oxideav-aac

Pure-Rust AAC-LC (MPEG-4 Audio / ISO/IEC 14496-3 Object Type 2) decoder and encoder. ADTS framing, Huffman codebooks 1-11, IMDCT, M/S stereo, TNS, PNS, and pulse data. No C dependencies, no FFI, no *-sys crates.

Part of the oxideav framework but usable standalone.

Installation

[dependencies]
oxideav-core = "0.1"
oxideav-codec = "0.1"
oxideav-aac = "0.0"

Quick use

The crate registers a single codec id, "aac", that maps to both the decoder and the encoder.

use oxideav_core::{CodecId, CodecParameters, RuntimeContext};

let mut ctx = RuntimeContext::new();
oxideav_aac::register(&mut ctx);

// Decode an ADTS stream
let mut params = CodecParameters::audio(CodecId::new("aac"));
params.sample_rate = Some(44_100);
params.channels = Some(2);
let mut dec = ctx.codecs.make_decoder(&params)?;
// dec.send_packet(&adts_packet)?;
// dec.receive_frame()?  -> Frame::Audio (interleaved S16, FRAME_LEN=1024)

// Encode S16 PCM into ADTS
let mut enc_params = CodecParameters::audio(CodecId::new("aac"));
enc_params.sample_rate = Some(44_100);
enc_params.channels = Some(1);
enc_params.bit_rate = Some(128_000);
let mut enc = codecs.make_encoder(&enc_params)?;
// enc.send_frame(&audio_frame_s16)?;
// enc.receive_packet()?  -> ADTS-framed AAC frame
# Ok::<(), oxideav_core::Error>(())

For raw raw_data_block() payloads coming out of an MP4 demuxer, set params.extradata to the AudioSpecificConfig (ASC) blob; the decoder will read it instead of looking for an ADTS sync word on the first packet.

Decode support

Feature	Status
Object types	AAC-LC (AOT 2) decode + encode; AAC-LD (AOT 23) / AAC-ELD (AOT 39) ASC parse only (frame decode deferred)
Containers	ADTS (with or without CRC) and raw + ASC
Channel configurations	1..=7 (mono, stereo, 3.0, 4.0, 5.0, 5.1, 7.1)
Sample rates	All 13 standard SF indices (96k - 7350)
Window sequences	Long, LongStart, LongStop, EightShort
Window shapes	sine and KBD
Huffman spectral books 1-11	Yes (escape book 11 included)
Scalefactor Huffman book	Yes
M/S stereo (§4.6.13)	Yes (long + short)
PNS / Perceptual Noise Sub. (§4.6.13)	Yes (long + short, correlated noise)
Intensity stereo (§4.6.8.2.3)	Yes (cb 14/15, sign from ms_used)
TNS (§4.6.9)	Long AND short-window filters
Pulse data (§4.6.5)	Yes (long-window only); short-window rejects as non-conformant
Fill / DSE elements	Skipped cleanly; FIL count==15, esc_count==0 off-by-one fixed (v0.1.1)
LD/ELD AudioSpecificConfig	asc.ld_config / asc.eld_config populated for AOT 23/39; SWB_LD_512 / SWB_LD_480 tables in ld_eld module
LD/ELD filterbank kernels	480- and 512-sample MDCT/IMDCT (crate::mdct::mdct_ld_512 / crate::imdct::imdct_ld_480) + sine half-windows (crate::window::sine_ld_512 / sine_ld_480) + crate::ld_eld::imdct_and_overlap_ld overlap-add filterbank with LdChannelState. Two-frame TDAC round-trip <5e-3 max err on sine input at both frame sizes. ER raw_data_block decoder pending.
USAC / xHE-AAC AudioSpecificConfig	asc.usac_config populated for AOT 42; crate::usac::parse_usac_config captures sample rate (incl. 24-bit explicit escape), coreSbrFrameLengthIndex, channelConfigurationIndex, and the first usacElementType (SCE/CPE/LFE/Ext, ISO/IEC 23003-3 Table 9). Frame decode pending.
LFE element (§4.6.10)	Yes (long-window SCE-like path)
PCE (Program Config Element)	Parsed (channel mapping reserved for future use)
Gain control / SSR / Main / LTP	Refused (Error::Unsupported)
CCE elements	Refused (Error::Unsupported)
HE-AAC v1 (SBR) decode (mono + CPE)	Yes (independent + coupled CPE)
HE-AAC v2 (PS) decode	Yes (QMF-domain upmix incl. IPD/OPD)

The decoder advertises max_channels = 8 and max_sample_rate = 96_000 in CodecCapabilities. PCM output is interleaved in AAC element order (C, L, R, Ls, Rs, LFE for 5.1) — downstream muxers may remap.

