aboutsummaryrefslogtreecommitdiffstats
path: root/libavcodec/wmavoice.c
blob: 259755343d950923bb7a60c45df1df07c10f2ff2 (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
/*
 * Windows Media Audio Voice decoder.
 * Copyright (c) 2009 Ronald S. Bultje
 *
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * FFmpeg is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with FFmpeg; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */

/**
 * @file libavcodec/wmavoice.c
 * @brief Windows Media Audio Voice compatible decoder
 * @author Ronald S. Bultje <rsbultje@gmail.com>
 */

#include <math.h>
#include "avcodec.h"
#include "get_bits.h"
#include "put_bits.h"
#include "wmavoice_data.h"
#include "celp_math.h"
#include "celp_filters.h"
#include "acelp_vectors.h"
#include "acelp_filters.h"
#include "lsp.h"
#include "libavutil/lzo.h"

#define MAX_BLOCKS           8   ///< maximum number of blocks per frame
#define MAX_LSPS             16  ///< maximum filter order
#define MAX_FRAMES           3   ///< maximum number of frames per superframe
#define MAX_FRAMESIZE        160 ///< maximum number of samples per frame
#define MAX_SIGNAL_HISTORY   416 ///< maximum excitation signal history
#define MAX_SFRAMESIZE       (MAX_FRAMESIZE * MAX_FRAMES)
                                 ///< maximum number of samples per superframe
#define SFRAME_CACHE_MAXSIZE 256 ///< maximum cache size for frame data that
                                 ///< was split over two packets
#define VLC_NBITS            6   ///< number of bits to read per VLC iteration

/**
 * Frame type VLC coding.
 */
static VLC frame_type_vlc;

/**
 * Adaptive codebook types.
 */
enum {
    ACB_TYPE_NONE       = 0, ///< no adaptive codebook (only hardcoded fixed)
    ACB_TYPE_ASYMMETRIC = 1, ///< adaptive codebook with per-frame pitch, which
                             ///< we interpolate to get a per-sample pitch.
                             ///< Signal is generated using an asymmetric sinc
                             ///< window function
                             ///< @note see #wmavoice_ipol1_coeffs
    ACB_TYPE_HAMMING    = 2  ///< Per-block pitch with signal generation using
                             ///< a Hamming sinc window function
                             ///< @note see #wmavoice_ipol2_coeffs
};

/**
 * Fixed codebook types.
 */
enum {
    FCB_TYPE_SILENCE    = 0, ///< comfort noise during silence
                             ///< generated from a hardcoded (fixed) codebook
                             ///< with per-frame (low) gain values
    FCB_TYPE_HARDCODED  = 1, ///< hardcoded (fixed) codebook with per-block
                             ///< gain values
    FCB_TYPE_AW_PULSES  = 2, ///< Pitch-adaptive window (AW) pulse signals,
                             ///< used in particular for low-bitrate streams
    FCB_TYPE_EXC_PULSES = 3, ///< Innovation (fixed) codebook pulse sets in
                             ///< combinations of either single pulses or
                             ///< pulse pairs
};

/**
 * Description of frame types.
 */
static const struct frame_type_desc {
    uint8_t n_blocks;     ///< amount of blocks per frame (each block
                          ///< (contains 160/#n_blocks samples)
    uint8_t log_n_blocks; ///< log2(#n_blocks)
    uint8_t acb_type;     ///< Adaptive codebook type (ACB_TYPE_*)
    uint8_t fcb_type;     ///< Fixed codebook type (FCB_TYPE_*)
    uint8_t dbl_pulses;   ///< how many pulse vectors have pulse pairs
                          ///< (rather than just one single pulse)
                          ///< only if #fcb_type == #FCB_TYPE_EXC_PULSES
    uint16_t frame_size;  ///< the amount of bits that make up the block
                          ///< data (per frame)
} frame_descs[17] = {
    { 1, 0, ACB_TYPE_NONE,       FCB_TYPE_SILENCE,    0,   0 },
    { 2, 1, ACB_TYPE_NONE,       FCB_TYPE_HARDCODED,  0,  28 },
    { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_AW_PULSES,  0,  46 },
    { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2,  80 },
    { 2, 1, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 104 },
    { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 0, 108 },
    { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 2, 132 },
    { 4, 2, ACB_TYPE_ASYMMETRIC, FCB_TYPE_EXC_PULSES, 5, 168 },
    { 2, 1, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 0,  64 },
    { 2, 1, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 2,  80 },
    { 2, 1, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 5, 104 },
    { 4, 2, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 0, 108 },
    { 4, 2, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 2, 132 },
    { 4, 2, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 5, 168 },
    { 8, 3, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 0, 176 },
    { 8, 3, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 2, 208 },
    { 8, 3, ACB_TYPE_HAMMING,    FCB_TYPE_EXC_PULSES, 5, 256 }
};

/**
 * WMA Voice decoding context.
 */
typedef struct {
    /**
     * @defgroup struct_global Global values
     * Global values, specified in the stream header / extradata or used
     * all over.
     * @{
     */
    GetBitContext gb;             ///< packet bitreader. During decoder init,
                                  ///< it contains the extradata from the
                                  ///< demuxer. During decoding, it contains
                                  ///< packet data.
    int8_t vbm_tree[25];          ///< converts VLC codes to frame type

    int spillover_bitsize;        ///< number of bits used to specify
                                  ///< #spillover_nbits in the packet header
                                  ///< = ceil(log2(ctx->block_align << 3))
    int history_nsamples;         ///< number of samples in history for signal
                                  ///< prediction (through ACB)

    int do_apf;                   ///< whether to apply the averaged
                                  ///< projection filter (APF)

    int lsps;                     ///< number of LSPs per frame [10 or 16]
    int lsp_q_mode;               ///< defines quantizer defaults [0, 1]
    int lsp_def_mode;             ///< defines different sets of LSP defaults
                                  ///< [0, 1]
    int frame_lsp_bitsize;        ///< size (in bits) of LSPs, when encoded
                                  ///< per-frame (independent coding)
    int sframe_lsp_bitsize;       ///< size (in bits) of LSPs, when encoded
                                  ///< per superframe (residual coding)

    int min_pitch_val;            ///< base value for pitch parsing code
    int max_pitch_val;            ///< max value + 1 for pitch parsing
    int pitch_nbits;              ///< number of bits used to specify the
                                  ///< pitch value in the frame header
    int block_pitch_nbits;        ///< number of bits used to specify the
                                  ///< first block's pitch value
    int block_pitch_range;        ///< range of the block pitch
    int block_delta_pitch_nbits;  ///< number of bits used to specify the
                                  ///< delta pitch between this and the last
                                  ///< block's pitch value, used in all but
                                  ///< first block
    int block_delta_pitch_hrange; ///< 1/2 range of the delta (full range is
                                  ///< from -this to +this-1)
    uint16_t block_conv_table[4]; ///< boundaries for block pitch unit/scale
                                  ///< conversion

    /**
     * @}
     * @defgroup struct_packet Packet values
     * Packet values, specified in the packet header or related to a packet.
     * A packet is considered to be a single unit of data provided to this
     * decoder by the demuxer.
     * @{
     */
    int spillover_nbits;          ///< number of bits of the previous packet's
                                  ///< last superframe preceeding this
                                  ///< packet's first full superframe (useful
                                  ///< for re-synchronization also)
    int has_residual_lsps;        ///< if set, superframes contain one set of
                                  ///< LSPs that cover all frames, encoded as
                                  ///< independent and residual LSPs; if not
                                  ///< set, each frame contains its own, fully
                                  ///< independent, LSPs
    int skip_bits_next;           ///< number of bits to skip at the next call
                                  ///< to #wmavoice_decode_packet() (since
                                  ///< they're part of the previous superframe)

    uint8_t sframe_cache[SFRAME_CACHE_MAXSIZE + FF_INPUT_BUFFER_PADDING_SIZE];
                                  ///< cache for superframe data split over
                                  ///< multiple packets
    int sframe_cache_size;        ///< set to >0 if we have data from an
                                  ///< (incomplete) superframe from a previous
                                  ///< packet that spilled over in the current
                                  ///< packet; specifies the amount of bits in
                                  ///< #sframe_cache
    PutBitContext pb;             ///< bitstream writer for #sframe_cache

