Enhanced AC-3 coding tools

Enhanced AC-3 keeps the AC-3 filter bank, exponents and bit allocation, and adds four tools, each switched on in the audio frame or block: the Adaptive Hybrid Transform for steady sounds, spectral extension and enhanced coupling for cheaper high frequencies, and transient pre-noise processing after the transform. This page follows clause E.2 of ETSI TS 102 366.

Adaptive Hybrid Transform (AHT)

For steady signals a second transform, a six-point DCT-II across the six blocks of a frame, is applied to each bin’s MDCT coefficients, so one bin covers 1,536 samples instead of 256. It needs six-block frames (ahte exists only then) and exponents sent once in the frame: a channel may use it (chahtinu) only if exactly one of its six exponent strategies is new (so block 0), the coupling channel (cplahtinu) only if coupling is on in all six blocks and its strategy or exponents are new in just one block, and the LFE likewise (lfeahtinu). All six blocks’ mantissas for such a channel come in block 0, where the channel’s mantissas would be.

High-efficiency bit allocation

The bit allocation runs as in AC-3 up to the masking curve and its delta adjustments, but the last look-up uses hebaptab instead of baptab, giving a pointer from 0 to 19:

Address	hebap	Address	hebap	Address	hebap
0 to 7	same as the address	19–22	11	35–38	15
8–11	8	23–26	12	39–42	16
12–14	9	27–30	13	43–46	17
15–18	10	31–34	14	47–54	18
				55–63	19

Quantisers by hebap (Table E.2.2)
hebap	Quantiser	Bits	hebap	Quantiser	Levels	Bits
0	none	0	10	symmetric + GAQ	31	5
1	vector, 4 rows	2 for six	11	symmetric + GAQ	63	6
2	vector, 8 rows	3 for six	12	symmetric + GAQ	127	7
3	vector, 16 rows	4 for six	13	symmetric + GAQ	255	8
4	vector, 32 rows	5 for six	14	symmetric + GAQ	511	9
5	vector, 128 rows	7 for six	15	symmetric + GAQ	1,023	10
6	vector, 256 rows	8 for six	16	symmetric + GAQ	2,047	11
7	vector, 512 rows	9 for six	17	symmetric + GAQ	4,095	12
8	symmetric + GAQ	3 (7 levels)	18	symmetric + GAQ	16,383	14
9	symmetric + GAQ	4 (15 levels)	19	symmetric + GAQ	65,535	16

Vector quantisation

A bin with hebap 1 to 7 carries one index for all six of its values. The index picks a row of six 16-bit two’s complement values (fractions of 32,768) from the table for that hebap: Tables E.3.1 to E.3.7, with 4, 8, 16, 32, 128, 256 and 512 rows, 956 in all. The encoder sends the row nearest its six values. The tables are printed in full in Annex E.3 of TS 102 366, and this site’s decoder holds them in explorer/formats/shared/audio/ac3-vq.js; the first row of the hebap 1 table, for example, is 1BFF 1283 0452 10AD 28AC 12D4.

Gain-adaptive quantisation (GAQ)

Bins with hebap 8 or more carry six separate mantissas. Each AHT channel sends a 2-bit mode (chgaqmod, cplgaqmod or lfegaqmod) and then, for the bins in the mode’s range, gains: a gain Gk of 2 or 4 lets the encoder send the small values of that bin with fewer bits.

gaqmod	Gains	Bins with a gain	Gain words
0	no GAQ	none	none
1	Gk 1 or 2	hebap 8 to 11	1 bit a bin: 0 for 1, 1 for 2
2	Gk 1 or 4	hebap 8 to 16	1 bit a bin: 0 for 1, 1 for 4
3	Gk 1, 2 or 4	hebap 8 to 16	5 bits for three bins: M1 = g / 9, M2 = (g % 9) / 3, M3 = g % 3; 0, 1, 2 mean 1, 2, 4

The gains come first, in bin order, then the mantissas of every bin: nothing for hebap 0, one index for 1 to 7, six values for 8 to 19. With m bits for the hebap:

	Gk = 1 (or no gain)	Gk = 2: small	Gk = 2: large	Gk = 4: small	Gk = 4: large
Bits	m	m − 1	m − 1 more	m − 2	m more
Levels	2^m − 1	2^m−1 − 1	2^m−1	2^m−2 − 1	2^m
Step	2 / (2^m − 1)	1 / (2^m−1)	1 / (2^m−1 − 1)	1 / (2^m−1)	3 / (2^m+1 − 2)

With a gain of 2 or 4, a small codeword equal to the most negative value (100…0) is a tag: a large mantissa follows. A small value is divided by Gk. A large one, and every value coded with Gk = 1, is remapped as y = x + a·x + b, x being the codeword as a signed fraction and a and b 16-bit fractions; b depends on the sign of x.