Encode support

Feature	Status
Object types	AAC-LC (AOT 2)
Containers	ADTS only (one raw_data_block per frame, no CRC)
Channels	1..=7 (1, 2, 3, 4, 5, 6, 8 input channels)
Element orchestration	SCE + CPE + LFE sequence per AAC channel_configuration §1.6.3
Sample rates	Any of the 13 standard SF indices that match a known SWB table; tested at 44.1 kHz / 48 kHz
Input sample format	S16 and F32 interleaved
Window sequence	Long-only by default (short-block toolkit in place, state-machine wiring pending)
Window shape	Sine
MDCT scaling	Matches ffmpeg's aacenc.c 32768x convention
Spectral codebook selection	Per-band cheapest of books 1-11 (incl. escape)
Section data	Run-length compressed; merges adjacent same-cb bands
Scalefactors	Huffman-coded deltas with global_gain anchor; 3-accumulator path (g_gain / g_noise / g_is) for NOISE / IS bands
M/S stereo (§4.6.13)	Per-band L/R-vs-M/S decision by bit cost + activity gate (energy-balance ∈ [1/8, 8] AND |corr| ≥ 0.4) + magnitude-weighted sign-agreement gate (≥ 55 % per-line polarity agreement); blocks M/S on partially anti-phased bands
TNS (§4.6.9)	LPC analysis on SCE long blocks; 4-bit parcor quantisation; adaptive filter order 2–8 per Spectral Flatness Measure (SFM); adaptive gain threshold 1.38–1.80 (raises on noise-like bands to suppress low-value parcor spend)
PNS encode (§4.6.12)	Yes (long windows; peak-to-RMS ≤ 2.6 + SFM ≥ 0.25 noise gate, ≥ 4 kHz band-centre gate; trimmed-mean energy gain matches source RMS within ±1 dB)
Intensity stereo encode (§4.6.8.1.4)	Yes (long windows; |corr| ≥ 0.95 + per-line sign-agreement ≥ 80 % + energy ratio ∈ [1/256, 256] in CPE common-window path; ≥ 4 kHz band-centre gate; corpus PSNR delta +1.7 dB on aac-lc-intensity-stereo after round-#523 tuning)
Pulse data encode (§4.6.10)	Yes (up to 4 per frame; sign-preserving outlier extraction, amp capped at |residual| - 1)
Short blocks (building blocks)	TransientDetector, mdct_short_eightshort, analyse_and_quantise_short, write_single_ics_short — all tested; emit_block state-machine integration pending
HE-AACv1 (SBR) encode	Mono (HeAacMonoEncoder, psy-on default); stereo CPE (HeAacStereoEncoder, independent coupling, §4.6.18.3.5, psy-on default as of round-27 M/S CPE side-lobe fix); v2 (HeAacV2Encoder, mono SCE + PS, psy-on default); SBR noise-floor scalefactor derived from per-band QMF energy (tonal → sf 18, noise-dominated → sf 10)
Gapless playback metadata	gapless::GaplessInfo triple (delay/padding/valid_samples) + Apple iTunSMPB-format string emitter; AAC-LC reports 2112-sample priming, HE-AAC reports 2624 (high rate). End-of-file padding rounded to the next packet boundary so an MP4 edts/elst writer or ID3v2 TXXX:iTunSMPB wrapper can round-trip the source PCM sample-accurately.
Psychoacoustic model	Bark-band PE/SMR allocator (psy::PsyModel, default-on for AAC-LC + every HE-AAC variant); per-band tonality-driven target_max, sub-baseline coarsening gated on tonality < 0.15 (avoids +17 % bytes on noise-only fixtures). Tonality-adaptive self-masking: -10 dB for noise bands, -16 dB for tonal bands (ISO 11172-3 Annex D). Above-baseline target_max clamped to baseline=7 in the CPE LR/MS path (TNS off). Override via AacEncoder::set_enable_psy_model or env OXIDEAV_AAC_PSY_MODEL=0. Corpus-validated to within +0.08 dB mean PSNR / -0.42 dB worst; HE-AAC corpus-validated mean +1.92 dB / worst +0.07 dB.
Gain control	Not implemented
CBR / VBR	Bit-reservoir CBR allocator (AacEncoder::set_cbr_target_bitrate) drives a per-frame scalefactor bias from a 6144-bit reservoir to bound output frames at the configured rate; default off (VBR)