    /**
     * @}
     * @defgroup struct_frame Frame and superframe values
     * Superframe and frame data - these can change from frame to frame,
     * although some of them do in that case serve as a cache / history for
     * the next frame or superframe.
     * @{
     */
    double prev_lsps[MAX_LSPS];   ///< LSPs of the last frame of the previous
                                  ///< superframe
    int last_pitch_val;           ///< pitch value of the previous frame
    int last_acb_type;            ///< frame type [0-2] of the previous frame
    int pitch_diff_sh16;          ///< ((cur_pitch_val - #last_pitch_val)
                                  ///< << 16) / #MAX_FRAMESIZE
    float silence_gain;           ///< set for use in blocks if #ACB_TYPE_NONE

    int aw_idx_is_ext;            ///< whether the AW index was encoded in
                                  ///< 8 bits (instead of 6)
    int aw_pulse_range;           ///< the range over which #aw_pulse_set1()
                                  ///< can apply the pulse, relative to the
                                  ///< value in aw_first_pulse_off. The exact
                                  ///< position of the first AW-pulse is within
                                  ///< [pulse_off, pulse_off + this], and
                                  ///< depends on bitstream values; [16 or 24]
    int aw_n_pulses[2];           ///< number of AW-pulses in each block; note
                                  ///< that this number can be negative (in
                                  ///< which case it basically means "zero")
    int aw_first_pulse_off[2];    ///< index of first sample to which to
                                  ///< apply AW-pulses, or -0xff if unset
    int aw_next_pulse_off_cache;  ///< the position (relative to start of the
                                  ///< second block) at which pulses should
                                  ///< start to be positioned, serves as a
                                  ///< cache for pitch-adaptive window pulses
                                  ///< between blocks

    int frame_cntr;               ///< current frame index [0 - 0xFFFE]; is
                                  ///< only used for comfort noise in #pRNG()
    float gain_pred_err[6];       ///< cache for gain prediction
    float excitation_history[MAX_SIGNAL_HISTORY];
                                  ///< cache of the signal of previous
                                  ///< superframes, used as a history for
                                  ///< signal generation
    float synth_history[MAX_LSPS]; ///< see #excitation_history
    /**
     * @}
     */
} WMAVoiceContext;

/**
 * Sets up the variable bit mode (VBM) tree from container extradata.
 * @param gb bit I/O context.
 *           The bit context (s->gb) should be loaded with byte 23-46 of the
 *           container extradata (i.e. the ones containing the VBM tree).
 * @param vbm_tree pointer to array to which the decoded VBM tree will be
 *                 written.
 * @return 0 on success, <0 on error.
 */
static av_cold int decode_vbmtree(GetBitContext *gb, int8_t vbm_tree[25])
{
    static const uint8_t bits[] = {
         2,  2,  2,  4,  4,  4,
         6,  6,  6,  8,  8,  8,
        10, 10, 10, 12, 12, 12,
        14, 14, 14, 14
    };
    static const uint16_t codes[] = {
          0x0000, 0x0001, 0x0002,        //              00/01/10
          0x000c, 0x000d, 0x000e,        //           11+00/01/10
          0x003c, 0x003d, 0x003e,        //         1111+00/01/10
          0x00fc, 0x00fd, 0x00fe,        //       111111+00/01/10
          0x03fc, 0x03fd, 0x03fe,        //     11111111+00/01/10
          0x0ffc, 0x0ffd, 0x0ffe,        //   1111111111+00/01/10
          0x3ffc, 0x3ffd, 0x3ffe, 0x3fff // 111111111111+xx
    };
    int cntr[8], n, res;

    memset(vbm_tree, 0xff, sizeof(vbm_tree));
    memset(cntr,     0,    sizeof(cntr));
    for (n = 0; n < 17; n++) {
        res = get_bits(gb, 3);
        if (cntr[res] > 3) // should be >= 3 + (res == 7))
            return -1;
        vbm_tree[res * 3 + cntr[res]++] = n;
    }
    INIT_VLC_STATIC(&frame_type_vlc, VLC_NBITS, sizeof(bits),
                    bits, 1, 1, codes, 2, 2, 132);
    return 0;
}

/**
 * Set up decoder with parameters from demuxer (extradata etc.).
 */
static av_cold int wmavoice_decode_init(AVCodecContext *ctx)
{
    int n, flags, pitch_range, lsp16_flag;
    WMAVoiceContext *s = ctx->priv_data;

    /**
     * Extradata layout:
     * - byte  0-18: WMAPro-in-WMAVoice extradata (see wmaprodec.c),
     * - byte 19-22: flags field (annoyingly in LE; see below for known
     *               values),
     * - byte 23-46: variable bitmode tree (really just 17 * 3 bits,
     *               rest is 0).
     */
    if (ctx->extradata_size != 46) {
        av_log(ctx, AV_LOG_ERROR,
               "Invalid extradata size %d (should be 46)\n",
               ctx->extradata_size);
        return -1;
    }
    flags                = AV_RL32(ctx->extradata + 18);
    s->spillover_bitsize = 3 + av_ceil_log2(ctx->block_align);
    s->do_apf            =    flags & 0x1;
    s->lsp_q_mode        = !!(flags & 0x2000);
    s->lsp_def_mode      = !!(flags & 0x4000);
    lsp16_flag           =    flags & 0x1000;
    if (lsp16_flag) {
        s->lsps               = 16;
        s->frame_lsp_bitsize  = 34;
        s->sframe_lsp_bitsize = 60;
    } else {
        s->lsps               = 10;
        s->frame_lsp_bitsize  = 24;
        s->sframe_lsp_bitsize = 48;
    }
    for (n = 0; n < s->lsps; n++)
        s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);

    init_get_bits(&s->gb, ctx->extradata + 22, (ctx->extradata_size - 22) << 3);
    if (decode_vbmtree(&s->gb, s->vbm_tree) < 0) {
        av_log(ctx, AV_LOG_ERROR, "Invalid VBM tree; broken extradata?\n");
        return -1;
    }

    s->min_pitch_val    = ((ctx->sample_rate << 8)      /  400 + 50) >> 8;
    s->max_pitch_val    = ((ctx->sample_rate << 8) * 37 / 2000 + 50) >> 8;
    pitch_range         = s->max_pitch_val - s->min_pitch_val;
    s->pitch_nbits      = av_ceil_log2(pitch_range);
    s->last_pitch_val   = 40;
    s->last_acb_type    = ACB_TYPE_NONE;
    s->history_nsamples = s->max_pitch_val + 8;

    if (s->min_pitch_val < 1 || s->history_nsamples > MAX_SIGNAL_HISTORY) {
        int min_sr = ((((1 << 8) - 50) * 400) + 0xFF) >> 8,
            max_sr = ((((MAX_SIGNAL_HISTORY - 8) << 8) + 205) * 2000 / 37) >> 8;

        av_log(ctx, AV_LOG_ERROR,
               "Unsupported samplerate %d (min=%d, max=%d)\n",
               ctx->sample_rate, min_sr, max_sr); // 322-22097 Hz

        return -1;
    }

    s->block_conv_table[0]      = s->min_pitch_val;
    s->block_conv_table[1]      = (pitch_range * 25) >> 6;
    s->block_conv_table[2]      = (pitch_range * 44) >> 6;
    s->block_conv_table[3]      = s->max_pitch_val - 1;
    s->block_delta_pitch_hrange = (pitch_range >> 3) & ~0xF;
    s->block_delta_pitch_nbits  = 1 + av_ceil_log2(s->block_delta_pitch_hrange);
    s->block_pitch_range        = s->block_conv_table[2] +
                                  s->block_conv_table[3] + 1 +
                                  2 * (s->block_conv_table[1] - 2 * s->min_pitch_val);
    s->block_pitch_nbits        = av_ceil_log2(s->block_pitch_range);

    ctx->sample_fmt             = SAMPLE_FMT_FLT;

    return 0;
}

/**
 * Dequantize LSPs
 * @param lsps output pointer to the array that will hold the LSPs
 * @param num number of LSPs to be dequantized
 * @param values quantized values, contains n_stages values
 * @param sizes range (i.e. max value) of each quantized value
 * @param n_stages number of dequantization runs
 * @param table dequantization table to be used
 * @param mul_q LSF multiplier
 * @param base_q base (lowest) LSF values
 */
static void dequant_lsps(double *lsps, int num,
                         const uint16_t *values,
                         const uint16_t *sizes,
                         int n_stages, const uint8_t *table,
                         const double *mul_q,
                         const double *base_q)
{
    int n, m;

