const double mul_lsf[3] = {
960 1.2232979501e-3, 1.4062241527e-3, 1.6114744851e-3
962 static const double base_lsf[3] = {
963 M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
965 const float (*ipol_tab)[2][16] = q_mode ?
966 wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
967 uint16_t interpol, v[3];
970 dequant_lsp16i(bc, i_lsps);
972 interpol = bitstream_read(bc, 5);
973 v[0] = bitstream_read(bc, 7);
974 v[1] = bitstream_read(bc, 7);
975 v[2] = bitstream_read(bc, 7);
977 for (n = 0; n < 16; n++) {
978 double delta = old[n] - i_lsps[n];
979 a1[n] = ipol_tab[interpol][0][n] * delta + i_lsps[n];
980 a1[16 + n] = ipol_tab[interpol][1][n] * delta + i_lsps[n];
983 dequant_lsps( a2, 10, v, vec_sizes, 1,
984 wmavoice_dq_lsp16r1, mul_lsf, base_lsf);
985 dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
986 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
987 dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
988 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
993 * @name Pitch-adaptive window coding functions
994 * The next few functions are for pitch-adaptive window coding.
998 * Parse the offset of the first pitch-adaptive window pulses, and
999 * the distribution of pulses between the two blocks in this frame.
1000 * @param s WMA Voice decoding context private data
1001 * @param bc bit I/O context
1002 * @param pitch pitch for each block in this frame
1004 static void aw_parse_coords(WMAVoiceContext *s, BitstreamContext *bc,
1007 static const int16_t start_offset[94] = {
1008 -11, -9, -7, -5, -3, -1, 1, 3, 5, 7, 9, 11,
1009 13, 15, 18, 17, 19, 20, 21, 22, 23, 24, 25, 26,
1010 27, 28, 29, 30, 31, 32, 33, 35, 37, 39, 41, 43,
1011 45, 47, 49, 51, 53, 55, 57, 59, 61, 63, 65, 67,
1012 69, 71, 73, 75, 77, 79, 81, 83, 85, 87, 89, 91,
1013 93, 95, 97, 99, 101, 103, 105, 107, 109, 111, 113, 115,
1014 117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
1015 141, 143, 145, 147, 149, 151, 153, 155, 157, 159
1019 /* position of pulse */
1020 s->aw_idx_is_ext = 0;
1021 if ((bits = bitstream_read(bc, 6)) >= 54) {
1022 s->aw_idx_is_ext = 1;
1023 bits += (bits - 54) * 3 + bitstream_read(bc, 2);
1026 /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
1027 * the distribution of the pulses in each block contained in this frame. */
1028 s->aw_pulse_range = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
1029 for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
1030 s->aw_n_pulses[0] = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
1031 s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
1032 offset += s->aw_n_pulses[0] * pitch[0];
1033 s->aw_n_pulses[1] = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
1034 s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;
1036 /* if continuing from a position before the block, reset position to
1037 * start of block (when corrected for the range over which it can be
1038 * spread in aw_pulse_set1()). */
1039 if (start_offset[bits] < MAX_FRAMESIZE / 2) {
1040 while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
1041 s->aw_first_pulse_off[1] -= pitch[1];
1042 if (start_offset[bits] < 0)
1043 while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
1044 s->aw_first_pulse_off[0] -= pitch[0];
1049 * Apply second set of pitch-adaptive window pulses.
1050 * @param s WMA Voice decoding context private data
1051 * @param bc bit I/O context
1052 * @param block_idx block index in frame [0, 1]
1053 * @param fcb structure containing fixed codebook vector info
1054 * @return -1 on error, 0 otherwise
1056 static int aw_pulse_set2(WMAVoiceContext *s, BitstreamContext *bc,
1057 int block_idx, AMRFixed *fcb)
1059 uint16_t use_mask_mem[9]; // only 5 are used, rest is padding
1060 uint16_t *use_mask = use_mask_mem + 2;
1061 /* in this function, idx is the index in the 80-bit (+ padding) use_mask
1062 * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
1063 * of idx are the position of the bit within a particular item in the
1064 * array (0 being the most significant bit, and 15 being the least
1065 * significant bit), and the remainder (>> 4) is the index in the
1066 * use_mask[]-array. This is faster and uses less memory than using a
1067 * 80-byte/80-int array. */
1068 int pulse_off = s->aw_first_pulse_off[block_idx],
1069 pulse_start, n, idx, range, aidx, start_off = 0;
1071 /* set offset of first pulse to within this block */
1072 if (s->aw_n_pulses[block_idx] > 0)
1073 while (pulse_off + s->aw_pulse_range < 1)
1074 pulse_off += fcb->pitch_lag;
1076 /* find range per pulse */
1077 if (s->aw_n_pulses[0] > 0) {
1078 if (block_idx == 0) {
1080 } else /* block_idx = 1 */ {
1082 if (s->aw_n_pulses[block_idx] > 0)
1083 pulse_off = s->aw_next_pulse_off_cache;
1087 pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;
1089 /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
1090 * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
1091 * we exclude that range from being pulsed again in this function. */
