oxideav-aac

A pure-Rust AAC (Advanced Audio Coding) codec for the oxideav framework.

Every numeric constant, bit layout, and clause reference is sourced from the staged ISO/IEC 13818-7 and ISO/IEC 14496-3 specifications under docs/audio/aac/.

Status

The crate implements the full AAC-LC decode chain end to end — from ADTS bitstream parse through the per-tool reconstruction to interleaved 16-bit PCM — plus the complete §4.6.18 SBR back-end (HE-AAC v1) and the subpart-8 Parametric Stereo tool (HE-AAC v2), and wires them into the framework's runtime Decoder trait (register() installs an AAC decoder under id "aac"; see codec_decoder below). The PCM is validated byte-exactly (within the 1-LSB IMDCT-rounding bound) against the staged expected.wav corpus — including the HE-AAC v1 SBR fixture, which decodes 99.98% sample-exact with a max error of 1 LSB at the doubled output rate, and the HE-AAC v2 fixture, whose PS stereo reconstruction lands at a 5e-5 per-channel error-to-signal RMS against the reference decode. The §4.6.18.4.3 downsampled output mode (core-rate SBR, auto-selected from a core-rate extensionSamplingFrequency ASC) and the §4.6.18.8 low power SBR tool (real-valued filterbanks + aliasing detection/reduction) are selectable on every decode entry point. The crate also ships an end-to-end AAC-LC encoder (encoder / codec_encoder): PCM → §4.6.11.3.1 forward-MDCT analysis with §4.6.11.3.2 block switching (transient-driven ONLY_LONG → LONG_START → EIGHT_SHORT → LONG_STOP, short frames grouped per §4.5.2.3.4 on the band-envelope similarity of adjacent windows), the exact §4.6.2 inverse quantizer under a masking-spread psychoacoustics-lite model with a bidirectional rate loop, measured-bit-cost codebook/section choice (a DP over section boundaries priced with the real tuple writer), per-band §4.6.8.1 M/S joint stereo (long frames per sfb, short frames per (window group, sfb) under the pair's joint grouping), and every Table 1.19 default channel layout — 1–6 and 8 (7.1) channels as SCE / common_window-CPE / §4.5.2.1.3-conforming LFE element plans — assembled into ADTS through the Phase-2 bit-exact wire writers. Every stream is round-tripped through the crate's own decoder (multitone 128 kbps at 0.016 err/sig RMS; staged-fixture transcodes at 0.0008–0.003; multichannel layouts pinned with one distinct tone per speaker); register() installs the encoder alongside the decoder under id "aac". On top of the LC encoder the crate now ships an HE-AAC v1 (SBR) encoder (he_aac_encoder / sbr_encoder): a crossover-cut downsampler feeds the LC core at half rate while the Annex 4.B.18 parameter estimator writes one EXT_SBR_DATA(_CRC) fill payload per frame — implicit ADTS signalling plus both explicit AudioSpecificConfig forms and a LOAS/LATM writer; the registry factory selects the profile from an SBR-signalling extradata ASC (make_he_aac_encoder is the direct path).

ISO/IEC 14496-26 conformance (normative corpus)

tests/iso_14496_26_conformance.rs decodes members of the normative MPEG-4 Audio conformance corpus end to end against their reference waveforms (corpus located via OXIDEAV_ISO_14496_26_DIR, skip-if-absent; sourcing, per-member checksums and the member-level fetch recipe are in docs/audio/aac/iso-14496-26-conformance.md — the ISO-copyright bitstreams are never committed). Measured state:

• ER AAC LD — 15 vectors across 22.05/24/32/44.1/48 kHz at both frame lengths: 47 003 / 47 004 access units decode (the single residual is er_ad1103_22_ep0 AU 367, which the staged corpus screen records as failing under every width hypothesis). PCM: the LD-512 er_ad1000* family is reference-exact at err/sig ≈ 4.4e-5; the LD-480 er_ad1103np* family lands at ≈ 1.3e-4 outside its TNS/PNS access units (PNS noise phase is generator-defined; the deployed LD TNS record is a still-untraced extra-spec wire — see "Not yet supported"). The 32 kHz members pin the §4.5.4 Tables 4.144/4.145 band tables end to end.
• CCE — the twelve am05_* vectors (AAC Main + one coupling_channel_element() in every access unit): 1 370 / 1 370 AUs decode, and all six am05_48 output channels match their per-speaker references at ≈ 1e-4 err/sig — pinning the CCE gain path (conformance-settled cc_scale^(−ge) exponent), the §4.6.6 Main-profile predictor at its normative fixed-precision arithmetic, M/S + intensity + TNS interplay, and the §8.5.2.2 PCE reorder.
• SBR-CRC — the four al_sbr_{e,i}_32_* vectors (the corpus's only EXT_SBR_DATA_CRC carriers): 1 600 / 1 600 payload CRCs verify during a full decode, including the §4.5.2.8.1 pre-header prefix (upsampling-only state) and the whole-payload coverage region (bs_fill_bits included).