    memset(lsps, 0, num * sizeof(*lsps));
    for (n = 0; n < n_stages; n++) {
        const uint8_t *t_off = &table[values[n] * num];
        double base = base_q[n], mul = mul_q[n];

        for (m = 0; m < num; m++)
            lsps[m] += base + mul * t_off[m];

        table += sizes[n] * num;
    }
}

/**
 * @defgroup lsp_dequant LSP dequantization routines
 * LSP dequantization routines, for 10/16LSPs and independent/residual coding.
 * @note we assume enough bits are available, caller should check.
 * lsp10i() consumes 24 bits; lsp10r() consumes an additional 24 bits;
 * lsp16i() consumes 34 bits; lsp16r() consumes an additional 26 bits.
 * @{
 */
/**
 * Parse 10 independently-coded LSPs.
 */
static void dequant_lsp10i(GetBitContext *gb, double *lsps)
{
    static const uint16_t vec_sizes[4] = { 256, 64, 32, 32 };
    static const double mul_lsf[4] = {
        5.2187144800e-3,    1.4626986422e-3,
        9.6179549166e-4,    1.1325736225e-3
    };
    static const double base_lsf[4] = {
        M_PI * -2.15522e-1, M_PI * -6.1646e-2,
        M_PI * -3.3486e-2,  M_PI * -5.7408e-2
    };
    uint16_t v[4];

    v[0] = get_bits(gb, 8);
    v[1] = get_bits(gb, 6);
    v[2] = get_bits(gb, 5);
    v[3] = get_bits(gb, 5);

    dequant_lsps(lsps, 10, v, vec_sizes, 4, wmavoice_dq_lsp10i,
                 mul_lsf, base_lsf);
}

/**
 * Parse 10 independently-coded LSPs, and then derive the tables to
 * generate LSPs for the other frames from them (residual coding).
 */
static void dequant_lsp10r(GetBitContext *gb,
                           double *i_lsps, const double *old,
                           double *a1, double *a2, int q_mode)
{
    static const uint16_t vec_sizes[3] = { 128, 64, 64 };
    static const double mul_lsf[3] = {
        2.5807601174e-3,    1.2354460219e-3,   1.1763821673e-3
    };
    static const double base_lsf[3] = {
        M_PI * -1.07448e-1, M_PI * -5.2706e-2, M_PI * -5.1634e-2
    };
    const float (*ipol_tab)[2][10] = q_mode ?
        wmavoice_lsp10_intercoeff_b : wmavoice_lsp10_intercoeff_a;
    uint16_t interpol, v[3];
    int n;

    dequant_lsp10i(gb, i_lsps);

    interpol = get_bits(gb, 5);
    v[0]     = get_bits(gb, 7);
    v[1]     = get_bits(gb, 6);
    v[2]     = get_bits(gb, 6);

    for (n = 0; n < 10; n++) {
        double delta = old[n] - i_lsps[n];
        a1[n]        = ipol_tab[interpol][0][n] * delta + i_lsps[n];
        a1[10 + n]   = ipol_tab[interpol][1][n] * delta + i_lsps[n];
    }

    dequant_lsps(a2, 20, v, vec_sizes, 3, wmavoice_dq_lsp10r,
                 mul_lsf, base_lsf);
}

/**
 * Parse 16 independently-coded LSPs.
 */
static void dequant_lsp16i(GetBitContext *gb, double *lsps)
{
    static const uint16_t vec_sizes[5] = { 256, 64, 128, 64, 128 };
    static const double mul_lsf[5] = {
        3.3439586280e-3,    6.9908173703e-4,
        3.3216608306e-3,    1.0334960326e-3,
        3.1899104283e-3
    };
    static const double base_lsf[5] = {
        M_PI * -1.27576e-1, M_PI * -2.4292e-2,
        M_PI * -1.28094e-1, M_PI * -3.2128e-2,
        M_PI * -1.29816e-1
    };
    uint16_t v[5];

    v[0] = get_bits(gb, 8);
    v[1] = get_bits(gb, 6);
    v[2] = get_bits(gb, 7);
    v[3] = get_bits(gb, 6);
    v[4] = get_bits(gb, 7);

    dequant_lsps( lsps,     5,  v,     vec_sizes,    2,
                 wmavoice_dq_lsp16i1,  mul_lsf,     base_lsf);
    dequant_lsps(&lsps[5],  5, &v[2], &vec_sizes[2], 2,
                 wmavoice_dq_lsp16i2, &mul_lsf[2], &base_lsf[2]);
    dequant_lsps(&lsps[10], 6, &v[4], &vec_sizes[4], 1,
                 wmavoice_dq_lsp16i3, &mul_lsf[4], &base_lsf[4]);
}

/**
 * Parse 16 independently-coded LSPs, and then derive the tables to
 * generate LSPs for the other frames from them (residual coding).
 */
static void dequant_lsp16r(GetBitContext *gb,
                           double *i_lsps, const double *old,
                           double *a1, double *a2, int q_mode)
{
    static const uint16_t vec_sizes[3] = { 128, 128, 128 };
    static const double mul_lsf[3] = {
        1.2232979501e-3,   1.4062241527e-3,   1.6114744851e-3
    };
    static const double base_lsf[3] = {
        M_PI * -5.5830e-2, M_PI * -5.2908e-2, M_PI * -5.4776e-2
    };
    const float (*ipol_tab)[2][16] = q_mode ?
        wmavoice_lsp16_intercoeff_b : wmavoice_lsp16_intercoeff_a;
    uint16_t interpol, v[3];
    int n;

    dequant_lsp16i(gb, i_lsps);

    interpol = get_bits(gb, 5);
    v[0]     = get_bits(gb, 7);
    v[1]     = get_bits(gb, 7);
    v[2]     = get_bits(gb, 7);

    for (n = 0; n < 16; n++) {
        double delta = old[n] - i_lsps[n];
        a1[n]        = ipol_tab[interpol][0][n] * delta + i_lsps[n];
        a1[16 + n]   = ipol_tab[interpol][1][n] * delta + i_lsps[n];
    }

    dequant_lsps( a2,     10,  v,     vec_sizes,    1,
                 wmavoice_dq_lsp16r1,  mul_lsf,     base_lsf);
    dequant_lsps(&a2[10], 10, &v[1], &vec_sizes[1], 1,
                 wmavoice_dq_lsp16r2, &mul_lsf[1], &base_lsf[1]);
    dequant_lsps(&a2[20], 12, &v[2], &vec_sizes[2], 1,
                 wmavoice_dq_lsp16r3, &mul_lsf[2], &base_lsf[2]);
}

/**
 * @}
 * @defgroup aw Pitch-adaptive window coding functions
 * The next few functions are for pitch-adaptive window coding.
 * @{
 */
/**
 * Parse the offset of the first pitch-adaptive window pulses, and
 * the distribution of pulses between the two blocks in this frame.
 * @param s WMA Voice decoding context private data
 * @param gb bit I/O context
 * @param pitch pitch for each block in this frame
 */
static void aw_parse_coords(WMAVoiceContext *s, GetBitContext *gb,
                            const int *pitch)
{
    static const int16_t start_offset[94] = {
        -11,  -9,  -7,  -5,  -3,  -1,   1,   3,   5,   7,   9,  11,
         13,  15,  18,  17,  19,  20,  21,  22,  23,  24,  25,  26,
         27,  28,  29,  30,  31,  32,  33,  35,  37,  39,  41,  43,
         45,  47,  49,  51,  53,  55,  57,  59,  61,  63,  65,  67,
         69,  71,  73,  75,  77,  79,  81,  83,  85,  87,  89,  91,
         93,  95,  97,  99, 101, 103, 105, 107, 109, 111, 113, 115,
        117, 119, 121, 123, 125, 127, 129, 131, 133, 135, 137, 139,
        141, 143, 145, 147, 149, 151, 153, 155, 157, 159
    };
    int bits, offset;

    /* position of pulse */
    s->aw_idx_is_ext = 0;
    if ((bits = get_bits(gb, 6)) >= 54) {
        s->aw_idx_is_ext = 1;
        bits += (bits - 54) * 3 + get_bits(gb, 2);
    }