1092 memset(&use_mask[-2], 0, 2 * sizeof(use_mask[0]));
1093 memset( use_mask, -1, 5 * sizeof(use_mask[0]));
1094 memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
1095 if (s->aw_n_pulses[block_idx] > 0)
1096 for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
1097 int excl_range = s->aw_pulse_range; // always 16 or 24
1098 uint16_t *use_mask_ptr = &use_mask[idx >> 4];
1099 int first_sh = 16 - (idx & 15);
1100 *use_mask_ptr++ &= 0xFFFFu << first_sh;
1101 excl_range -= first_sh;
1102 if (excl_range >= 16) {
1103 *use_mask_ptr++ = 0;
1104 *use_mask_ptr &= 0xFFFF >> (excl_range - 16);
1106 *use_mask_ptr &= 0xFFFF >> excl_range;
1109 /* find the 'aidx'th offset that is not excluded */
1110 aidx = bitstream_read(bc, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
1111 for (n = 0; n <= aidx; pulse_start++) {
1112 for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
1113 if (idx >= MAX_FRAMESIZE / 2) { // find from zero
1114 if (use_mask[0]) idx = 0x0F;
1115 else if (use_mask[1]) idx = 0x1F;
1116 else if (use_mask[2]) idx = 0x2F;
1117 else if (use_mask[3]) idx = 0x3F;
1118 else if (use_mask[4]) idx = 0x4F;
1120 idx -= av_log2_16bit(use_mask[idx >> 4]);
1122 if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
1123 use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
1129 fcb->x[fcb->n] = start_off;
1130 fcb->y[fcb->n] = bitstream_read_bit(bc) ? -1.0 : 1.0;
1133 /* set offset for next block, relative to start of that block */
1134 n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
1135 s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
1140 * Apply first set of pitch-adaptive window pulses.
1141 * @param s WMA Voice decoding context private data
1142 * @param bc bit I/O context
1143 * @param block_idx block index in frame [0, 1]
1144 * @param fcb storage location for fixed codebook pulse info
1146 static void aw_pulse_set1(WMAVoiceContext *s, BitstreamContext *bc,
1147 int block_idx, AMRFixed *fcb)
1149 int val = bitstream_read(bc, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
1152 if (s->aw_n_pulses[block_idx] > 0) {
1153 int n, v_mask, i_mask, sh, n_pulses;
1155 if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
1160 } else { // 4 pulses, 1:sign + 2:index each
1167 for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
1168 fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
1169 fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
1170 s->aw_first_pulse_off[block_idx];
1171 while (fcb->x[fcb->n] < 0)
1172 fcb->x[fcb->n] += fcb->pitch_lag;
1173 if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
1177 int num2 = (val & 0x1FF) >> 1, delta, idx;
1179 if (num2 < 1 * 79) { delta = 1; idx = num2 + 1; }
1180 else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
1181 else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
1182 else { delta = 7; idx = num2 + 1 - 3 * 75; }
1183 v = (val & 0x200) ? -1.0 : 1.0;
1185 fcb->no_repeat_mask |= 3 << fcb->n;
1186 fcb->x[fcb->n] = idx - delta;
1188 fcb->x[fcb->n + 1] = idx;
1189 fcb->y[fcb->n + 1] = (val & 1) ? -v : v;
1197 * Generate a random number from frame_cntr and block_idx, which will live
1198 * in the range [0, 1000 - block_size] (so it can be used as an index in a
1199 * table of size 1000 of which you want to read block_size entries).
1201 * @param frame_cntr current frame number
1202 * @param block_num current block index
1203 * @param block_size amount of entries we want to read from a table
1204 * that has 1000 entries
1205 * @return a (non-)random number in the [0, 1000 - block_size] range.
1207 static int pRNG(int frame_cntr, int block_num, int block_size)
1209 /* array to simplify the calculation of z:
1210 * y = (x % 9) * 5 + 6;
1211 * z = (49995 * x) / y;
1212 * Since y only has 9 values, we can remove the division by using a
1213 * LUT and using FASTDIV-style divisions. For each of the 9 values
1214 * of y, we can rewrite z as:
1215 * z = x * (49995 / y) + x * ((49995 % y) / y)
1216 * In this table, each col represents one possible value of y, the
1217 * first number is 49995 / y, and the second is the FASTDIV variant
1218 * of 49995 % y / y. */
1219 static const unsigned int div_tbl[9][2] = {
1220 { 8332, 3 * 715827883U }, // y = 6
1221 { 4545, 0 * 390451573U }, // y = 11
1222 { 3124, 11 * 268435456U }, // y = 16
1223 { 2380, 15 * 204522253U }, // y = 21
1224 { 1922, 23 * 165191050U }, // y = 26
1225 { 1612, 23 * 138547333U }, // y = 31
1226 { 1388, 27 * 119304648U }, // y = 36
1227 { 1219, 16 * 104755300U }, // y = 41
1228 { 1086, 39 * 93368855U } // y = 46
1230 unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
1231 if (x >= 0xFFFF) x -= 0xFFFF; // max value of x is 8*1877+0xFFFE=0x13AA6,
1232 // so this is effectively a modulo (%)
1233 y = x - 9 * MULH(477218589, x); // x % 9
1234 z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
1235 // z = x * 49995 / (y * 5 + 6)
1236 return z % (1000 - block_size);
1240 * Parse hardcoded signal for a single block.