Bitstream parsing

• ADTS fixed header (adts) — ISO/IEC 13818-7 §1.A.2: sync, profile, sampling-frequency index, channel configuration, frame length, raw-data-block count, CRC presence flag.
• ADTS error_check() + SBR CRC verification (adts_crc) — the ISO/IEC 13818-7:2004 §8.1.1.1 protected-bit region walk (all 56 header bits; the first 192 bits of every SCE / CPE / CCE / LFE with the 3-bit id_syn_ele excluded and zero-padding of short elements; the additional first-128-bits of every CPE's second individual_channel_stream; all PCE / DSE bits) fed into the ISO/IEC 11172-3 §2.4.3.1 CRC-16 (0x8005, all-ones init) that 13818-7 §8.1.1.2 cites. Both frame forms verify: the Table 1.A.8 single-raw-data-block crc_check and the Table 1.A.9 / 1.A.10 multi-RDB split (headers + raw_data_block_position table under one CRC, one CRC per block). Wired into StreamDecoder::decode_adts_frame / decode_all and the runtime Decoder; protect_adts_frame / protect_adts_stream produce the protected form (a protected rewrite of a staged fixture decodes byte-identically through a black-box validator binary — which, notably, does not verify the CRC value, so the code convention is pinned to the documented §2.4.3.1 parameters). The same module hosts the SBR bs_sbr_crc_bits CRC-10 (G10, zero init) computed over the Table 4.62 coverage region; the FIL walk verifies every EXT_SBR_DATA_CRC payload. Corruption of any covered bit surfaces Error::AdtsCrcMismatch / Error::SbrCrcMismatch; fill bits and the beyond-window element bits are provably uncovered.
• AudioSpecificConfig (asc) — ISO/IEC 14496-3 §1.6.2.1 including the §4.4.1 GASpecificConfig body for all General Audio object types, the hierarchical SBR (AOT 5) / PS (AOT 29) wrappers, the extensionFlag subtree, the epConfig field, and the Table 1.15 trailing syncExtensionType == 0x2b7 implicit-SBR probe. A carrier-bounded parse_bits_bounded entry point is exposed for LATM StreamMuxConfig callers.
• program_config_element (pce) — §4.4.1.1, used standalone and inline inside asc.
• raw_data_block() walker (raw_data_block) — §4.4.2.1: visits each id_syn_ele and stops at END. FIL / DSE / PCE bodies are fully consumed; the channel-element body is composed by the modules below.
• Channel-element body (ics_body) — Table 4.50: global_gain → ics_info → section_data → scale_factor_data → optional pulse_data / tns_data / gain_control_data, surfacing the start bit-offset for the spectral data.
• spectral_data() (spectral_data) — Table 4.56 wire walker and bit-exact writer, dispatching onto the Huffman codebooks.
• extension_payload() (extension_payload) — §4.4.2.7 / Table 4.51 parser + encoder for the EXT_FILL, EXT_FILL_DATA, and EXT_DYNAMIC_RANGE branches. The two SBR-data extension types decode through the parse_with_sbr entry (feeding the §4.6.18 back-end); the plain parse entry without an SBR context rejects them (Error::UnsupportedExtensionSbr).
• Error-protection CRC generator (crc) — §1.8.4.5: the full family of MPEG-4 Audio CRC generation polynomials (CRC4..CRC32, including the CRC8 LATM StreamMuxConfig() crcCheckSum and the 16-bit x¹⁶+x¹⁵+x²+1), a zero-init MSB-first shift-register (crc_bits / crc_bytes) implementing the §1.8.4.5 M(x)·xᵏ = Q(x)·G(x) + R(x) remainder with the normative output-bit inversion ("written in a reversed manner, i.e. each bit is inverted"), and the [crc::stream_mux_config_crc] LATM helper. Cross-checked against an independent GF(2) long-division reference and the codeword-divisibility property. The ADTS adts_error_check() region-selection CRC uses a different code convention (ISO/IEC 11172-3 §2.4.3.1, all-ones init, no output inversion) and lives in the dedicated adts_crc module above.
• RVLC error-resilient scalefactor coding (rvlc, scale_factor_data::ErScaleFactorData) — §4.6.16.2 the reversible-variable-length-coding replacement for the §4.6.3 noiseless coding of scalefactors, used when aacScalefactorDataResilienceFlag == 1. The rvlc module transcribes the symmetric (bit-palindrome) RVLC codebook (Table 4.166, deltas -7..