    /* for a repeated pulse at pulse_off with a pitch_lag of pitch[], count
     * the distribution of the pulses in each block contained in this frame. */
    s->aw_pulse_range        = FFMIN(pitch[0], pitch[1]) > 32 ? 24 : 16;
    for (offset = start_offset[bits]; offset < 0; offset += pitch[0]) ;
    s->aw_n_pulses[0]        = (pitch[0] - 1 + MAX_FRAMESIZE / 2 - offset) / pitch[0];
    s->aw_first_pulse_off[0] = offset - s->aw_pulse_range / 2;
    offset                  += s->aw_n_pulses[0] * pitch[0];
    s->aw_n_pulses[1]        = (pitch[1] - 1 + MAX_FRAMESIZE - offset) / pitch[1];
    s->aw_first_pulse_off[1] = offset - (MAX_FRAMESIZE + s->aw_pulse_range) / 2;

    /* if continuing from a position before the block, reset position to
     * start of block (when corrected for the range over which it can be
     * spread in aw_pulse_set1()). */
    if (start_offset[bits] < MAX_FRAMESIZE / 2) {
        while (s->aw_first_pulse_off[1] - pitch[1] + s->aw_pulse_range > 0)
            s->aw_first_pulse_off[1] -= pitch[1];
        if (start_offset[bits] < 0)
            while (s->aw_first_pulse_off[0] - pitch[0] + s->aw_pulse_range > 0)
                s->aw_first_pulse_off[0] -= pitch[0];
    }
}

/**
 * Apply second set of pitch-adaptive window pulses.
 * @param s WMA Voice decoding context private data
 * @param gb bit I/O context
 * @param block_idx block index in frame [0, 1]
 * @param fcb structure containing fixed codebook vector info
 */
static void aw_pulse_set2(WMAVoiceContext *s, GetBitContext *gb,
                          int block_idx, AMRFixed *fcb)
{
    uint16_t use_mask[7]; // only 5 are used, rest is padding
    /* in this function, idx is the index in the 80-bit (+ padding) use_mask
     * bit-array. Since use_mask consists of 16-bit values, the lower 4 bits
     * of idx are the position of the bit within a particular item in the
     * array (0 being the most significant bit, and 15 being the least
     * significant bit), and the remainder (>> 4) is the index in the
     * use_mask[]-array. This is faster and uses less memory than using a
     * 80-byte/80-int array. */
    int pulse_off = s->aw_first_pulse_off[block_idx],
        pulse_start, n, idx, range, aidx, start_off = 0;

    /* set offset of first pulse to within this block */
    if (s->aw_n_pulses[block_idx] > 0)
        while (pulse_off + s->aw_pulse_range < 1)
            pulse_off += fcb->pitch_lag;

    /* find range per pulse */
    if (s->aw_n_pulses[0] > 0) {
        if (block_idx == 0) {
            range = 32;
        } else /* block_idx = 1 */ {
            range = 8;
            if (s->aw_n_pulses[block_idx] > 0)
                pulse_off = s->aw_next_pulse_off_cache;
        }
    } else
        range = 16;
    pulse_start = s->aw_n_pulses[block_idx] > 0 ? pulse_off - range / 2 : 0;

    /* aw_pulse_set1() already applies pulses around pulse_off (to be exactly,
     * in the range of [pulse_off, pulse_off + s->aw_pulse_range], and thus
     * we exclude that range from being pulsed again in this function. */
    memset( use_mask,   -1, 5 * sizeof(use_mask[0]));
    memset(&use_mask[5], 0, 2 * sizeof(use_mask[0]));
    if (s->aw_n_pulses[block_idx] > 0)
        for (idx = pulse_off; idx < MAX_FRAMESIZE / 2; idx += fcb->pitch_lag) {
            int excl_range         = s->aw_pulse_range; // always 16 or 24
            uint16_t *use_mask_ptr = &use_mask[idx >> 4];
            int first_sh           = 16 - (idx & 15);
            *use_mask_ptr++       &= 0xFFFF << first_sh;
            excl_range            -= first_sh;
            if (excl_range >= 16) {
                *use_mask_ptr++    = 0;
                *use_mask_ptr     &= 0xFFFF >> (excl_range - 16);
            } else
                *use_mask_ptr     &= 0xFFFF >> excl_range;
        }

    /* find the 'aidx'th offset that is not excluded */
    aidx = get_bits(gb, s->aw_n_pulses[0] > 0 ? 5 - 2 * block_idx : 4);
    for (n = 0; n <= aidx; pulse_start++) {
        for (idx = pulse_start; idx < 0; idx += fcb->pitch_lag) ;
        if (idx >= MAX_FRAMESIZE / 2) { // find from zero
            if (use_mask[0])      idx = 0x0F;
            else if (use_mask[1]) idx = 0x1F;
            else if (use_mask[2]) idx = 0x2F;
            else if (use_mask[3]) idx = 0x3F;
            else if (use_mask[4]) idx = 0x4F;
            else                  return;
            idx -= av_log2_16bit(use_mask[idx >> 4]);
        }
        if (use_mask[idx >> 4] & (0x8000 >> (idx & 15))) {
            use_mask[idx >> 4] &= ~(0x8000 >> (idx & 15));
            n++;
            start_off = idx;
        }
    }

    fcb->x[fcb->n] = start_off;
    fcb->y[fcb->n] = get_bits1(gb) ? -1.0 : 1.0;
    fcb->n++;

    /* set offset for next block, relative to start of that block */
    n = (MAX_FRAMESIZE / 2 - start_off) % fcb->pitch_lag;
    s->aw_next_pulse_off_cache = n ? fcb->pitch_lag - n : 0;
}

/**
 * Apply first set of pitch-adaptive window pulses.
 * @param s WMA Voice decoding context private data
 * @param gb bit I/O context
 * @param block_idx block index in frame [0, 1]
 * @param fcb storage location for fixed codebook pulse info
 */
static void aw_pulse_set1(WMAVoiceContext *s, GetBitContext *gb,
                          int block_idx, AMRFixed *fcb)
{
    int val = get_bits(gb, 12 - 2 * (s->aw_idx_is_ext && !block_idx));
    float v;

    if (s->aw_n_pulses[block_idx] > 0) {
        int n, v_mask, i_mask, sh, n_pulses;

        if (s->aw_pulse_range == 24) { // 3 pulses, 1:sign + 3:index each
            n_pulses = 3;
            v_mask   = 8;
            i_mask   = 7;
            sh       = 4;
        } else { // 4 pulses, 1:sign + 2:index each
            n_pulses = 4;
            v_mask   = 4;
            i_mask   = 3;
            sh       = 3;
        }

        for (n = n_pulses - 1; n >= 0; n--, val >>= sh) {
            fcb->y[fcb->n] = (val & v_mask) ? -1.0 : 1.0;
            fcb->x[fcb->n] = (val & i_mask) * n_pulses + n +
                                 s->aw_first_pulse_off[block_idx];
            while (fcb->x[fcb->n] < 0)
                fcb->x[fcb->n] += fcb->pitch_lag;
            if (fcb->x[fcb->n] < MAX_FRAMESIZE / 2)
                fcb->n++;
        }
    } else {
        int num2 = (val & 0x1FF) >> 1, delta, idx;

        if (num2 < 1 * 79)      { delta = 1; idx = num2 + 1; }
        else if (num2 < 2 * 78) { delta = 3; idx = num2 + 1 - 1 * 77; }
        else if (num2 < 3 * 77) { delta = 5; idx = num2 + 1 - 2 * 76; }
        else                    { delta = 7; idx = num2 + 1 - 3 * 75; }
        v = (val & 0x200) ? -1.0 : 1.0;

        fcb->no_repeat_mask |= 3 << fcb->n;
        fcb->x[fcb->n]       = idx - delta;
        fcb->y[fcb->n]       = v;
        fcb->x[fcb->n + 1]   = idx;
        fcb->y[fcb->n + 1]   = (val & 1) ? -v : v;
        fcb->n              += 2;
    }
}