1241 * @note see #synth_block().
1243 static void synth_block_hardcoded(WMAVoiceContext *s, BitstreamContext *bc,
1244 int block_idx, int size,
1245 const struct frame_type_desc *frame_desc,
1251 assert(size <= MAX_FRAMESIZE);
1253 /* Set the offset from which we start reading wmavoice_std_codebook */
1254 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1255 r_idx = pRNG(s->frame_cntr, block_idx, size);
1256 gain = s->silence_gain;
1257 } else /* FCB_TYPE_HARDCODED */ {
1258 r_idx = bitstream_read(bc, 8);
1259 gain = wmavoice_gain_universal[bitstream_read(bc, 6)];
1262 /* Clear gain prediction parameters */
1263 memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));
1265 /* Apply gain to hardcoded codebook and use that as excitation signal */
1266 for (n = 0; n < size; n++)
1267 excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
1271 * Parse FCB/ACB signal for a single block.
1272 * @note see #synth_block().
1274 static void synth_block_fcb_acb(WMAVoiceContext *s, BitstreamContext *bc,
1275 int block_idx, int size,
1276 int block_pitch_sh2,
1277 const struct frame_type_desc *frame_desc,
1280 static const float gain_coeff[6] = {
1281 0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
1283 float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
1284 int n, idx, gain_weight;
1287 assert(size <= MAX_FRAMESIZE / 2);
1288 memset(pulses, 0, sizeof(*pulses) * size);
1290 fcb.pitch_lag = block_pitch_sh2 >> 2;
1291 fcb.pitch_fac = 1.0;
1292 fcb.no_repeat_mask = 0;
1295 /* For the other frame types, this is where we apply the innovation
1296 * (fixed) codebook pulses of the speech signal. */
1297 if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1298 aw_pulse_set1(s, bc, block_idx, &fcb);
1299 if (aw_pulse_set2(s, bc, block_idx, &fcb)) {
1300 /* Conceal the block with silence and return.
1301 * Skip the correct amount of bits to read the next
1302 * block from the correct offset. */
1303 int r_idx = pRNG(s->frame_cntr, block_idx, size);
1305 for (n = 0; n < size; n++)
1307 wmavoice_std_codebook[r_idx + n] * s->silence_gain;
1308 bitstream_skip(bc, 7 + 1);
1311 } else /* FCB_TYPE_EXC_PULSES */ {
1312 int offset_nbits = 5 - frame_desc->log_n_blocks;
1314 fcb.no_repeat_mask = -1;
1315 /* similar to ff_decode_10_pulses_35bits(), but with single pulses
1316 * (instead of double) for a subset of pulses */
1317 for (n = 0; n < 5; n++) {
1321 sign = bitstream_read_bit(bc) ? 1.0 : -1.0;
1322 pos1 = bitstream_read(bc, offset_nbits);
1323 fcb.x[fcb.n] = n + 5 * pos1;
1324 fcb.y[fcb.n++] = sign;
1325 if (n < frame_desc->dbl_pulses) {
1326 pos2 = bitstream_read(bc, offset_nbits);
1327 fcb.x[fcb.n] = n + 5 * pos2;
1328 fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
1332 ff_set_fixed_vector(pulses, &fcb, 1.0, size);
1334 /* Calculate gain for adaptive & fixed codebook signal.
1335 * see ff_amr_set_fixed_gain(). */
1336 idx = bitstream_read(bc, 7);
1337 fcb_gain = expf(avpriv_scalarproduct_float_c(s->gain_pred_err,
1339 5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
1340 acb_gain = wmavoice_gain_codebook_acb[idx];
1341 pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
1342 -2.9957322736 /* log(0.05) */,
1343 1.6094379124 /* log(5.0) */);
1345 gain_weight = 8 >> frame_desc->log_n_blocks;
1346 memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
1347 sizeof(*s->gain_pred_err) * (6 - gain_weight));
1348 for (n = 0; n < gain_weight; n++)
1349 s->gain_pred_err[n] = pred_err;
1351 /* Calculation of adaptive codebook */
1352 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1354 for (n = 0; n < size; n += len) {
1356 int abs_idx = block_idx * size + n;
1357 int pitch_sh16 = (s->last_pitch_val << 16) +
1358 s->pitch_diff_sh16 * abs_idx;
1359 int pitch = (pitch_sh16 + 0x6FFF) >> 16;
1360 int idx_sh16 = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
1361 idx = idx_sh16 >> 16;
1362 if (s->pitch_diff_sh16) {
1363 if (s->pitch_diff_sh16 > 0) {
1364 next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
1366 next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
1367 len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
1372 ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
1373 wmavoice_ipol1_coeffs, 17,
1376 } else /* ACB_TYPE_HAMMING */ {
1377 int block_pitch = block_pitch_sh2 >> 2;
1378 idx = block_pitch_sh2 & 3;
1380 ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
1381 wmavoice_ipol2_coeffs, 4,
1384 av_memcpy_backptr((uint8_t *) excitation, sizeof(float) * block_pitch,
1385 sizeof(float) * size);
1388 /* Interpolate ACB/FCB and use as excitation signal */
1389 ff_weighted_vector_sumf(excitation, excitation, pulses,
1390 acb_gain, fcb_gain, size);
1394 * Parse data in a single block.