=+7 with ±7 the ESC_FLAG), the eight asymmetric forbidden codewords (Table 4.167) whose appearance is surfaced as the §4.6.16.2.1 in-band error-detection event, and the 54-entry RVLC-ESC Huffman codebook (Table 4.168) — every codebook proven prefix-free and round-tripping, and independently cross-validated against the staged packed binary-tree node tables. ErScaleFactorData::parse / ::write decode and re-encode the whole Table 4.53 RVLC branch: the sf_concealment / rev_global_gain / length_of_rvlc_sf (11 bits for EIGHT_SHORT_SEQUENCE, else 9) header, the RVLC base-delta band loop (first PNS band keeping the 9-bit PCM seed), the optional sf_escapes_present / length_of_rvlc_escapes second pass folding each escape into its ±ESC_FLAG base (+7 + esc / -7 - esc per §4.6.16.2.1), and the dpcm_is_last_position / dpcm_noise_last_position backward seeds. Both length_of_* fields are validated against the bits actually consumed. The reconstructed records share the non-resilient ScaleFactorData shape, so the §4.6.2.3.2 forward DPCM accumulate() pass consumes them unchanged — pinned by a test that an RVLC stream and the Huffman stream carrying the same deltas accumulate to identical absolute scalefactors. The resilience-flag dispatch from ics_body is now wired (see the error-resilient ICS body below); the RVLC bitstream path itself is decoded end to end.
• Error-resilient channel-element body (ics_body::IcsBody::parse_er / ::parse_with_ics_info_er, section_data::SectionData::parse_er / ::write_er) — ISO/IEC 14496-3 §4.4.6 Tables 4.50 / 4.52, the ER General-Audio object types (AOTs 17 / 19 / 20 / 23). Drives all three resilience branches off the AacResilienceFlags triplet: section_data() through the 5-bit sect_cb branch (carrying the §4.6.16.4 virtual codebooks 16..=31, whose ESC_HCB / >= 16 runs take the fixed sect_len_incr = 1 single-band coding) when aacSectionDataResilienceFlag is set; scale_factor_data() through the RVLC ErScaleFactorData branch (its reconstruction mirrored into the shared scale_factor_data field so the §4.6.2.3.2 accumulate pass is branch-agnostic, with the RVLC seeds retained in er_scale_factor_data) when aacScalefactorDataResilienceFlag is set; and the length_of_reordered_spectral_data (14-bit) + length_of_longest_codeword (6-bit) HCR length fields in reordered_spectral_lengths when aacSpectralDataResilienceFlag is set. The trailing reordered_spectral_data() (HCR) payload is the caller's responsibility, exactly as spectral_data() is on the non-resilient path.
• HCR segmentation / pre-sorting scaffold (hcr) — ISO/IEC 14496-3 §4.6.16.3.3 / §4.6.16.3.5. The deterministic, header-only half of Huffman codeword reordering: the Table 4.170 maxCwLen table, the §4.6.16.3.3.1 codebookPriority[32] table + the assignedUnitNr pre-sorting metric, the segmentWidth = min(maxCwLen, length_of_longest_codeword) derivation, the §4.6.16.3.2 length-field clamps, and the Segmentation layout that instantiates PCW segments until the length_of_reordered_spectral_data buffer is exhausted (folding the trailing bits into the last segment). ReorderPlan::build then runs the §4.6.16.3.3.4 ReorderSpectralData() writing scheme (PCWs forward from each segment start, then the non-PCW set / trial loop with the per-set ToggleWriteDirection() and the modulo-shift segment = (trial + codewordBase) % numberOfSegments) to resolve, for each codeword, the ordered global buffer bit positions (MSB-first) that carry its bits — pinned by a bijection invariant (every buffer bit covered exactly once).
• HCR payload codec (hcr_decode) — §4.6.16.3.3.4 / §4.6.16.3.4, both directions of the reordered_spectral_data() payload itself. encode_reordered_spectral_data enumerates the frame's codeword units (the §4.5.2.3.2 unit: Huffman codeword + sign bits + escape sequences, two or four lines) in the §4.6.16.3.3.1 pre-sorted order — the unit-based window interleave (Table 4.169; the §4.5.2.3.5 grouping interleave does not apply under HCR) stably ordered by assignedUnitNr — encodes each unit and scatters the bits over the segment grid via ReorderPlan. decode_reordered_spectral_data inverts the walk without transmitted lengths: PCWs decode forward from their own segment starts, then the non-PCW sets run the same direction-toggling modulo-shift trial loop, each codeword consuming segment free-region bits until its Huffman unit completes (prefix codes make "incomplete" exactly detectable, so lengths are discovered where the writer defined them). The §4.6.16.4 virtual codebooks (16..=31) decode as book 11 with their own maxCwLen segment widths. Round-tripped bit-exactly over long / eight-short (two window groups), spectrum-less-band mixes, escape-bearing book 11, virtual codebooks, and slack-padded buffers; corrupt payloads surface errors, never panics. Threading the ER triplet from GASpecificConfig through the stream drivers (the ER top-level payloads) remains open, and an HCR-bearing conformance stream is still wanted as an external cross-check.