/**
 * @}
 *
 * Generate a random number from frame_cntr and block_idx, which will lief
 * in the range [0, 1000 - block_size] (so it can be used as an index in a
 * table of size 1000 of which you want to read block_size entries).
 *
 * @param frame_cntr current frame number
 * @param block_num current block index
 * @param block_size amount of entries we want to read from a table
 *                   that has 1000 entries
 * @return a (non-)random number in the [0, 1000 - block_size] range.
 */
static int pRNG(int frame_cntr, int block_num, int block_size)
{
    /* array to simplify the calculation of z:
     * y = (x % 9) * 5 + 6;
     * z = (49995 * x) / y;
     * Since y only has 9 values, we can remove the division by using a
     * LUT and using FASTDIV-style divisions. For each of the 9 values
     * of y, we can rewrite z as:
     * z = x * (49995 / y) + x * ((49995 % y) / y)
     * In this table, each col represents one possible value of y, the
     * first number is 49995 / y, and the second is the FASTDIV variant
     * of 49995 % y / y. */
    static const unsigned int div_tbl[9][2] = {
        { 8332,  3 * 715827883U }, // y =  6
        { 4545,  0 * 390451573U }, // y = 11
        { 3124, 11 * 268435456U }, // y = 16
        { 2380, 15 * 204522253U }, // y = 21
        { 1922, 23 * 165191050U }, // y = 26
        { 1612, 23 * 138547333U }, // y = 31
        { 1388, 27 * 119304648U }, // y = 36
        { 1219, 16 * 104755300U }, // y = 41
        { 1086, 39 *  93368855U }  // y = 46
    };
    unsigned int z, y, x = MUL16(block_num, 1877) + frame_cntr;
    if (x >= 0xFFFF) x -= 0xFFFF;   // max value of x is 8*1877+0xFFFE=0x13AA6,
                                    // so this is effectively a modulo (%)
    y = x - 9 * MULH(477218589, x); // x % 9
    z = (uint16_t) (x * div_tbl[y][0] + UMULH(x, div_tbl[y][1]));
                                    // z = x * 49995 / (y * 5 + 6)
    return z % (1000 - block_size);
}

/**
 * Parse hardcoded signal for a single block.
 * @note see #synth_block().
 */
static void synth_block_hardcoded(WMAVoiceContext *s, GetBitContext *gb,
                                 int block_idx, int size,
                                 const struct frame_type_desc *frame_desc,
                                 float *excitation)
{
    float gain;
    int n, r_idx;

    assert(size <= MAX_FRAMESIZE);

    /* Set the offset from which we start reading wmavoice_std_codebook */
    if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
        r_idx = pRNG(s->frame_cntr, block_idx, size);
        gain  = s->silence_gain;
    } else /* FCB_TYPE_HARDCODED */ {
        r_idx = get_bits(gb, 8);
        gain  = wmavoice_gain_universal[get_bits(gb, 6)];
    }

    /* Clear gain prediction parameters */
    memset(s->gain_pred_err, 0, sizeof(s->gain_pred_err));

    /* Apply gain to hardcoded codebook and use that as excitation signal */
    for (n = 0; n < size; n++)
        excitation[n] = wmavoice_std_codebook[r_idx + n] * gain;
}

/**
 * Parse FCB/ACB signal for a single block.
 * @note see #synth_block().
 */
static void synth_block_fcb_acb(WMAVoiceContext *s, GetBitContext *gb,
                                int block_idx, int size,
                                int block_pitch_sh2,
                                const struct frame_type_desc *frame_desc,
                                float *excitation)
{
    static const float gain_coeff[6] = {
        0.8169, -0.06545, 0.1726, 0.0185, -0.0359, 0.0458
    };
    float pulses[MAX_FRAMESIZE / 2], pred_err, acb_gain, fcb_gain;
    int n, idx, gain_weight;
    AMRFixed fcb;

    assert(size <= MAX_FRAMESIZE / 2);
    memset(pulses, 0, sizeof(*pulses) * size);

    fcb.pitch_lag      = block_pitch_sh2 >> 2;
    fcb.pitch_fac      = 1.0;
    fcb.no_repeat_mask = 0;
    fcb.n              = 0;

    /* For the other frame types, this is where we apply the innovation
     * (fixed) codebook pulses of the speech signal. */
    if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
        aw_pulse_set1(s, gb, block_idx, &fcb);
        aw_pulse_set2(s, gb, block_idx, &fcb);
    } else /* FCB_TYPE_EXC_PULSES */ {
        int offset_nbits = 5 - frame_desc->log_n_blocks;

        fcb.no_repeat_mask = -1;
        /* similar to ff_decode_10_pulses_35bits(), but with single pulses
         * (instead of double) for a subset of pulses */
        for (n = 0; n < 5; n++) {
            float sign;
            int pos1, pos2;

            sign           = get_bits1(gb) ? 1.0 : -1.0;
            pos1           = get_bits(gb, offset_nbits);
            fcb.x[fcb.n]   = n + 5 * pos1;
            fcb.y[fcb.n++] = sign;
            if (n < frame_desc->dbl_pulses) {
                pos2           = get_bits(gb, offset_nbits);
                fcb.x[fcb.n]   = n + 5 * pos2;
                fcb.y[fcb.n++] = (pos1 < pos2) ? -sign : sign;
            }
        }
    }
    ff_set_fixed_vector(pulses, &fcb, 1.0, size);

    /* Calculate gain for adaptive & fixed codebook signal.
     * see ff_amr_set_fixed_gain(). */
    idx = get_bits(gb, 7);
    fcb_gain = expf(ff_dot_productf(s->gain_pred_err, gain_coeff, 6) -
                    5.2409161640 + wmavoice_gain_codebook_fcb[idx]);
    acb_gain = wmavoice_gain_codebook_acb[idx];
    pred_err = av_clipf(wmavoice_gain_codebook_fcb[idx],
                        -2.9957322736 /* log(0.05) */,
                         1.6094379124 /* log(5.0)  */);

    gain_weight = 8 >> frame_desc->log_n_blocks;
    memmove(&s->gain_pred_err[gain_weight], s->gain_pred_err,
            sizeof(*s->gain_pred_err) * (6 - gain_weight));
    for (n = 0; n < gain_weight; n++)
        s->gain_pred_err[n] = pred_err;

    /* Calculation of adaptive codebook */
    if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
        int len;
        for (n = 0; n < size; n += len) {
            int next_idx_sh16;
            int abs_idx    = block_idx * size + n;
            int pitch_sh16 = (s->last_pitch_val << 16) +
                             s->pitch_diff_sh16 * abs_idx;
            int pitch      = (pitch_sh16 + 0x6FFF) >> 16;
            int idx_sh16   = ((pitch << 16) - pitch_sh16) * 8 + 0x58000;
            idx            = idx_sh16 >> 16;
            if (s->pitch_diff_sh16) {
                if (s->pitch_diff_sh16 > 0) {
                    next_idx_sh16 = (idx_sh16) &~ 0xFFFF;
                } else
                    next_idx_sh16 = (idx_sh16 + 0x10000) &~ 0xFFFF;
                len = av_clip((idx_sh16 - next_idx_sh16) / s->pitch_diff_sh16 / 8,
                              1, size - n);
            } else
                len = size;

            ff_acelp_interpolatef(&excitation[n], &excitation[n - pitch],
                                  wmavoice_ipol1_coeffs, 17,
                                  idx, 9, len);
        }
    } else /* ACB_TYPE_HAMMING */ {
        int block_pitch = block_pitch_sh2 >> 2;
        idx             = block_pitch_sh2 & 3;
        if (idx) {
            ff_acelp_interpolatef(excitation, &excitation[-block_pitch],
                                  wmavoice_ipol2_coeffs, 4,
                                  idx, 8, size);
        } else
            av_memcpy_backptr(excitation, sizeof(float) * block_pitch,
                              sizeof(float) * size);
    }

    /* Interpolate ACB/FCB and use as excitation signal */
    ff_weighted_vector_sumf(excitation, excitation, pulses,
                            acb_gain, fcb_gain, size);
}

/**
 * Parse data in a single block.
 * @note we assume enough bits are available, caller should check.
 *
 * @param s WMA Voice decoding context private data
 * @param gb bit I/O context
 * @param block_idx index of the to-be-read block
 * @param size amount of samples to be read in this block
 * @param block_pitch_sh2 pitch for this block << 2
 * @param lsps LSPs for (the end of) this frame
 * @param prev_lsps LSPs for the last frame
 * @param frame_desc frame type descriptor
 * @param excitation target memory for the ACB+FCB interpolated signal
 * @param synth target memory for the speech synthesis filter output
 * @return 0 on success, <0 on error.
 */
static void synth_block(WMAVoiceContext *s, GetBitContext *gb,
                        int block_idx, int size,
                        int block_pitch_sh2,
                        const double *lsps, const double *prev_lsps,
                        const struct frame_type_desc *frame_desc,
                        float *excitation, float *synth)
{
    double i_lsps[MAX_LSPS];
    float lpcs[MAX_LSPS];
    float fac;
    int n;

    if (frame_desc->acb_type == ACB_TYPE_NONE)
        synth_block_hardcoded(s, gb, block_idx, size, frame_desc, excitation);
    else
        synth_block_fcb_acb(s, gb, block_idx, size, block_pitch_sh2,
                            frame_desc, excitation);