1395 * @note we assume enough bits are available, caller should check.
1397 * @param s WMA Voice decoding context private data
1398 * @param bc bit I/O context
1399 * @param block_idx index of the to-be-read block
1400 * @param size amount of samples to be read in this block
1401 * @param block_pitch_sh2 pitch for this block << 2
1402 * @param lsps LSPs for (the end of) this frame
1403 * @param prev_lsps LSPs for the last frame
1404 * @param frame_desc frame type descriptor
1405 * @param excitation target memory for the ACB+FCB interpolated signal
1406 * @param synth target memory for the speech synthesis filter output
1407 * @return 0 on success, <0 on error.
1409 static void synth_block(WMAVoiceContext *s, BitstreamContext *bc,
1410 int block_idx, int size,
1411 int block_pitch_sh2,
1412 const double *lsps, const double *prev_lsps,
1413 const struct frame_type_desc *frame_desc,
1414 float *excitation, float *synth)
1416 double i_lsps[MAX_LSPS];
1417 float lpcs[MAX_LSPS];
1421 if (frame_desc->acb_type == ACB_TYPE_NONE)
1422 synth_block_hardcoded(s, bc, block_idx, size, frame_desc, excitation);
1424 synth_block_fcb_acb(s, bc, block_idx, size, block_pitch_sh2,
1425 frame_desc, excitation);
1427 /* convert interpolated LSPs to LPCs */
1428 fac = (block_idx + 0.5) / frame_desc->n_blocks;
1429 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1430 i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
1431 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1433 /* Speech synthesis */
1434 ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
1438 * Synthesize output samples for a single frame.
1439 * @note we assume enough bits are available, caller should check.
1441 * @param ctx WMA Voice decoder context
1442 * @param bc bit I/O context (s->bc or one for cross-packet superframes)
1443 * @param frame_idx Frame number within superframe [0-2]
1444 * @param samples pointer to output sample buffer, has space for at least 160
1446 * @param lsps LSP array
1447 * @param prev_lsps array of previous frame's LSPs
1448 * @param excitation target buffer for excitation signal
1449 * @param synth target buffer for synthesized speech data
1450 * @return 0 on success, <0 on error.
1452 static int synth_frame(AVCodecContext *ctx, BitstreamContext *bc,
1453 int frame_idx, float *samples,
1454 const double *lsps, const double *prev_lsps,
1455 float *excitation, float *synth)
1457 WMAVoiceContext *s = ctx->priv_data;
1458 int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
1459 int pitch[MAX_BLOCKS], last_block_pitch;
1461 /* Parse frame type ("frame header"), see frame_descs */
1462 int bd_idx = s->vbm_tree[bitstream_read_vlc(bc, frame_type_vlc.table, 6, 3)], block_nsamples;
1465 av_log(ctx, AV_LOG_ERROR,
1466 "Invalid frame type VLC code, skipping\n");
1467 return AVERROR_INVALIDDATA;
1470 block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;
1472 /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
1473 if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
1474 /* Pitch is provided per frame, which is interpreted as the pitch of
1475 * the last sample of the last block of this frame. We can interpolate
1476 * the pitch of other blocks (and even pitch-per-sample) by gradually
1477 * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
1478 n_blocks_x2 = frame_descs[bd_idx].n_blocks << 1;
1479 log_n_blocks_x2 = frame_descs[bd_idx].log_n_blocks + 1;
1480 cur_pitch_val = s->min_pitch_val + bitstream_read(bc, s->pitch_nbits);
1481 cur_pitch_val = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
1482 if (s->last_acb_type == ACB_TYPE_NONE ||
1483 20 * abs(cur_pitch_val - s->last_pitch_val) >
1484 (cur_pitch_val + s->last_pitch_val))
1485 s->last_pitch_val = cur_pitch_val;
1487 /* pitch per block */
1488 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1489 int fac = n * 2 + 1;
1491 pitch[n] = (MUL16(fac, cur_pitch_val) +
1492 MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
1493 frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
1496 /* "pitch-diff-per-sample" for calculation of pitch per sample */
1497 s->pitch_diff_sh16 =
1498 ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
1501 /* Global gain (if silence) and pitch-adaptive window coordinates */
1502 switch (frame_descs[bd_idx].fcb_type) {
1503 case FCB_TYPE_SILENCE:
1504 s->silence_gain = wmavoice_gain_silence[bitstream_read(bc, 8)];