Numeric reconstruction (AAC-LC tool chain)

• Spectrum Huffman codebooks 1..=11 (spectrum_huffman, spectral_codebook) — the complete Annex 4.A spectrum book set, including the ESC book 11, with §4.6.3.3 index↔tuple translation and sign-bit / escape-sequence handling.
• Inverse quantization + scalefactors (dequant) — §4.6.1.3 non-uniform inverse quantizer and §4.6.2.3.3 scalefactor gain.
• Decoded spectrum (decoded_spectrum) — §4.6.3.3 quant_to_spec() de-interleaver plus the per-channel pipeline composing pulse fix-up → scalefactor accumulation → inverse quantization + rescale → de-interleave → TNS.
• TNS (tns_data, tns_coef, tns_frame, tns_max, swb_offset) — §4.6.9 Temporal Noise Shaping: wire parse, coefficient inverse-quantisation + conversion to LPC, the all-pole IIR pass, and the per-frame region-slicing orchestration.
• Filterbank (filterbank) — §4.6.11 stateful per-channel IMDCT with sine / KBD windows, all four window_sequence shapes, eight-short internal overlap-add, and inter-frame overlap-add. Pinned by streaming TDAC perfect-reconstruction tests. Covers all four §4.5.1.1 frame-length families: the 1024/128- and 960/120-line block-switching families (N = 2048/256 and 1920/240) and the long-only ER AAC LD 512/480-line families (N = 1024/960), where the window_shape == 1 bit selects the §4.6.17.2.3 Table 4.171 low-overlap window in place of KBD (power-complementarity and streaming TDAC pinned per family).
• Channel-pair / noise tools — M/S stereo de-matrix (ms_stereo, §4.6.8.1), intensity stereo (intensity_stereo, §4.6.8.2), and Perceptual Noise Substitution (pns, §4.6.13). PNS produces energy-exact bands; only the per-coefficient phase is RNG-defined per §4.6.13.3, so its output is not byte-exact against any one decoder — the staged docs/audio/aac/pns-gen-rand-vector.md analysis pins the normative half (band selection, energy DPCM, measured-energy normalisation, correlated-CPE same-vector rule — all implemented) and shows the generator recurrence/seed/threading to be deliberately unspecified, so cross-decoder PNS checks are energy-domain by design.
• Coupling channel element (cce) — §4.6.8.3 / Table 4.8. The CCE coupling header (CouplingHeader: ind_sw_cce_flag, num_coupled_elements, the per-target cc_target_is_cpe / cc_target_tag_select / cc_l / cc_r list with the Table 4.153 shared-vs-split num_gain_element_lists derivation, cc_domain / gain_element_sign / gain_element_scale) and the trailing gain-list block (CouplingGains: per-target common_gain_element or per-(g, sfb) dpcm_gain_element running-sum lists — the §4.6.8.3.3 ind_sw_cce_flag ⇒ common-gain-only constraint and the embedded-SCE sfb_cb ZERO_HCB skip — reusing the §4.A.1 scalefactor Huffman codebook hcod_sf). CouplingChannelElement ties the whole Table 4.8 element (header → embedded individual_channel_stream(0,0) body + spectrum → gain lists) together, and CouplingGains::cc_gain computes the §4.6.8.3.3 couple_channel() factor cc_gain = cc_sign · cc_scale^gain (Table 4.154 cc_scale_table, implicit list-0 natural scaling). CouplingGains::couple_channel applies the §4.6.8.3.3 per-band scale-and-add — the spec's group / window-group / sfb / coefficient loop multiplying the embedded-SCE spectrum by the per-(g, sfb) cc_gain and adding it onto a target channel's window-major spectrum (implicit list 0 in natural scaling, ZERO_HCB bands skipped). The cross-element application is wired into the stream decoder: the raw-data-block walk is two-pass (parse every channel element, then decode), each CCE's embedded SCE is decoded through a per-instance-tag CceDecoder slot (its own pulse / dequant / PNS / TNS, plus its own persistent §4.6.11 filterbank for the independently-switched case), and the decode_coupling_channel() target walk matches cc_target_is_cpe / cc_target_tag_select, assigns the Table 4.153 gain lists (shared / left / right / both), and injects the scaled spectrum at the signalled cc_domain stage (before / after the target's TNS; window-state match enforced) or — for an independently switched CCE — the scaled time signal after the target's filterbank. Validated end to end with writer-assembled CCE streams