    /* convert interpolated LSPs to LPCs */
    fac = (block_idx + 0.5) / frame_desc->n_blocks;
    for (n = 0; n < s->lsps; n++) // LSF -> LSP
        i_lsps[n] = cos(prev_lsps[n] + fac * (lsps[n] - prev_lsps[n]));
    ff_acelp_lspd2lpc(i_lsps, lpcs, s->lsps >> 1);

    /* Speech synthesis */
    ff_celp_lp_synthesis_filterf(synth, lpcs, excitation, size, s->lsps);
}

/**
 * Synthesize output samples for a single frame.
 * @note we assume enough bits are available, caller should check.
 *
 * @param ctx WMA Voice decoder context
 * @param gb bit I/O context (s->gb or one for cross-packet superframes)
 * @param samples pointer to output sample buffer, has space for at least 160
 *                samples
 * @param lsps LSP array
 * @param prev_lsps array of previous frame's LSPs
 * @param excitation target buffer for excitation signal
 * @param synth target buffer for synthesized speech data
 * @return 0 on success, <0 on error.
 */
static int synth_frame(AVCodecContext *ctx, GetBitContext *gb,
                       float *samples,
                       const double *lsps, const double *prev_lsps,
                       float *excitation, float *synth)
{
    WMAVoiceContext *s = ctx->priv_data;
    int n, n_blocks_x2, log_n_blocks_x2, cur_pitch_val;
    int pitch[MAX_BLOCKS], last_block_pitch;

    /* Parse frame type ("frame header"), see frame_descs */
    int bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)],
        block_nsamples = MAX_FRAMESIZE / frame_descs[bd_idx].n_blocks;

    if (bd_idx < 0) {
        av_log(ctx, AV_LOG_ERROR,
               "Invalid frame type VLC code, skipping\n");
        return -1;
    }

    /* Pitch calculation for ACB_TYPE_ASYMMETRIC ("pitch-per-frame") */
    if (frame_descs[bd_idx].acb_type == ACB_TYPE_ASYMMETRIC) {
        /* Pitch is provided per frame, which is interpreted as the pitch of
         * the last sample of the last block of this frame. We can interpolate
         * the pitch of other blocks (and even pitch-per-sample) by gradually
         * incrementing/decrementing prev_frame_pitch to cur_pitch_val. */
        n_blocks_x2      = frame_descs[bd_idx].n_blocks << 1;
        log_n_blocks_x2  = frame_descs[bd_idx].log_n_blocks + 1;
        cur_pitch_val    = s->min_pitch_val + get_bits(gb, s->pitch_nbits);
        cur_pitch_val    = FFMIN(cur_pitch_val, s->max_pitch_val - 1);
        if (s->last_acb_type == ACB_TYPE_NONE ||
            20 * abs(cur_pitch_val - s->last_pitch_val) >
                (cur_pitch_val + s->last_pitch_val))
            s->last_pitch_val = cur_pitch_val;

        /* pitch per block */
        for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
            int fac = n * 2 + 1;

            pitch[n] = (MUL16(fac,                 cur_pitch_val) +
                        MUL16((n_blocks_x2 - fac), s->last_pitch_val) +
                        frame_descs[bd_idx].n_blocks) >> log_n_blocks_x2;
        }

        /* "pitch-diff-per-sample" for calculation of pitch per sample */
        s->pitch_diff_sh16 =
            ((cur_pitch_val - s->last_pitch_val) << 16) / MAX_FRAMESIZE;
    }

    /* Global gain (if silence) and pitch-adaptive window coordinates */
    switch (frame_descs[bd_idx].fcb_type) {
    case FCB_TYPE_SILENCE:
        s->silence_gain = wmavoice_gain_silence[get_bits(gb, 8)];
        break;
    case FCB_TYPE_AW_PULSES:
        aw_parse_coords(s, gb, pitch);
        break;
    }

    for (n = 0; n < frame_descs[bd_idx].n_blocks; n++) {
        int bl_pitch_sh2;

        /* Pitch calculation for ACB_TYPE_HAMMING ("pitch-per-block") */
        switch (frame_descs[bd_idx].acb_type) {
        case ACB_TYPE_HAMMING: {
            /* Pitch is given per block. Per-block pitches are encoded as an
             * absolute value for the first block, and then delta values
             * relative to this value) for all subsequent blocks. The scale of
             * this pitch value is semi-logaritmic compared to its use in the
             * decoder, so we convert it to normal scale also. */
            int block_pitch,
                t1 = (s->block_conv_table[1] - s->block_conv_table[0]) << 2,
                t2 = (s->block_conv_table[2] - s->block_conv_table[1]) << 1,
                t3 =  s->block_conv_table[3] - s->block_conv_table[2] + 1;

            if (n == 0) {
                block_pitch = get_bits(gb, s->block_pitch_nbits);
            } else
                block_pitch = last_block_pitch - s->block_delta_pitch_hrange +
                                 get_bits(gb, s->block_delta_pitch_nbits);
            /* Convert last_ so that any next delta is within _range */
            last_block_pitch = av_clip(block_pitch,
                                       s->block_delta_pitch_hrange,
                                       s->block_pitch_range -
                                           s->block_delta_pitch_hrange);

            /* Convert semi-log-style scale back to normal scale */
            if (block_pitch < t1) {
                bl_pitch_sh2 = (s->block_conv_table[0] << 2) + block_pitch;
            } else {
                block_pitch -= t1;
                if (block_pitch < t2) {
                    bl_pitch_sh2 =
                        (s->block_conv_table[1] << 2) + (block_pitch << 1);
                } else {
                    block_pitch -= t2;
                    if (block_pitch < t3) {
                        bl_pitch_sh2 =
                            (s->block_conv_table[2] + block_pitch) << 2;
                    } else
                        bl_pitch_sh2 = s->block_conv_table[3] << 2;
                }
            }
            pitch[n] = bl_pitch_sh2 >> 2;
            break;
        }

        case ACB_TYPE_ASYMMETRIC: {
            bl_pitch_sh2 = pitch[n] << 2;
            break;
        }

        default: // ACB_TYPE_NONE has no pitch
            bl_pitch_sh2 = 0;
            break;
        }

        synth_block(s, gb, n, block_nsamples, bl_pitch_sh2,
                    lsps, prev_lsps, &frame_descs[bd_idx],
                    &excitation[n * block_nsamples],
                    &synth[n * block_nsamples]);
    }

    /* Averaging projection filter, if applicable. Else, just copy samples
     * from synthesis buffer */
    if (s->do_apf) {
        // FIXME this is where APF would take place, currently not implemented
        av_log_missing_feature(ctx, "APF", 0);
        s->do_apf = 0;
    } //else
        for (n = 0; n < 160; n++)
            samples[n] = av_clipf(synth[n], -1.0, 1.0);

    /* Cache values for next frame */
    s->frame_cntr++;
    if (s->frame_cntr >= 0xFFFF) s->frame_cntr -= 0xFFFF; // i.e. modulo (%)
    s->last_acb_type = frame_descs[bd_idx].acb_type;
    switch (frame_descs[bd_idx].acb_type) {
    case ACB_TYPE_NONE:
        s->last_pitch_val = 0;
        break;
    case ACB_TYPE_ASYMMETRIC:
        s->last_pitch_val = cur_pitch_val;
        break;
    case ACB_TYPE_HAMMING:
        s->last_pitch_val = pitch[frame_descs[bd_idx].n_blocks - 1];
        break;
    }

    return 0;
}

/**
 * Ensure minimum value for first item, maximum value for last value,
 * proper spacing between each value and proper ordering.
 *
 * @param lsps array of LSPs
 * @param num size of LSP array
 *
 * @note basically a double version of #ff_acelp_reorder_lsf(), might be
 *       useful to put in a generic location later on. Parts are also
 *       present in #ff_set_min_dist_lsf() + #ff_sort_nearly_sorted_floats(),
 *       which is in float.
 */
static void stabilize_lsps(double *lsps, int num)
{
    int n, m, l;