The encoder advertises max_channels = 8 and max_sample_rate = 48_000. Multi-channel output emits elements in AAC element order (C, L, R for 3.0; C, L, R, Ls, Rs, LFE for 5.1; etc.) — round-trip validated via the self-decoder for 5.1 and 7.1 layouts in tests/encode_roundtrip.rs. TNS on stereo (CPE) is gated off until per-band M/S decisions can run on TNS-flattened coefficients.

Round-trip verification

tests/encode_roundtrip.rs runs the encoder end-to-end for:

• 44.1 kHz mono sine through our own decoder + ffmpeg (Goertzel ratio >= 50x)
• 44.1 kHz stereo sine through our own decoder + ffmpeg (both channels)
• 48 kHz mono sine through ffmpeg
• 0.5 s mono and stereo silence through our own decoder (RMS < 1e-3)
• 44.1 kHz 5.1 sine-per-channel through our own decoder (each of the 6 channels recovers its tone above a 20x Goertzel floor)
• 44.1 kHz 7.1 sine-per-channel through our own decoder (all 8 channels)
• 44.1 kHz 5.1 sine-per-channel through ffmpeg (encode_51_roundtrip_ffmpeg). All 6 channels survive ffmpeg's bitstream→WAVE-5.1 reorder (C/L/R/Ls/Rs/LFE → L/R/C/LFE/Ls/Rs); per-channel PSNR floors at 20 dB (matching the AC-3 5.1 acceptance pattern), with five of six channels clearing 25 dB on the synthetic tone fixture. The L/R-CPE R channel running an octave-paired tone (R = 880 Hz vs L = 440 Hz) sits at ~22 dB because M/S coding biases bit allocation toward the side signal.
• 44.1 kHz 7.1 sine-per-channel through ffmpeg (encode_71_roundtrip_ffmpeg, task #154). All 8 channels survive ffmpeg's bitstream→WAVE-7.1 reorder (AAC bitstream order C/L/R/Ls/Rs/Lb/Rb/LFE per §1.6.3 Table 1.19 → WAVE FL/FR/FC/LFE/BL/BR/SL/SR, inverse mapping [2, 0, 1, 6, 7, 4, 5, 3]); per-channel PSNR floors at 22 dB, with seven of eight channels clearing 24 dB on the synthetic tone fixture. The L/R-CPE R channel hits the M/S-bias floor for the same reason as 5.1.

r19 — AAC-LC ffmpeg-interop RMS audit (2026-04-26)

tests/lc_rms_interop_r19.rs (new) cross-checks all four directions:

direction	RMS	ratio
ours-encode → ours-decode	6718	0.97x
ours-encode → ffmpeg-decode	6644	0.96x
ffmpeg-encode → ours-decode	6881	0.99x
ffmpeg-encode → ffmpeg-decode (ref)	6950	1.00x

Test signal: 1 s of 440 Hz sine at amplitude 0.3 through 44.1 kHz mono; expected RMS = 0.3 * 32767 / sqrt(2) = 6951. All four within ±5% of unity — the AAC-LC core spectrum scale matches ffmpeg end-to-end on the informationally-correct (RMS) metric. Audited and confirmed spec-correct:

• §4.6.1.3 inverse quant sign(q) * |q|^(4/3) (matches ics::inv_quant).
• §4.6.2.3.3 scalefactor gain 2^(0.25 * (sf - 100)), SF_OFFSET = 100 (matches ics::sf_to_gain).
• §4.6.11.3.1 IMDCT scale 2/N with N = 2 * input_n — our IMDCT uses 2/input_n (= 4/N, doubled) and the decoder's S16 output stage multiplies by 0.5, giving an effective 2/N matching the spec.
• §4.5.2.3.6 "the integer part of the output of the IMDCT can be used directly as a 16-bit PCM audio output" — encoder forward-MDCT scale derivation: 2 * 32768 = 65 536, encoded in MDCT_FORWARD_SCALE.