against the filterbank-linearity identity decode([target, CCE]) = decode([target]) + cc_gain·decode([embedded]) (natural scaling, gain-list ×2, independently-switched, and CPE-left-only layouts, ≤ 2 LSB stacked-rounding deviation). A writer-assembled CCE fixture cycling all three coupling shapes (dpcm + sign split / ind-switched / shared natural, both domains, PCE-declared valid_cc_elements) is staged in the docs corpus (aac-cce-writer-assembled). The two long-standing §4.6.8.3.3 wire questions are both settled: the exponent is negated — cc_gain = cc_sign · cc_scale^(−gain_element) — confirmed by the ISO/IEC 14496-26 am05_* conformance vectors (all three editions print the positive exponent, which misses every coupled target by ~1e-1 err/sig), and the gain_element_sign split follows the 2001 / 13818-7:2004 couple_channel() text as ruled in docs/audio/aac/cce-gain-sign-split.md §3 — cc_sign off each transmitted dpcm delta (1 − 2·(dpcm & 1)), accumulator fed with dpcm >> 1, and a common_gain_element never sign-split (the 14496-3:2009 page prints two conflicting fragments; its accumulated-value variant is an editorial defect of that edition).
• Frequency-domain prediction (predictor) — §4.6.6 MPEG-2 backward-adaptive intra-channel predictor for the AAC Main object type (AOT 1). A bank of second-order lattice predictors (one per MDCT line up to the §4.6.6.2 PRED_SFB_MAX limit) reconstructs x_rec = x_est + y_rec on the signalled bands. Implements the §4.6.6.3.2.1 lattice predict() + LMS adaptation (α = 0.90625, a = b = 0.953125), the §4.6.6.3.2.3 flt_round_inf() 16-bit-float rounding applied to every stored state variable and the predicted value, and the §4.6.6.3.3 reset (the 30 Table 4.97 cyclic groups + the short-block reset-all). Wired into element_decode: the bank runs every long frame before TNS (and is mutually exclusive with LTP by object type), persisting the backward-adaptive state across frames.
• Long-Term Prediction (ltp) — §4.6.7 long-window LTP: the Table 4.98 coefficient codebook, the §4.6.7.3 predict() single-tap time-domain predictor (x_est(i) = ltp_coef·x_rec(i − ltp_lag)) over a per-channel x_rec reconstruction history, the windowed analysis MDCT(x_est) (the §4.6.15.3.3 / §4.6.11.3.1 forward transform, now a reusable filterbank primitive), and the per-sfb X_rec = X_est + Y_rec combination on the bands flagged by ltp_long_used. LTP is restricted to long windows for the AAC LTP object type (§4.6.7.1, 2009 edition). The ISO/IEC 14496-3:2001 short-window synthesis (LtpState::apply_short_2001) is also implemented per the 2001 §4.6.7.3 pseudo-code — per flagged subwindow, lag_w = ltp_lag + ltp_short_lag[w], the 256-point windowed MDCT(x_est), and the X_rec = X_est + Y_rec add on the first 8 SFBs — with the one quantity the 2001 text never fixes (the per-subwindow x_rec index origin; see the staged docs/audio/aac/short-window-ltp-blocked.md §5) taken as an explicit caller parameter rather than an invented convention. The ER AAC LD branch is implemented: the 10-bit lag with the ltp_lag_update repeat state (ltp_prev_lag) and the §4.6.7.3 M = N/2 lag offset, applied at the LD transform lengths.
• TNS analysis filter (tns_coef::tns_ma_filter, tns_frame::tns_analysis_frame) — §4.6.7.4.1 / Figure 4.30: the all-zero (moving-average, FIR) inverse of the §4.6.9.3 all-pole synthesis filter, y(n) = x(n) + Σ lpc[k]·x(n−k). Run over the same per-window region walk as tns_decode_frame; analysis ∘ synthesis is the identity over a shared region, which is the §4.6.7.4.1 noise-shaping invariant.
• Element decode driver (element_decode) — ElementDecoder chains the whole stack per element: decode_sce for SCE / LFE and decode_cpe for a CPE (pulse → dequant → quant_to_spec() → M/S → intensity → PNS → LTP → TNS → filterbank), carrying the per-channel overlap-add tail and the §4.6.7.3 LTP reconstruction history across frames. LTP runs in the §4.6.7.4.1 / Figure 4.30 block order — long-term synthesis (with the all-zero TNS analysis filter applied to X_est) before the §4.6.9 TNS synthesis filter, so the single synthesis pass shapes the residual while undoing the analysis on the LTP contribution.