1506 case FCB_TYPE_AW_PULSES:
1507 aw_parse_coords(s, bc, pitch);
1511 for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
1514 /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
1515 switch (frame_descs[bd_idx].acb_type) {
1516 case ACB_TYPE_HAMMING: {
1517 /* Pitch is given per block. Per-block pitches are encoded as an
1518 * absolute value for the first block, and then delta values
1519 * relative to this value) for all subsequent blocks. The scale of
1520 * this pitch value is semi-logarithmic compared to its use in the
1521 * decoder, so we convert it to normal scale also. */
1523 t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
1524 t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
1525 t3 = s->block_conv_table[3] - s->block_conv_table[2] + 1;
1528 block_pitch = bitstream_read(bc, s->block_pitch_nbits);
1530 block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
1531 bitstream_read(bc, s->block_delta_pitch_nbits);
1532 /* Convert last_ so that any next delta is within _range */
1533 last_block_pitch = av_clip(block_pitch,
1534 s->block_delta_pitch_hrange,
1535 s->block_pitch_range -
1536 s->block_delta_pitch_hrange);
1538 /* Convert semi-log-style scale back to normal scale */
1539 if (block_pitch < t1) {
1540 bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
1543 if (block_pitch < t2) {
1545 (s->block_conv_table[1] << 2) + (block_pitch << 1);
1548 if (block_pitch < t3) {
1550 (s->block_conv_table[2] + block_pitch) << 2;
1552 bl_pitch_sh2 = s->block_conv_table[3] << 2;
1555 pitch[n] = bl_pitch_sh2 >> 2;
1559 case ACB_TYPE_ASYMMETRIC: {
1560 bl_pitch_sh2 = pitch[n] << 2;
1564 default: // ACB_TYPE_NONE has no pitch
1569 synth_block(s, bc, n, block_nsamples, bl_pitch_sh2,
1570 lsps, prev_lsps, &frame_descs[bd_idx],
1571 &excitation[n * block_nsamples],
1572 &synth[n * block_nsamples]);
1575 /* Averaging projection filter, if applicable. Else, just copy samples
1576 * from synthesis buffer */
1578 double i_lsps[MAX_LSPS];
1579 float lpcs[MAX_LSPS];
1581 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1582 i_lsps[n] = cos(0.5 * (prev_lsps[n] + lsps[n]));
1583 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1584 postfilter(s, synth, samples, 80, lpcs,
1585 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx],
1586 frame_descs[bd_idx].fcb_type, pitch[0]);
1588 for (n = 0; n < s->lsps; n++) // LSF -> LSP
1589 i_lsps[n] = cos(lsps[n]);
1590 ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);
1591 postfilter(s, &synth[80], &samples[80], 80, lpcs,
1592 &s->zero_exc_pf[s->history_nsamples + MAX_FRAMESIZE * frame_idx + 80],
1593 frame_descs[bd_idx].fcb_type, pitch[0]);
1595 memcpy(samples, synth, 160 * sizeof(synth[0]));
1597 /* Cache values for next frame */
1599 if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
1600 s->last_acb_type = frame_descs[bd_idx].acb_type;
1601 switch (frame_descs[bd_idx].acb_type) {
1603 s->last_pitch_val = 0;
1605 case ACB_TYPE_ASYMMETRIC:
1606 s->last_pitch_val = cur_pitch_val;
1608 case ACB_TYPE_HAMMING:
1609 s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
1617 * Ensure minimum value for first item, maximum value for last value,
1618 * proper spacing between each value and proper ordering.
1620 * @param lsps array of LSPs
1621 * @param num size of LSP array
1623 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
1624 * useful to put in a generic location later on. Parts are also
1625 * present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
1626 * which is in float.
1628 static void stabilize_lsps(double *lsps, int num)
1632 /* set minimum value for first, maximum value for last and minimum
1633 * spacing between LSF values.
1634 * Very similar to ff_set_min_dist_lsf(), but in double. */
1635 lsps[0] = FFMAX(lsps[0], 0.0015 * M_PI);
1636 for (n = 1; n < num; n++)
1637 lsps[n] = FFMAX(lsps[n], lsps[n - 1] + 0.0125 * M_PI);
1638 lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);
1640 /* reorder (looks like one-time / non-recursed bubblesort).
1641 * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
1642 for (n = 1; n < num; n++) {
1643 if (lsps[n] < lsps[n - 1]) {
1644 for (m = 1; m < num; m++) {
1645 double tmp = lsps[m];
1646 for (l = m - 1; l >= 0; l--) {
1647 if (lsps[l] <= tmp) break;
1648 lsps[l + 1] = lsps[l];
1658 * Test if there's enough bits to read 1 superframe.