Remapping constants (Table E.2.6), hexadecimal
hebap	Gk 1: a (b = 0)	Gk 2: a	Gk 2: b, x ≥ 0	Gk 2: b, x < 0	Gk 4: a	Gk 4: b, x ≥ 0	Gk 4: b, x < 0
8	1249	D555	4000	EAAB	EDB7	2000	FB6E
9	0889	C925	4000	D249	E666	2000	ECCD
10	0421	C444	4000	C889	E319	2000	E632
11	0208	C211	4000	C421	E186	2000	E30C
12	0102	C104	4000	C208	E0C2	2000	E183
13	0081	C081	4000	C102	E060	2000	E0C1
14	0040	C040	4000	C081	E030	2000	E060
15	0020	C020	4000	C040	E018	2000	E030
16	0010	C010	4000	C020	E00C	2000	E018
17	0008	no gains
18	0002	no gains
19	0000	no gains

The inverse DCT

The six values X(k, j) of bin k become the bin’s MDCT coefficient in each block m:

C(k, m) = √2 × Σ_j=0…5 R_j × X(k, j) × cos(π(2m + 1)j / 12),   R₀ = 1/√2, R_j = 1 otherwise

then each is scaled by the bin’s exponent and goes on through coupling, rematrixing and the filter bank as in AC-3. (The radicals in this equation are hard to read in the published PDF; the form here is the one FFmpeg’s decoder implements.)

Spectral extension (SPX)

Above a chosen frequency, a channel’s coefficients are not sent but made up: lower coefficients are copied up, blended with noise, and scaled to a coarse envelope the encoder sends. Bins 25 to 228 form 17 sub-bands of 12 bins:

Sub-band	Bins	spxbegf	spxendf	Sub-band	Bins	spxbegf	spxendf
0	25–36			9	133–144	6	3
1	37–48			10	145–156
2	49–60	0		11	157–168	7	4
3	61–72	1		12	169–180
4	73–84	2		13	181–192		5
5	85–96	3	0	14	193–204
6	97–108	4	1	15	205–216		6
7	109–120	5	2	16	217–228
8	121–132			(17)	from 229		7

spx_begin_subbnd = spxbegf < 6 ? spxbegf + 2 : spxbegf × 2 − 3     (first synthesised sub-band)
spx_end_subbnd   = spxendf < 3 ? spxendf + 5 : spxendf × 2 + 3     (one past the last)
copy start       = sub-band spxstrtf (bin 25 + 12 × spxstrtf)

Sub-bands are joined into bands by spxbndstrc, 1 joining a sub-band to the band before. The default structure, used when the first block with spectral extension sends none, joins sub-bands 8, 10, 12, 14 and 16 to the ones before them. With coupling also in use, coupling ends where spectral extension begins (cplendf is then derived from spxbegf: spxbegf − 2, or spxbegf × 2 − 7, and may be negative), and a channel in spectral extension but not in coupling ends at the start bin of spx_begin_subbnd, sending no chbwcod.

Coordinates are a 4-bit exponent and a 2-bit mantissa with an implied leading 1, ranged by mstrspxco (×1, ×2⁻³, ×2⁻⁶, ×2⁻⁹):

spxco = (spxcoexp == 15 ? spxcomant / 4 : (spxcomant + 4) / 8) × 2^−(spxcoexp + 3 × mstrspxco)

For each band the decoder then:

copies coefficients from the copy start upwards into the band, wrapping back to the copy start when it reaches the first synthesised bin (a band that would not fit is started again from the copy start, and marked as a wrap point);
measures the band’s RMS energy;
if the channel has spectral extension attenuation (chinspxatten), applies a five-bin notch centred on the first synthesised bin and on each wrap point, the three values from Table E.2.12 for spxattencod mirrored about the centre (a₀, a₁, a₂, a₁, a₀);
blends with zero-mean, unit-variance noise: with nratio = clamp((first bin of the band + size / 2) / (bin of spx_end_subbnd) − spxblnd / 32, 0, 1), each coefficient becomes tc × √(1 − nratio) + noise × RMS × √nratio;
scales the result by spxco × 32.

Notch attenuations (Table E.2.12), by spxattencod
Code	a₀	a₁	a₂	Code	a₀	a₁	a₂
0	0.954841604	0.911722489	0.870550563	16	0.455861244	0.207809474	0.094732285
1	0.911722489	0.831237896	0.757858283	17	0.435275282	0.189464571	0.082469244
2	0.870550563	0.757858283	0.659753955	18	0.415618948	0.172739110	0.071793647
3	0.831237896	0.690956440	0.574349177	19	0.396850263	0.157490131	0.062500000
4	0.793700526	0.629960525	0.500000000	20	0.378929142	0.143587294	0.054409410
5	0.757858283	0.574349177	0.435275282	21	0.361817309	0.130911765	0.047366143
6	0.723634619	0.523647061	0.378929142	22	0.345478220	0.119355200	0.041234622
7	0.690956440	0.477420802	0.329876978	23	0.329876978	0.108818820	0.035896824
8	0.659753955	0.435275282	0.287174589	24	0.314980262	0.099212566	0.031250000
9	0.629960525	0.396850263	0.250000000	25	0.300756259	0.090454327	0.027204705
10	0.601512518	0.361817309	0.217637641	26	0.287174589	0.082469244	0.023683071
11	0.574349177	0.329876978	0.189464571	27	0.274206245	0.075189065	0.020617311
12	0.548412490	0.300756259	0.164938489	28	0.261823531	0.068551561	0.017948412
13	0.523647061	0.274206245	0.143587294	29	0.250000000	0.062500000	0.015625000
14	0.500000000	0.250000000	0.125000000	30	0.238710401	0.056982656	0.013602353
15	0.477420802	0.227930622	0.108818820	31	0.227930622	0.051952369	0.011841536