Frame-length families — §4.5.1.1 / §4.6.17 (960, LD 512/480)

All four frameLengthFlag frame geometries decode end to end, keyed by swb_offset::FrameFamily (resolved from the AOT + flag; the LATM/LOAS driver installs it per layer from the ASC, StreamDecoder::set_frame_family serves raw callers — ADTS can only carry the default 1024-line family):

• AAC-LC at 960/120 lines (frameLengthFlag == 1) — the bracketed "values for 1920/240" columns of Tables 4.129–4.141, the N = 1920/240 transform pair with all four window sequences, both window shapes, grouping, TNS and the full joint-stereo/noise tool chain; 960 PCM samples per channel per frame. Verified bit-exact against a black-box decoder binary on writer-assembled streams, and staged with mutation coverage as aac-lc-960-writer-loas.
• ER AAC LD at 512/480 lines (AOT 23, §4.6.17) — Tables 4.142–4.147 (with the §4.5.1.1 nearest-defined-table rule for the rates those tables omit), long-only frames (a non-ONLY_LONG window_sequence is rejected, §4.6.17.2.2), the §4.6.17.2.3 Table 4.171 low-overlap window on the window_shape == 1 bit, the §4.6.17.2.5 LD TNS_MAX_BANDS tables, the §4.6.7 LD LTP branch (10-bit lag, ltp_lag_update repeat via a per-channel ltp_prev_lag, M = N/2 lag offset at the LD transform lengths), and the Table 4.19 er_raw_data_block() element walk shared with ER AAC LC; 512/480 PCM samples per channel per frame. The LD-512 geometry (every swb band boundary, both window shapes, the transform/overlap-add) is verified bit-exact against two independent black-box decoder binaries; LD-480 against one (the other binary decodes 480-line streams on the wrong 512-line frequency grid — probed and documented in the fixture notes). Staged fixtures: aac-ld-512-writer-loas, aac-ld-480-writer-loas.
• LD TNS wire — RESOLVED against the conformance corpus: the divergence between the literal Table 4.54 / Table 4.155 field widths and the deployed LD TNS wire was settled by the ISO/IEC 14496-26 screen recorded in docs/audio/aac/er-ld-tns-divergence.md §0 — the normative LD wire transmits n_filt in 1 bit (the reduced Table 4.155 column; the literal 2-bit keying hard-fails 792 of the corpus's 2 017 TNS-bearing AUs). The LD families read the 1 / 4 / 3 column via TnsData::parse_family / write_family; because the corpus never transmits a length / order field (n_filt == 0 throughout, so 4 / 3 vs 6 / 5 is undetermined), the §0.6 configurability recommendation is kept via the explicit-width TnsData::parse_widths / write_widths entry points.
• SBR over the 960-line core — §4.6.18.2.6 numTimeSlots = 15 ("16 for a 1024 AAC frame and 15 for a 960 AAC frame"): the SBR decoder (SbrDecoder::new_slots) and the Annex 8.A PS chain (numQMFSlots = numTimeSlots · RATE = 30: hybrid analysis history, envelope interpolation, mixing) run over 30-slot frames, so a 960-line LATM/ASC stream with SBR decodes to 1920 samples per channel at the doubled rate; a payload-less frame holds the state through pure upsampling. The SBR encoder is laid out over the same family (SbrEncoderConfig::num_time_slots = 15: 48 analysis columns per frame, every grid builder and the onset detector over 15 slots, trailing borders in 15..=18), and a writer-assembled LC-960 stream carrying its payloads decodes end to end. The available reference decoder binary does not implement SBR over 960-line frames (it says so under explicit signalling and decodes the core alone under implicit signalling), so the black-box check for this family is the core band: no diagnostics, and the per-band energies below the crossover match this crate's decode in shape (one common scale). An SBR payload on an LD stream is still rejected before its body is parsed (Error::SbrUnsupportedFrameFamily) — the §4.6.19 LD SBR tool belongs to ELD (out of scope).

Scalable AAC (AOT 6) / ER AAC scalable (AOT 20) — §4.4.2.2 / §4.5.2.2

The AAC-only scalable combinations decode end to end (scalable): one aac_scalable_main_element() plus up to seven extension elements, each on its own elementary stream / LATM layer (mono-only, stereo-only and mixed mono→stereo stacks, Table 4.87).

• Syntax — Tables 4.13–4.18: the main/extension headers (window geometry hoisted out of ics_info(), per-channel TNS on the first mono and first stereo layer, per-channel LTP on the main layer, the §4.6.8.1.4 incremental ms_data() over last_max_sfb_ms..max_sfb, per-channel diff_control_data_lr() with the Table 4.18 ms_used gating), and the Table 4.50 scale_flag == 1 ICS form (IcsBody::parse_scale — no inline ics_info(), no tool dispatch trio). For AOT 20 the §4.4.6 resilience triplet selects the ER wire branches per channel (5-bit-sect_cb sections, RVLC scalefactors, inline HCR reordered_spectral_data()); the element syntax itself is unchanged (§4.5.2.4), pinned bit-identical to the AOT-6 decode of the same spectra. ScalableFrame::parse / ::write round-trip the whole per-layer payload stack byte-exactly.
• Layer combination (§4.5.2.2.4 SIAQ, ScalableDecoder): the dequantized spectra of all layers sum per output path under the Table 4.91–4.93 per-band tool rules — a lower layer's PNS band survives only while every higher layer decodes the band to zero (§4.6.13.6), intensity accumulates the M/L channel with positions from the highest layer, invalid combinations surface Error::ScalableLayerCombination — then the §4.6.14.2.1 FSS merges the combined mono spectrum into the stereo pair (L/R += 2·M'' on clear diff_control_lr bits, M = M'' + M' on M/S bands; long and short windows), the cumulative-mask M/S butterfly, the scalable-invariant intensity reconstruction (invert_intensity() = +1), correlated PNS via ms_used, the §4.6.9.5 / Table 4.158 serial TNS layout (first mono layer's filter serves the low bands up to the highest mono max_sfb, first stereo layer's filters serve L/R, with the lower-boundary override rule) and the §4.6.11 filterbank. §4.6.7.5 base-layer LTP runs on the lowest layer only, its reconstruction history fed by a parallel first-layer-alone synthesis chain (pinned by a history-isolation test). Both the 1024- and 960-line families.
• Transport: the LATM/LOAS driver recognises AOT-6/20 layers, collects each program's layer payloads per access unit and decodes them combined through a persistent per-program ScalableDecoder (ScalableConfig::from_layer_ascs validates the layer stack; dependsOnCoreCoder == 1 — a CELP core — and TwinVQ lower layers are out of scope, Error::ScalableUnsupportedCore).
• Single-layer scalable streams are pinned bit-identical to the equivalent SCE / common-window CPE decodes; multi-layer stacks are pinned against references composed from the crate's own reconstruction primitives; every branch carries a deterministic bit-flip / truncation battery (tests/scalable_*.rs).