1660 * @param orig_bc bit I/O context used for reading. This function
1661 * does not modify the state of the bitreader; it
1662 * only uses it to copy the current stream position
1663 * @param s WMA Voice decoding context private data
1664 * @return < 0 on error, 1 on not enough bits or 0 if OK.
1666 static int check_bits_for_superframe(BitstreamContext *orig_bc,
1669 BitstreamContext s_bc, *bc = &s_bc;
1670 int n, need_bits, bd_idx;
1671 const struct frame_type_desc *frame_desc;
1673 /* initialize a copy */
1676 /* superframe header */
1677 if (bitstream_bits_left(bc) < 14)
1679 if (!bitstream_read_bit(bc))
1680 return AVERROR(ENOSYS); // WMAPro-in-WMAVoice superframe
1681 if (bitstream_read_bit(bc)) bitstream_skip(bc, 12); // number of samples in superframe
1682 if (s->has_residual_lsps) { // residual LSPs (for all frames)
1683 if (bitstream_bits_left(bc) < s->sframe_lsp_bitsize)
1685 bitstream_skip(bc, s->sframe_lsp_bitsize);
1689 for (n = 0; n < MAX_FRAMES; n++) {
1690 int aw_idx_is_ext = 0;
1692 if (!s->has_residual_lsps) { // independent LSPs (per-frame)
1693 if (bitstream_bits_left(bc) < s->frame_lsp_bitsize)
1695 bitstream_skip(bc, s->frame_lsp_bitsize);
1697 bd_idx = s->vbm_tree[bitstream_read_vlc(bc, frame_type_vlc.table, 6, 3)];
1699 return AVERROR_INVALIDDATA; // invalid frame type VLC code
1700 frame_desc = &frame_descs[bd_idx];
1701 if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
1702 if (bitstream_bits_left(bc) < s->pitch_nbits)
1704 bitstream_skip(bc, s->pitch_nbits);
1706 if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
1707 bitstream_skip(bc, 8);
1708 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1709 int tmp = bitstream_read(bc, 6);
1711 bitstream_skip(bc, 2);
1717 if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
1718 need_bits = s->block_pitch_nbits +
1719 (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
1720 } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
1721 need_bits = 2 * !aw_idx_is_ext;
1724 need_bits += frame_desc->frame_size;
1725 if (bitstream_bits_left(bc) < need_bits)
1727 bitstream_skip(bc, need_bits);
1734 * Synthesize output samples for a single superframe. If we have any data
1735 * cached in s->sframe_cache, that will be used instead of whatever is loaded
1738 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
1739 * to give a total of 480 samples per frame. See #synth_frame() for frame
1740 * parsing. In addition to 3 frames, superframes can also contain the LSPs
1741 * (if these are globally specified for all frames (residually); they can
1742 * also be specified individually per-frame. See the s->has_residual_lsps
1743 * option), and can specify the number of samples encoded in this superframe
1744 * (if less than 480), usually used to prevent blanks at track boundaries.
1746 * @param ctx WMA Voice decoder context
1747 * @return 0 on success, <0 on error or 1 if there was not enough data to
1748 * fully parse the superframe
1750 static int synth_superframe(AVCodecContext *ctx, AVFrame *frame,
1753 WMAVoiceContext *s = ctx->priv_data;
1754 BitstreamContext *bc = &s->bc, s_bc;
1755 int n, res, n_samples = 480;
1756 double lsps[MAX_FRAMES][MAX_LSPS];
1757 const double *mean_lsf = s->lsps == 16 ?
1758 wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
1759 float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
1760 float synth[MAX_LSPS + MAX_SFRAMESIZE];
1763 memcpy(synth, s->synth_history,
1764 s->lsps * sizeof(*synth));
1765 memcpy(excitation, s->excitation_history,
1766 s->history_nsamples * sizeof(*excitation));
1768 if (s->sframe_cache_size > 0) {
1770 bitstream_init(bc, s->sframe_cache, s->sframe_cache_size);
1771 s->sframe_cache_size = 0;
1774 if ((res = check_bits_for_superframe(bc, s)) == 1) {
1780 /* First bit is speech/music bit, it differentiates between WMAVoice
1781 * speech samples (the actual codec) and WMAVoice music samples, which
1782 * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
1784 if (!bitstream_read_bit(bc)) {
1785 avpriv_request_sample(ctx, "WMAPro-in-WMAVoice");
1786 return AVERROR_PATCHWELCOME;
1789 /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
1790 if (bitstream_read_bit(bc)) {
1791 if ((n_samples = bitstream_read(bc, 12)) > 480) {
1792 av_log(ctx, AV_LOG_ERROR,
1793 "Superframe encodes >480 samples (%d), not allowed\n",
1795 return AVERROR_INVALIDDATA;
1798 /* Parse LSPs, if global for the superframe (can also be per-frame). */
1799 if (s->has_residual_lsps) {
1800 double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];