    /* set minimum value for first, maximum value for last and minimum
     * spacing between LSF values.
     * Very similar to ff_set_min_dist_lsf(), but in double. */
    lsps[0]       = FFMAX(lsps[0],       0.0015 * M_PI);
    for (n = 1; n < num; n++)
        lsps[n]   = FFMAX(lsps[n],       lsps[n - 1] + 0.0125 * M_PI);
    lsps[num - 1] = FFMIN(lsps[num - 1], 0.9985 * M_PI);

    /* reorder (looks like one-time / non-recursed bubblesort).
     * Very similar to ff_sort_nearly_sorted_floats(), but in double. */
    for (n = 1; n < num; n++) {
        if (lsps[n] < lsps[n - 1]) {
            for (m = 1; m < num; m++) {
                double tmp = lsps[m];
                for (l = m - 1; l >= 0; l--) {
                    if (lsps[l] <= tmp) break;
                    lsps[l + 1] = lsps[l];
                }
                lsps[l + 1] = tmp;
            }
            break;
        }
    }
}

/**
 * Test if there's enough bits to read 1 superframe.
 *
 * @param orig_gb bit I/O context used for reading. This function
 *                does not modify the state of the bitreader; it
 *                only uses it to copy the current stream position
 * @param s WMA Voice decoding context private data
 * @return -1 if unsupported, 1 on not enough bits or 0 if OK.
 */
static int check_bits_for_superframe(GetBitContext *orig_gb,
                                     WMAVoiceContext *s)
{
    GetBitContext s_gb, *gb = &s_gb;
    int n, need_bits, bd_idx;
    const struct frame_type_desc *frame_desc;

    /* initialize a copy */
    init_get_bits(gb, orig_gb->buffer, orig_gb->size_in_bits);
    skip_bits_long(gb, get_bits_count(orig_gb));
    assert(get_bits_left(gb) == get_bits_left(orig_gb));

    /* superframe header */
    if (get_bits_left(gb) < 14)
        return 1;
    if (!get_bits1(gb))
        return -1;                        // WMAPro-in-WMAVoice superframe
    if (get_bits1(gb)) skip_bits(gb, 12); // number of  samples in superframe
    if (s->has_residual_lsps) {           // residual LSPs (for all frames)
        if (get_bits_left(gb) < s->sframe_lsp_bitsize)
            return 1;
        skip_bits_long(gb, s->sframe_lsp_bitsize);
    }

    /* frames */
    for (n = 0; n < MAX_FRAMES; n++) {
        int aw_idx_is_ext = 0;

        if (!s->has_residual_lsps) {     // independent LSPs (per-frame)
           if (get_bits_left(gb) < s->frame_lsp_bitsize) return 1;
           skip_bits_long(gb, s->frame_lsp_bitsize);
        }
        bd_idx = s->vbm_tree[get_vlc2(gb, frame_type_vlc.table, 6, 3)];
        if (bd_idx < 0)
            return -1;                   // invalid frame type VLC code
        frame_desc = &frame_descs[bd_idx];
        if (frame_desc->acb_type == ACB_TYPE_ASYMMETRIC) {
            if (get_bits_left(gb) < s->pitch_nbits)
                return 1;
            skip_bits_long(gb, s->pitch_nbits);
        }
        if (frame_desc->fcb_type == FCB_TYPE_SILENCE) {
            skip_bits(gb, 8);
        } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
            int tmp = get_bits(gb, 6);
            if (tmp >= 0x36) {
                skip_bits(gb, 2);
                aw_idx_is_ext = 1;
            }
        }

        /* blocks */
        if (frame_desc->acb_type == ACB_TYPE_HAMMING) {
            need_bits = s->block_pitch_nbits +
                (frame_desc->n_blocks - 1) * s->block_delta_pitch_nbits;
        } else if (frame_desc->fcb_type == FCB_TYPE_AW_PULSES) {
            need_bits = 2 * !aw_idx_is_ext;
        } else
            need_bits = 0;
        need_bits += frame_desc->frame_size;
        if (get_bits_left(gb) < need_bits)
            return 1;
        skip_bits_long(gb, need_bits);
    }

    return 0;
}

/**
 * Synthesize output samples for a single superframe. If we have any data
 * cached in s->sframe_cache, that will be used instead of whatever is loaded
 * in s->gb.
 *
 * WMA Voice superframes contain 3 frames, each containing 160 audio samples,
 * to give a total of 480 samples per frame. See #synth_frame() for frame
 * parsing. In addition to 3 frames, superframes can also contain the LSPs
 * (if these are globally specified for all frames (residually); they can
 * also be specified individually per-frame. See the s->has_residual_lsps
 * option), and can specify the number of samples encoded in this superframe
 * (if less than 480), usually used to prevent blanks at track boundaries.
 *
 * @param ctx WMA Voice decoder context
 * @param samples pointer to output buffer for voice samples
 * @param data_size pointer containing the size of #samples on input, and the
 *                  amount of #samples filled on output
 * @return 0 on success, <0 on error or 1 if there was not enough data to
 *         fully parse the superframe
 */
static int synth_superframe(AVCodecContext *ctx,
                            float *samples, int *data_size)
{
    WMAVoiceContext *s = ctx->priv_data;
    GetBitContext *gb = &s->gb, s_gb;
    int n, res, n_samples = 480;
    double lsps[MAX_FRAMES][MAX_LSPS];
    const double *mean_lsf = s->lsps == 16 ?
        wmavoice_mean_lsf16[s->lsp_def_mode] : wmavoice_mean_lsf10[s->lsp_def_mode];
    float excitation[MAX_SIGNAL_HISTORY + MAX_SFRAMESIZE + 12];
    float synth[MAX_LSPS + MAX_SFRAMESIZE];

    memcpy(synth,      s->synth_history,
           s->lsps             * sizeof(*synth));
    memcpy(excitation, s->excitation_history,
           s->history_nsamples * sizeof(*excitation));

    if (s->sframe_cache_size > 0) {
        gb = &s_gb;
        init_get_bits(gb, s->sframe_cache, s->sframe_cache_size);
        s->sframe_cache_size = 0;
    }

    if ((res = check_bits_for_superframe(gb, s)) == 1) return 1;

    /* First bit is speech/music bit, it differentiates between WMAVoice
     * speech samples (the actual codec) and WMAVoice music samples, which
     * are really WMAPro-in-WMAVoice-superframes. I've never seen those in
     * the wild yet. */
    if (!get_bits1(gb)) {
        av_log_missing_feature(ctx, "WMAPro-in-WMAVoice support", 1);
        return -1;
    }

    /* (optional) nr. of samples in superframe; always <= 480 and >= 0 */
    if (get_bits1(gb)) {
        if ((n_samples = get_bits(gb, 12)) > 480) {
            av_log(ctx, AV_LOG_ERROR,
                   "Superframe encodes >480 samples (%d), not allowed\n",
                   n_samples);
            return -1;
        }
    }
    /* Parse LSPs, if global for the superframe (can also be per-frame). */
    if (s->has_residual_lsps) {
        double prev_lsps[MAX_LSPS], a1[MAX_LSPS * 2], a2[MAX_LSPS * 2];

        for (n = 0; n < s->lsps; n++)
            prev_lsps[n] = s->prev_lsps[n] - mean_lsf[n];

        if (s->lsps == 10) {
            dequant_lsp10r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);
        } else /* s->lsps == 16 */
            dequant_lsp16r(gb, lsps[2], prev_lsps, a1, a2, s->lsp_q_mode);

        for (n = 0; n < s->lsps; n++) {
            lsps[0][n]  = mean_lsf[n] + (a1[n]           - a2[n * 2]);
            lsps[1][n]  = mean_lsf[n] + (a1[s->lsps + n] - a2[n * 2 + 1]);
            lsps[2][n] += mean_lsf[n];
        }
        for (n = 0; n < 3; n++)
            stabilize_lsps(lsps[n], s->lsps);
    }