Error protection (EP) tool — §1.8

The MPEG-4 Audio unequal-error-protection layer, from the out-of-band configuration to the LOAS EP carrier (ep_config / ep_fec / ep_rs / ep_frame):

• ErrorProtectionSpecificConfig() (Table 1.49) — parse + bit-exact write with reserved-field rejection, the §1.8.4.2 class_optional expansion (pinned against the spec's own Table 1.57/1.58 example), and ASC integration: epConfig == 2 / 3 now parse the inline config and the directMapping bit.
• SRCPC (§1.8.4.6) — the rate-1/4 systematic recursive convolutional encoder (Figure 1.10 equations), the Table 1.61 puncture family 8/8..8/32, §1.8.4.6.2 termination (the u = d tail rule, proven identical to the whole Table 1.60 listing) and a hard-decision 16-state Viterbi decoder correcting channel errors.
• In-band header FEC (§1.8.4.3 Table 1.59) — majority, BCH(7,4), BCH(15,7), Golay(23,12), BCH(31,16) with the normative generators and bounded-distance correction; CRC4 + terminated SRCPC 8/16 for 17+ bits; the extended header_protection path.
• Shortened Reed-Solomon (§1.8.4.7) — SRS(255−l, 255−2k−l) over the spec's GF(2⁸) (m(x) = x⁸+x⁴+x³+x²+1; the generated antilog table is pinned against Table 1.62 rows), the part split with zero-padded last part, lowest-order-first parity, and the syndrome / Berlekamp-Massey / Chien / Forney correction chain (k byte errors per part corrected, k+1 rejected).
• ep_frame() (§1.8.2.2, EpFrameCodec) — the FEC-protected choice_of_pred + class_attrib() header (in-band Table 1.55 rate / Table 1.56 CRC escapes, num_stuffing_bits), per-class §1.8.4.5 CRC (the family now reaches down to CRC1) + SRCPC / SRS protection, §1.8.4.4 RS chains, the "until the end" class-length recovery (§1.8.4.1), §1.8.4.9 class-reordered transmission, and the §1.8.4.8 recursive interleaver (k = m·D + min(m, d) + n; bitwise for SRCPC, bytewise for RS) in modes 0 / 1 / 2 with the per-class mode-2 interleave_switch. Encode ↔ decode round-trips across the configuration matrix; errors are corrected through the whole frame; a full bit-flip battery never panics. Two under-specified corners (an escaped rate on an RS class; byte-wise interleave over a non-octet-aligned Y stream) are rejected rather than guessed.
• LOAS EP carrier (§1.7) — the EPAudioSyncStream() BCH(36,18) headerParity (§1.7.2.2.2 generator; generate + verify), EPMuxElement(1, 1) (majority-protected epUsePreviousMuxConfig, Golay-protected epSpecificConfigLength, Table 1.59-protected inline config with threaded reuse), and LoasDecoder::decode_all_ep — the recovered ep_frame() class concatenation is the plain AudioMuxElement() bit stream (§1.7.3.2.1: the sensitivity-category instances ride in syntax order), so payloads (scalable programs included) ride the existing decode paths. A writer-assembled EP stream decodes byte-identical to its plain LOAS equivalent and survives correctable channel errors.

SSR gain control (§4.6.12) — complete decode pipeline

The §4.6.12 SSR (Scalable Sample Rate, AOT 3) gain-control tool is implemented end to end — front-half filterbank, gain reconstruction, and IPQF synthesis — and wired into the decode driver (ADTS profile 2 routes every SCE / CPE channel through it), validated independent of any external SSR implementation:

• Gain-control reconstruction (gain_control) — §4.6.12.3.1–3. The §4.6.12.3.1 gain-control data decoding (the Table 4.108 AdjLoc() = 8·AC and Table 4.109 AdjLev() = AV − 4 tables, the NADW / ALOC / ALEV ladder with the step-(3) ALOC(0)=0 / ALEV(0) rule and the step-(4) per-window-sequence endpoint), the §4.6.12.3.2 gain-control function setting (the M_{W,B,j} index, the FMD fragment-modification function with the Inter(a,b,j) geometric-blend ramp, the per-sequence GMF composition threading the cross-frame PFMD, and the inversion AD(j) = 1/GMF(j)), and the §4.6.12.3.3 windowing + overlapping (GainBandState::window_overlap applies T = AD·U then overlap-adds per window_sequence into the band sample data V_B, threading the cross-frame PT_B tail). All four window_sequence shapes are covered; the spec initial values PFMD ≡ 1.0 / PT ≡ 0.0 are honoured, and the input-read vs produced PFMD lengths (which differ per sequence) are tracked separately with a persistent 256-entry carry.