1802 for (n = 0; n < s->lsps; n++)
1803 prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];
1805 if (s->lsps == 10) {
1806 dequant_lsp10r(bc, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1807 } else /* s->lsps == 16 */
1808 dequant_lsp16r(bc, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
1810 for (n = 0; n < s->lsps; n++) {
1811 lsps[0][n] = mean_lsf[n] + (a1[n] - a2[n * 2]);
1812 lsps[1][n] = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
1813 lsps[2][n] += mean_lsf[n];
1815 for (n = 0; n < 3; n++)
1816 stabilize_lsps(lsps[n], s->lsps);
1819 /* get output buffer */
1820 frame->nb_samples = 480;
1821 if ((res = ff_get_buffer(ctx, frame, 0)) < 0) {
1822 av_log(ctx, AV_LOG_ERROR, "get_buffer() failed\n");
1825 frame->nb_samples = n_samples;
1826 samples = (float *)frame->data[0];
1828 /* Parse frames, optionally preceded by per-frame (independent) LSPs. */
1829 for (n = 0; n < 3; n++) {
1830 if (!s->has_residual_lsps) {
1833 if (s->lsps == 10) {
1834 dequant_lsp10i(bc, lsps[n]);
1835 } else /* s->lsps == 16 */
1836 dequant_lsp16i(bc, lsps[n]);
1838 for (m = 0; m < s->lsps; m++)
1839 lsps[n][m] += mean_lsf[m];
1840 stabilize_lsps(lsps[n], s->lsps);
1843 if ((res = synth_frame(ctx, bc, n,
1844 &samples[n * MAX_FRAMESIZE],
1845 lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
1846 &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
1847 &synth[s->lsps + n * MAX_FRAMESIZE]))) {
1853 /* Statistics? FIXME - we don't check for length, a slight overrun
1854 * will be caught by internal buffer padding, and anything else
1855 * will be skipped, not read. */
1856 if (bitstream_read_bit(bc)) {
1857 res = bitstream_read(bc, 4);
1858 bitstream_skip(bc, 10 * (res + 1));
1863 /* Update history */
1864 memcpy(s->prev_lsps, lsps[2],
1865 s->lsps * sizeof(*s->prev_lsps));
1866 memcpy(s->synth_history, &synth[MAX_SFRAMESIZE],
1867 s->lsps * sizeof(*synth));
1868 memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
1869 s->history_nsamples * sizeof(*excitation));
1871 memmove(s->zero_exc_pf, &s->zero_exc_pf[MAX_SFRAMESIZE],
1872 s->history_nsamples * sizeof(*s->zero_exc_pf));
1878 * Parse the packet header at the start of each packet (input data to this
1881 * @param s WMA Voice decoding context private data
1882 * @return 1 if not enough bits were available, or 0 on success.
1884 static int parse_packet_header(WMAVoiceContext *s)
1886 BitstreamContext *bc = &s->bc;
1889 if (bitstream_bits_left(bc) < 11)
1891 bitstream_skip(bc, 4); // packet sequence number
1892 s->has_residual_lsps = bitstream_read_bit(bc);
1894 res = bitstream_read(bc, 6); // number of superframes per packet
1895 // (minus first one if there is spillover)
1896 if (bitstream_bits_left(bc) < 6 * (res == 0x3F) + s->spillover_bitsize)
1898 } while (res == 0x3F);
1899 s->spillover_nbits = bitstream_read(bc, s->spillover_bitsize);
1905 * Copy (unaligned) bits from bc/data/size to pb.
1907 * @param pb target buffer to copy bits into
1908 * @param data source buffer to copy bits from
1909 * @param size size of the source data, in bytes
1910 * @param bc bit I/O context specifying the current position in the source.
1911 * data. This function might use this to align the bit position to
1912 * a whole-byte boundary before calling #avpriv_copy_bits() on aligned
1914 * @param nbits the amount of bits to copy from source to target
1916 * @note after calling this function, the current position in the input bit
1917 * I/O context is undefined.
1919 static void copy_bits(PutBitContext *pb,
1920 const uint8_t *data, int size,
1921 BitstreamContext *bc, int nbits)
1923 int rmn_bytes, rmn_bits;
1925 rmn_bits = rmn_bytes = bitstream_bits_left(bc);
1926 if (rmn_bits < nbits)
1928 if (nbits > pb->size_in_bits - put_bits_count(pb))
1930 rmn_bits &= 7; rmn_bytes >>= 3;
1931 if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
1932 put_bits(pb, rmn_bits, bitstream_read(bc, rmn_bits));
1933 avpriv_copy_bits(pb, data + size - rmn_bytes,
1934 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
1938 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
1939 * and we expect that the demuxer / application provides it to us as such
1940 * (else you'll probably get garbage as output). Every packet has a size of
1941 * ctx->block_align bytes, starts with a packet header (see
1942 * #parse_packet_header()), and then a series of superframes. Superframe
1943 * boundaries may exceed packets, i.e. superframes can split data over
1944 * multiple (two) packets.
1946 * For more information about frames, see #synth_superframe().