The previously-claimed "~3.33x mid-stream amplitude gap" (rounds 17/18) was a peak-metric artefact: ffmpeg's encoder fills HF bands with PNS (codebook 13, §4.6.13) that our spec-compliant decoder reconstructs as additive noise. The PNS noise rides on the sine peak, inflating the peak ratio (1.79x for ffmpeg → ours) while the RMS stays at 0.99x. PNS is non-deterministic per-frame, so peak ratio is not a meaningful interop metric for tonal+noise content; RMS is.

examples/probe_lc_amp.rs and examples/spectrum_compare.rs are the diagnostic probes used to land this audit; both report RMS alongside peak so subsequent rounds don't re-chase the same phantom.

Note: tests/sbr_he_aac_ffmpeg_amplitude_r18.rs (HE-AACv1 SBR amplitude saturation at peak 32_768) remains ignored.

Round 21 audit data (1 kHz tone amp 0.3 mono SCE at 48 kHz):

input (peak / RMS)              :  9 830 / 6 951
ours-encode -> ours-decode       : 10 256 / 6 582  (within 5% of input)
ours-encode -> ffmpeg-decode     : 32 767 / 17 181 (saturated)
fdkaac-encode -> ffmpeg-decode   :  9 822 / 6 920  (within 1% of input)
fdkaac-encode -> ours-decode     : 13 009 / 7 061  (within 30% of input)

Round-21 probe of the bitstream-level envelope value confirms both fdkaac and our encoder transmit bs_data_env[0] = 0 for tonal content with no high-band energy (E_orig = 64, the spec minimum). The bitstream-level envelope value is therefore not the source of the saturation — both encoders emit the same value yet ffmpeg decodes the fdkaac stream cleanly while saturating ours. Audit of §4.6.18.4.2 (synthesis QMF gain 1/64), §4.6.18.7.1 (E_orig formula), §4.6.18.7.5 (limiter gain cap, Table 4.176), and the HF generator patches all confirms spec-correctness; the divergence must arise from a difference in how ffmpeg reads our specific header configuration vs fdkaac's. Round 22+ target.

r22 — HE-AACv1 SBR header diff-probe (2026-04-26)

Round 22 ran the methodical SBR header diff-probe specified in the round-21 brief. Result: no SBR header field is the saturation source. The fdkaac vs ours header values for a 1 kHz / 0.3-amp mono SCE at 48 kHz output (24 kHz core) were captured via our own SBR header parser:

field              fdkaac  ours
bs_amp_res         1       0       (effective per-frame value: both 0,
                                    forced to 0 for FIXFIX num_env=1)
bs_start_freq      13      5       (k0 = 22 vs k0 = 12)
bs_stop_freq       11      9
bs_xover_band      0       0       (same)
bs_freq_scale      1       2       (8 bands/octave vs 10)
bs_alter_scale     1       1       (same)
bs_noise_bands     2       2       (same)
bs_limiter_bands   2       2       (same)
bs_limiter_gains   2       2       (same)
bs_interpol_freq   1       1       (same)
bs_smoothing_mode  1       1       (same)

Forcing each differing field individually (bs_amp_res=1, bs_start_freq=13, bs_stop_freq=11, bs_freq_scale=1) and the combined fdkaac configuration all still produced ffmpeg peak 32 768 (full saturation). Forcing bs_stop_freq=11 and bs_freq_scale=1 with our bs_start_freq=5 triggers ffmpeg "Invalid bitstream, too many QMF subbands: 41" / "Invalid vDk0[1]: 0" — these are spec-consistent with the freq-table derivation (§4.6.18.3.2.1) and rule out the bs_start_freq / bs_stop_freq pair as a meaningful encoder-side fix.

The differential probe further extracted the per-band envelope and noise scalefactor data:

fdkaac: env[0] sf(14b) = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
        noise[0] sf(2b) = [14, 0]
ours:   env[0] sf(16b) = [29, -1, -1, -1, 0, -1, -1, 0, -1, -1, ...]
        noise[0] sf(4b) = [18, 0, 0, 0]

Setting our encoder's INT16_SCALE_SQ to 1.0 reproduces fdkaac's all-zero envelope (E_orig = 64, the spec minimum) — yet ffmpeg still saturates ours to peak 32 768. Even completely omitting the SBR FIL extension (pure AAC-LC core only, decoded as 24 kHz mono) still produces ffmpeg peak 32 768 / RMS 30 961, vs our own decoder's clean 10 683 / RMS ~6 000 on the identical stream.