    /* Parse frames, optionally preceeded by per-frame (independent) LSPs. */
    for (n = 0; n < 3; n++) {
        if (!s->has_residual_lsps) {
            int m;

            if (s->lsps == 10) {
                dequant_lsp10i(gb, lsps[n]);
            } else /* s->lsps == 16 */
                dequant_lsp16i(gb, lsps[n]);

            for (m = 0; m < s->lsps; m++)
                lsps[n][m] += mean_lsf[m];
            stabilize_lsps(lsps[n], s->lsps);
        }

        if ((res = synth_frame(ctx, gb,
                               &samples[n * MAX_FRAMESIZE],
                               lsps[n], n == 0 ? s->prev_lsps : lsps[n - 1],
                               &excitation[s->history_nsamples + n * MAX_FRAMESIZE],
                               &synth[s->lsps + n * MAX_FRAMESIZE])))
            return res;
    }

    /* Statistics? FIXME - we don't check for length, a slight overrun
     * will be caught by internal buffer padding, and anything else
     * will be skipped, not read. */
    if (get_bits1(gb)) {
        res = get_bits(gb, 4);
        skip_bits(gb, 10 * (res + 1));
    }

    /* Specify nr. of output samples */
    *data_size = n_samples * sizeof(float);

    /* Update history */
    memcpy(s->prev_lsps,           lsps[2],
           s->lsps             * sizeof(*s->prev_lsps));
    memcpy(s->synth_history,      &synth[MAX_SFRAMESIZE],
           s->lsps             * sizeof(*synth));
    memcpy(s->excitation_history, &excitation[MAX_SFRAMESIZE],
           s->history_nsamples * sizeof(*excitation));

    return 0;
}

/**
 * Parse the packet header at the start of each packet (input data to this
 * decoder).
 *
 * @param s WMA Voice decoding context private data
 * @return 1 if not enough bits were available, or 0 on success.
 */
static int parse_packet_header(WMAVoiceContext *s)
{
    GetBitContext *gb = &s->gb;
    unsigned int res;

    if (get_bits_left(gb) < 11)
        return 1;
    skip_bits(gb, 4);          // packet sequence number
    s->has_residual_lsps = get_bits1(gb);
    do {
        res = get_bits(gb, 6); // number of superframes per packet
                               // (minus first one if there is spillover)
        if (get_bits_left(gb) < 6 * (res == 0x3F) + s->spillover_bitsize)
            return 1;
    } while (res == 0x3F);
    s->spillover_nbits   = get_bits(gb, s->spillover_bitsize);

    return 0;
}

/**
 * Copy (unaligned) bits from gb/data/size to pb.
 *
 * @param pb target buffer to copy bits into
 * @param data source buffer to copy bits from
 * @param size size of the source data, in bytes
 * @param gb bit I/O context specifying the current position in the source.
 *           data. This function might use this to align the bit position to
 *           a whole-byte boundary before calling #ff_copy_bits() on aligned
 *           source data
 * @param nbits the amount of bits to copy from source to target
 *
 * @note after calling this function, the current position in the input bit
 *       I/O context is undefined.
 */
static void copy_bits(PutBitContext *pb,
                      const uint8_t *data, int size,
                      GetBitContext *gb, int nbits)
{
    int rmn_bytes, rmn_bits;

    rmn_bits = rmn_bytes = get_bits_left(gb);
    if (rmn_bits < nbits)
        return;
    rmn_bits &= 7; rmn_bytes >>= 3;
    if ((rmn_bits = FFMIN(rmn_bits, nbits)) > 0)
        put_bits(pb, rmn_bits, get_bits(gb, rmn_bits));
    ff_copy_bits(pb, data + size - rmn_bytes,
                 FFMIN(nbits - rmn_bits, rmn_bytes << 3));
}

/**
 * Packet decoding: a packet is anything that the (ASF) demuxer contains,
 * and we expect that the demuxer / application provides it to us as such
 * (else you'll probably get garbage as output). Every packet has a size of
 * ctx->block_align bytes, starts with a packet header (see
 * #parse_packet_header()), and then a series of superframes. Superframe
 * boundaries may exceed packets, i.e. superframes can split data over
 * multiple (two) packets.
 *
 * For more information about frames, see #synth_superframe().
 */
static int wmavoice_decode_packet(AVCodecContext *ctx, void *data,
                                  int *data_size, AVPacket *avpkt)
{
    WMAVoiceContext *s = ctx->priv_data;
    GetBitContext *gb = &s->gb;
    int size, res, pos;

    if (*data_size < 480 * sizeof(float)) {
        av_log(ctx, AV_LOG_ERROR,
               "Output buffer too small (%d given - %lu needed)\n",
               *data_size, 480 * sizeof(float));
        return -1;
    }
    *data_size = 0;

    /* Packets are sometimes a multiple of ctx->block_align, with a packet
     * header at each ctx->block_align bytes. However, FFmpeg's ASF demuxer
     * feeds us ASF packets, which may concatenate multiple "codec" packets
     * in a single "muxer" packet, so we artificially emulate that by
     * capping the packet size at ctx->block_align. */
    for (size = avpkt->size; size > ctx->block_align; size -= ctx->block_align);
    if (!size)
        return 0;
    init_get_bits(&s->gb, avpkt->data, size << 3);

    /* size == ctx->block_align is used to indicate whether we are dealing with
     * a new packet or a packet of which we already read the packet header
     * previously. */
    if (size == ctx->block_align) { // new packet header
        if ((res = parse_packet_header(s)) < 0)
            return res;

        /* If the packet header specifies a s->spillover_nbits, then we want
         * to push out all data of the previous packet (+ spillover) before
         * continuing to parse new superframes in the current packet. */
        if (s->spillover_nbits > 0) {
            if (s->sframe_cache_size > 0) {
                int cnt = get_bits_count(gb);
                copy_bits(&s->pb, avpkt->data, size, gb, s->spillover_nbits);
                flush_put_bits(&s->pb);
                s->sframe_cache_size += s->spillover_nbits;
                if ((res = synth_superframe(ctx, data, data_size)) == 0 &&
                    *data_size > 0) {
                    cnt += s->spillover_nbits;
                    s->skip_bits_next = cnt & 7;
                    return cnt >> 3;
                } else
                    skip_bits_long (gb, s->spillover_nbits - cnt +
                                    get_bits_count(gb)); // resync
            } else
                skip_bits_long(gb, s->spillover_nbits);  // resync
        }
    } else if (s->skip_bits_next)
        skip_bits(gb, s->skip_bits_next);

    /* Try parsing superframes in current packet */
    s->sframe_cache_size = 0;
    s->skip_bits_next = 0;
    pos = get_bits_left(gb);
    if ((res = synth_superframe(ctx, data, data_size)) < 0) {
        return res;
    } else if (*data_size > 0) {
        int cnt = get_bits_count(gb);
        s->skip_bits_next = cnt & 7;
        return cnt >> 3;
    } else if ((s->sframe_cache_size = pos) > 0) {
        /* rewind bit reader to start of last (incomplete) superframe... */
        init_get_bits(gb, avpkt->data, size << 3);
        skip_bits_long(gb, (size << 3) - pos);
        assert(get_bits_left(gb) == pos);

        /* ...and cache it for spillover in next packet */
        init_put_bits(&s->pb, s->sframe_cache, SFRAME_CACHE_MAXSIZE);
        copy_bits(&s->pb, avpkt->data, size, gb, s->sframe_cache_size);
        // FIXME bad - just copy bytes as whole and add use the
        // skip_bits_next field
    }

    return size;
}

static av_cold void wmavoice_flush(AVCodecContext *ctx)
{
    WMAVoiceContext *s = ctx->priv_data;
    int n;

    s->sframe_cache_size = 0;
    s->skip_bits_next    = 0;
    for (n = 0; n < s->lsps; n++)
        s->prev_lsps[n] = M_PI * (n + 1.0) / (s->lsps + 1.0);
    memset(s->excitation_history, 0,
           sizeof(*s->excitation_history) * MAX_SIGNAL_HISTORY);
    memset(s->synth_history,      0,
           sizeof(*s->synth_history)      * MAX_LSPS);
    memset(s->gain_pred_err,      0,
           sizeof(s->gain_pred_err));
}

AVCodec wmavoice_decoder = {
    "wmavoice",
    AVMEDIA_TYPE_AUDIO,
    CODEC_ID_WMAVOICE,
    sizeof(WMAVoiceContext),
    wmavoice_decode_init,
    NULL,
    NULL,
    wmavoice_decode_packet,
    CODEC_CAP_SUBFRAMES,
    .flush     = wmavoice_flush,
    .long_name = NULL_IF_CONFIG_SMALL("Windows Media Audio Voice"),
};