1948 static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
1949 int *got_frame_ptr, AVPacket *avpkt)
1951 WMAVoiceContext *s = ctx->priv_data;
1952 BitstreamContext *bc = &s->bc;
1955 /* Packets are sometimes a multiple of ctx->block_align, with a packet
1956 * header at each ctx->block_align bytes. However, Libav's ASF demuxer
1957 * feeds us ASF packets, which may concatenate multiple "codec" packets
1958 * in a single "muxer" packet, so we artificially emulate that by
1959 * capping the packet size at ctx->block_align. */
1960 for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
1965 bitstream_init8(&s->bc, avpkt->data, size);
1967 /* size == ctx->block_align is used to indicate whether we are dealing with
1968 * a new packet or a packet of which we already read the packet header
1970 if (size == ctx->block_align) { // new packet header
1971 if ((res = parse_packet_header(s)) < 0)
1974 /* If the packet header specifies a s->spillover_nbits, then we want
1975 * to push out all data of the previous packet (+ spillover) before
1976 * continuing to parse new superframes in the current packet. */
1977 if (s->spillover_nbits > 0) {
1978 if (s->sframe_cache_size > 0) {
1979 int cnt = bitstream_tell(bc);
1980 copy_bits(&s->pb, avpkt->data, size, bc, s->spillover_nbits);
1981 flush_put_bits(&s->pb);
1982 s->sframe_cache_size += s->spillover_nbits;
1983 if ((res = synth_superframe(ctx, data, got_frame_ptr)) == 0 &&
1985 cnt += s->spillover_nbits;
1986 s->skip_bits_next = cnt & 7;
1989 bitstream_skip (bc, s->spillover_nbits - cnt +
1990 bitstream_tell(bc)); // resync
1992 bitstream_skip(bc, s->spillover_nbits); // resync
1994 } else if (s->skip_bits_next)
1995 bitstream_skip(bc, s->skip_bits_next);
1997 /* Try parsing superframes in current packet */
1998 s->sframe_cache_size = 0;
1999 s->skip_bits_next = 0;
2000 pos = bitstream_bits_left(bc);
2001 if ((res = synth_superframe(ctx, data, got_frame_ptr)) < 0) {
2003 } else if (*got_frame_ptr) {
2004 int cnt = bitstream_tell(bc);
2005 s->skip_bits_next = cnt & 7;
2007 } else if ((s->sframe_cache_size = pos) > 0) {
2008 /* rewind bit reader to start of last (incomplete) superframe... */
2009 bitstream_init8(bc, avpkt->data, size);
2010 bitstream_skip(bc, (size << 3) - pos);
2011 assert(bitstream_bits_left(bc) == pos);
2013 /* ...and cache it for spillover in next packet */
2014 init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
2015 copy_bits(&s->pb, avpkt->data, size, bc, s->sframe_cache_size);
2016 // FIXME bad - just copy bytes as whole and add use the
2017 // skip_bits_next field
2023 static av_cold int wmavoice_decode_end(AVCodecContext *ctx)
2025 WMAVoiceContext *s = ctx->priv_data;
2028 ff_rdft_end(&s->rdft);
2029 ff_rdft_end(&s->irdft);
2030 ff_dct_end(&s->dct);
2031 ff_dct_end(&s->dst);
2037 static av_cold void wmavoice_flush(AVCodecContext *ctx)
2039 WMAVoiceContext *s = ctx->priv_data;
2042 s->postfilter_agc = 0;
2043 s->sframe_cache_size = 0;
2044 s->skip_bits_next = 0;
2045 for (n = 0; n < s->lsps; n++)
2046 s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
2047 memset(s->excitation_history, 0,
2048 sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
2049 memset(s->synth_history, 0,
2050 sizeof(*s->synth_history) * MAX_LSPS);
2051 memset(s->gain_pred_err, 0,
2052 sizeof(s->gain_pred_err));
2055 memset(&s->synth_filter_out_buf[MAX_LSPS_ALIGN16 - s->lsps], 0,
2056 sizeof(*s->synth_filter_out_buf) * s->lsps);
2057 memset(s->dcf_mem, 0,
2058 sizeof(*s->dcf_mem) * 2);
2059 memset(s->zero_exc_pf, 0,
2060 sizeof(*s->zero_exc_pf) * s->history_nsamples);
2061 memset(s->denoise_filter_cache, 0, sizeof(s->denoise_filter_cache));
2065 AVCodec ff_wmavoice_decoder = {
2067 .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
2068 .type = AVMEDIA_TYPE_AUDIO,
2069 .id = AV_CODEC_ID_WMAVOICE,
2070 .priv_data_size = sizeof(WMAVoiceContext),
2071 .init = wmavoice_decode_init,
2072 .init_static_data = wmavoice_init_static_data,
2073 .close = wmavoice_decode_end,
2074 .decode = wmavoice_decode_packet,
2075 .capabilities = AV_CODEC_CAP_SUBFRAMES | AV_CODEC_CAP_DR1,
2076 .flush = wmavoice_flush,