Round 23 (2026-04-30) audit: r22 MDCT-scale thesis refuted

Round 23 set MDCT_FORWARD_SCALE = 4 096 and re-ran the matrix:

• LC RMS interop test (44.1 kHz / 440 Hz): ours-encode → ours-decode ratio collapsed to 0.060 (16× too quiet). The test fails on the very first assertion. There is no ±5 % window that can absorb a 16× reduction.
• HE-AAC SBR amplitude (the r18 ignored test): peak only drops from 32 768 (saturated) to 25 287. Still saturated. The 16× scale change does not unblock HE-AAC interop.
• Pure AAC-LC mono at 24 kHz / 1 kHz with current SCALE = 65 536: ffmpeg-decoded peak 10 930 / RMS 6 955, within ±5 % of the input. Pure stereo LC at 24 kHz / (1k+2k) produces L=10 930/9 880, RMS 6 955/6 579 — also within ±5 % per channel. r22's claim that pure-LC-at-24 kHz saturates was wrong: it conflated the HE-AAC code path (which carries an SBR FIL extension) with pure-LC. New regression tests (r23_lc_24khz_probe.rs + r23_he_aac_isolation.rs) pin both cases at the spec-correct unity-RMS measurement.

Sweep on HE-AAC stereo with current SCALE values:

SCALE   L-peak  L-rms   R-peak  R-rms   verdict
65536   15767   6544    32768   30461   sat (R)
32768    7884   3272    32768   28126   sat (R)
16384    3942   1636    32768   23068   sat (R)
 8192    1972    818    32768   14711   sat (R)
 4096     986    409    25287    7563   passing-through
 2048     493    204    12643    3782   silent
 1024     247    102     6321    1891   silent

The L-peak strictly halves with SCALE (linear pass-through through the LC core). The R-peak stays clipped at 32 768 until the input drops below the clipping threshold, then itself halves. r22's "RMS = 6 951 at SCALE = 4 096 → unity" reading was a methodological error: a clipped square-wave at 32 768 has RMS ≈ 30 000, and reducing SCALE 16× simply lowers input below the clipping threshold — RMS passes through the input target on its way to silence (verified: SCALE = 2 048 → 1 891, SCALE = 1 024 → 947). There is no stable interop point in the sweep.

Conclusion (r23): MDCT_FORWARD_SCALE = 65 536 is the correct value (verified by 4 LC-only ffmpeg interop measurements at both 44.1 kHz and 24 kHz, mono and stereo) and remains in place. The HE-AAC ffmpeg-interop saturation lives in the SBR FIL extension itself (most likely bs_invf_mode, bs_add_harmonic, or an HF-generation-stage gain), not the LC core. Pinned for round 24.

Round 24 (2026-04-30): SBR FIL diff vs fdkaac

Round 24 built tests/r24_sbr_fil_diff.rs: encodes the r18 fixture (1 kHz / amp 0.3 / 0.5 s mono, 48 kHz, 48 kbps HE-AAC) through both our HeAacMonoEncoder and fdkaac -p 5 -f 2, parses every ADTS frame's SBR FIL element via the same oxideav_aac::sbr::bitstream::parse_* routines our decoder uses, and diffs the resulting SbrChannelData field-by-field. The harness walks the SCE element bit-cursor through the now-pub decoder::decode_ics + decoder::fill_spectrum so the FIL bit position is exact (no brute-force bit-offset scanning).

Field	ours	fdkaac
bs_amp_res (header)	0	1
bs_start_freq	5	13
bs_stop_freq	9	11
bs_freq_scale	2	1
derived n_high / nq	16 / 4	14 / 2
FIL payload bits / frame	96	80
bs_invf_mode totals	[0,0,0,0,0]	[5,2,0,0,0] (varies)
bs_add_harmonic_flag	0/12 frames	1/15 frames
bs_df_env totals	[0,0,0,0,0]	[2,4,2,1,0] (uses time-dir)
bs_df_noise totals	[0,0]	[9,2] (uses time-dir)
frame[0] env_sf[0]	[29,-1,-1,...]	[0,0,0,...] (E_orig minimum)
frame[0] noise_sf[0]	[18,0,0,0]	[14,0]