Enhanced coupling

Enhanced coupling can start lower than ordinary coupling, from bin 13. Its 22 sub-bands are six bins wide up to bin 36 and twelve above; sub-bands 0 to 7 always stand alone, and ecplbndstrc can join the higher ones, with a default structure (Table E.1.13) that joins sub-bands 9, 11, 13, 15, 16, 17, 19, 20 and 21 to the ones before them.

Sub-band	Bins	ecplbegf	ecplendf	Sub-band	Bins	ecplbegf	ecplendf
0	13–18	0		11	121–132	9	4
1	19–24			12	133–144	10	5
2	25–30	1		13	145–156	11	6
3	31–36			14	157–168	12	7
4	37–48	2		15	169–180		8
5	49–60	3		16	181–192	13	9
6	61–72	4		17	193–204		10
7	73–84	5	0	18	205–216	14	11
8	85–96	6	1	19	217–228		12
9	97–108	7	2	20	229–240	15	13
10	109–120	8	3	21	241–252		14
				(22)	from 253		15

ecpl_begin_subbnd = ecplbegf < 3 ? ecplbegf × 2 : ecplbegf < 13 ? ecplbegf + 2 : ecplbegf × 2 − 10
ecpl_end_subbnd   = ecplendf + 7;  with spectral extension: spxbegf < 6 ? spxbegf + 5 : spxbegf × 2

Each coupled channel has a 5-bit amplitude per band, sent in the first block with enhanced coupling and when ecplparam1e is 1. Codes 0 to 30 are 0 dB down to −45 dB in steps of about 1.5 dB (Table E.2.10: a mantissa 0x20, 0x1B, 0x17, 0x13 or 0x10 over 32, shifted right by 0 to 7), and 31 is silence:

amp = ecplamp == 31 ? 0 : (ecplampmanttab[ecplamp] / 32) × 2^−ecplampexptab[ecplamp]
chmant[ch][bin] = ecplmant[bin] × amp[ch][band of bin]

ecplamp	Exponent	Mantissa	Gain	ecplamp	Exponent	Mantissa	Gain
0	0	0x20	0 dB	16	3	0x10	−24.1 dB
1	0	0x1B	−1.5 dB	17	4	0x1B	−25.6 dB
2	0	0x17	−2.9 dB	18	4	0x17	−27.0 dB
3	0	0x13	−4.5 dB	19	4	0x13	−28.6 dB
4	0	0x10	−6.0 dB	20	4	0x10	−30.1 dB
5	1	0x1B	−7.5 dB	21	5	0x1B	−31.6 dB
6	1	0x17	−8.9 dB	22	5	0x17	−33.0 dB
7	1	0x13	−10.5 dB	23	5	0x13	−34.6 dB
8	1	0x10	−12.0 dB	24	5	0x10	−36.1 dB
9	2	0x1B	−13.5 dB	25	6	0x1B	−37.6 dB
10	2	0x17	−14.9 dB	26	6	0x17	−39.0 dB
11	2	0x13	−16.6 dB	27	6	0x13	−40.7 dB
12	2	0x10	−18.1 dB	28	6	0x10	−42.1 dB
13	3	0x1B	−19.5 dB	29	7	0x1B	−43.6 dB
14	3	0x17	−20.9 dB	30	7	0x17	−45.0 dB
15	3	0x13	−22.6 dB	31	–	0x00	silence

Every channel but the first coupled one also has a rsvdfieldse bit; when it is 1, 9 × (bands − 1) reserved bits follow, and each such channel ends with one more reserved bit. TS 102 366 defines only the amplitudes; whatever these reserved fields carry is not public, and a decoder skips them.

Transient pre-noise processing

Coarse quantisation smears a sharp attack backwards in time, as noise before it. The encoder finds the transient and sends, per channel and frame (chintransproc), its position transprocloc (in units of four samples from the frame’s first output sample, possibly in a later frame) and a length transproclen in samples. The decoder overwrites the noisy stretch with a copy of earlier, clean audio:

TC1 = 256, TC2 = 128
transloc = 4 × transprocloc;  pnlen = transloc − (start of the block before the transient’s block)
total = pnlen + transproclen + TC1
synth[s] = pcm[transloc − (2 × TC1 + 2 × pnlen) + s]          s = 0 … 2 × TC1 + pnlen − 1
from transloc − total: cross-fade from the original to synth over TC1 samples,
                       then synth, then cross-fade back over the last TC2 samples before transloc

Any pair of constant-amplitude cross-fade windows will do; Hann windows work well.

.AC3 File Format – E-AC-3 coding tools