Refuted thesis — the envelope value is NOT the saturation source. The diff first looked like our env_sf[0][0] = 29 (= 64 · 2^14.5 ≈ 1.5 M; QMF analysis-bank skirt leakage of the 1 kHz tone into the bottom-most SBR subband, amplified by INT16_SCALE_SQ = 2^30) was the cause. To pin this we added the OXIDEAV_AAC_SBR_ENV_FORCE_ZERO env-var probe in sbr/encode.rs::estimate_envelope: when set, every band gets value 0 (matching fdkaac's "no high-band content" output) and the noise floor gets value 14. Re-ran ffmpeg-decode of the r18 mono fixture under the override:

mono HE-AAC, no override:    peak=32768  rms=28739
mono HE-AAC, FORCE_ZERO=1:   peak=32768  rms=28739  (identical)

The override is verified to actually zero the envelope on every frame (force_zero_env_var_actually_zeros_envelope regression). Same saturation reading, byte-identical decode artefacts. The envelope is fully ruled out.

ffmpeg consistently logs No quantized data read for sbr_dequant on every decode, regardless of envelope value. This warning fires before any envelope arithmetic, suggesting ffmpeg's parser is giving up on our SBR data structurally — likely either:

1. Header-vs-grid bs_amp_res mismatch: we send bs_amp_res = 0 in the header but rely on the FIXFIX bs_num_env == 1 rule (§4.6.18.3.3) to override to 0 inside parse_sbr_grid. fdkaac sends bs_amp_res = 1 in the header and relies on the same override. ffmpeg may be reading the start-value bit count from the raw header value before the override fires.
2. bs_extended_data = 0 framing edge-case: our SCE always ends with bs_extended_data = 0 (1 bit) followed by zero fill bits to the FIL byte boundary. fdkaac's SCE is shorter (53 bits vs our 63), suggesting its grid + envelope shapes use fewer bits, but both end byte-aligned. The 4-bit extension_payload(cnt) count alignment may differ in some subtle way.
3. Time-vs-freq delta encoding: fdkaac uses time-direction delta on noise (bs_df_noise totals 9/2) and on envelope (2/4) while ours always uses freq-direction. Time-direction requires a previous-frame baseline; freq-direction encodes the first band as an absolute. ffmpeg may not accept freq-direction on the very first frame of a HE-AAC stream.

The diff harness is permanent infrastructure now; r25 can change any header field or encoding strategy and re-run the diff to verify convergence with fdkaac.

ffmpeg-dependent tests skip cleanly when ffmpeg is not on PATH. tests/encode_tns.rs confirms the encoder emits TNS on transient content and that TNS-bearing frames decode without error. tests/encode_pns_is_pulse.rs verifies:

• PNS fires on >=75% of >=4 kHz bands for white-noise input and round-trips within a factor of 4 in total RMS energy.
• Intensity stereo fires on >=2 HF bands for a stereo clip with a quiet correlated R channel; decoded R tracks L's sign.
• Pulse data is emitted on at least one frame of a loud 440 Hz tone and the round-trip Goertzel ratio stays >=50x.

tests/encode_pns_savings.rs (task #132) pins the bit-savings PNS buys on noise-rich content (cymbals + sax-like harmonic stack + room-tone broadband background, 1 s mono at 44.1 kHz / 96 kbps):

• PNS-active vs PNS-disabled (OXIDEAV_AAC_DISABLE_PNS=1) A/B encode: raw_data_block bytes drop 63.9% (8 749 B vs 24 256 B).
• Self-decoder RMS round-trip ratio: 0.977 (PNS preserves band energy within 2.3% of the input).
• ffmpeg cross-decode runs clean (no warnings, no errors). The ffmpeg-decoder RMS ratio of 6.06× is the same FAAD2-vs-ffmpeg dpcm_noise_nrg calibration delta documented in the r19 audit; not blocking the cross-decode-clean criterion.

Codec id

• "aac" — both encoder and decoder.

Frames are produced as interleaved SampleFormat::S16 at the stream's sample rate, 1024 samples per frame.

Caveats

• HE-AAC (SBR / PS) decode is supported (mono + stereo CPE, optional parametric stereo upmix). HE-AAC encode is mono + stereo only — no PS encoder yet, no SBR support for ≥3-channel CPE configurations.
• Bit_rate on the encoder is informational. The encoder picks scalefactors based on a fixed target quantisation magnitude; emitted frame size depends on signal complexity, not directly on the requested rate.
• The make_encoder registry path uses long blocks unconditionally. Sharp attacks pre-echo more than they would with proper short-block switching.
• MP4-side AAC streams need the AudioSpecificConfig blob in CodecParameters::extradata; otherwise the first packet must carry an ADTS sync word.

License

MIT - see LICENSE.