• IPQF synthesis filter (ipqf) — §4.6.12.3.4. The Table 4.110 length-96 prototype Q(j) (the symmetric Q(j) = Q(95 − j) half mirrored to 96), the cosine modulation Q_B(j) = Q(j)·cos((2B+1)(2j−3)π/16), the 4× upsampling Ṽ_B(j) = V_B(j/4), and the streaming convolution AS(n) = Σ_B Σ_j Q_B(j)·Ṽ_B(n−j) as a polyphase bank (Ipqf) that retains a 24-deep per-band history across frames — pinned by an impulse-response test against the direct §4.6.12.3.4 convolution (AS(n) = Q_0(n)).

• Per-channel driver (ssr) — SsrGainControl::decode_frame composes the four-band GainBandState and the Ipqf into one persistent per-channel pipeline: the four per-band IMDCT outputs U_{W,B} plus the decoded gain_control_data() → the §4.6.12.3.3 per-band windowing/overlap → the §4.6.12.3.4 IPQF synthesis → the PCM AS(n) (1024 samples/frame for the steady ONLY_LONG / EIGHT_SHORT case). PQF band 0 is never gain-controlled.

• Front-half filterbank (ssr_filterbank) — §4.6.12.1 / 13818-7 §16.1, closing the previously docs-gapped spectrum→band mapping: the frequency-ascending spectrum splits into four contiguous PQF-band quarters (the PQF's band B covers the B-th quarter, Annex C.2.1.1), the "even" bands — the spec's ordinal 2nd/4th, i.e. 0-based 1 and 3, exactly the bands the ×4 decimation spectrally inverts — are reversed, and each band runs a 256-line (long) / 8 × 32-line (short) IMDCT under the quarter-scale §4.6.11.3.2 window geometry (N_l/N_s = 512/64; the KBD windows are generated with the α = 4 / α = 6 running-sum construction and pinned against the normative Table 4.A.14 / 4.A.13 listings). The split + reversal convention is pinned by a tone-placement test against the Annex C.2.1.1 analysis-PQF definition.

• Per-channel pipeline + driver wiring (ssr::SsrChannelDecoder, element_decode) — the complete spectrum → PCM chain (front half → §4.6.12.3 gain compensation/overlap → IPQF), replacing the §4.6.11 filterbank for AOT 3 in the decode driver (per-channel-slot state, gain_control_data() from the channel body; note the §4.6.12.3.3 variable frame lengths — 1472 / 576 PCM samples for LONG_START / LONG_STOP). Validated by full round-trip tests against the Annex C.2.1.1 analysis PQF: steady long frames and a complete window-transition chain reconstruct at err/sig < 1e-3 (the PQF pair's near-perfect-reconstruction bound, both window shapes), gain ladders applied encoder-side cancel end to end, and an ADTS-profile-2 stream decodes through the public StreamDecoder (mono + stereo).

SBR bitstream decode (HE-AAC)

The full SBR side-info path is now decoded from the extension_payload SBR element down to the reconstructed quantized envelope / noise-floor scalefactors — every numeric table sourced from the ISO/IEC 14496-3 spec PDF (the §4.A normative Huffman grids and the §4.4.2.8 syntax tables), independent of any external SBR table extraction.

• SBR Huffman codebooks (sbr_huffman) — §4.A.6.1, all ten normative envelope / noise codebooks (Tables 4.A.79–4.A.88) transcribed from the spec codeword grids and validated complete + prefix-free. sbr_huff_dec() reads MSB-first and returns the signed DPCM delta (index − LAV); env_tables() / noise_tables() pick the (t_huff, f_huff) pair from the §4.6.18.3 coupling / channel / bs_amp_res selection (the freq-direction noise tables alias the 3.0 dB envelope freq tables per Table 4.A.78 Note 2).
• sbr_header() (sbr_header) — §4.4.2.8 Table 4.63: the fixed-width header plus the two optional extra blocks, with the Table 4.63 Note 3 defaults (Tables 4.105–4.111) applied when an extra flag is clear. band_geometry_changed() flags a §4.6.18.3.3 reset, and derive_bands() chains into the band-setup pipeline below.
• sbr_grid() / sbr_dtdf() / sbr_invf() (sbr_grid) — §4.4.2.8 Tables 4.69–4.71: all four bs_frame_class layouts (FIXFIX / FIXVAR / VARFIX / VARVAR) with the envelope count, variable / relative borders, ptr_bits = ceil(log2(num_env + 1)) pointer, reversed FIXVAR freq-res order, single-envelope FIXFIX bs_amp_res override, and bs_num_noise derivation; the delta-direction flags; and the per-noise-band 2-bit inverse-filtering modes.
• sbr_envelope() / sbr_noise() (sbr_envelope) — §4.4.2.8 Tables 4.72–4.73: the raw bs_data_* delta arrays, with the fixed-width absolute start value (5/6/7-bit per the coupling / channel / bs_amp_res context; noise always 5-bit) and the frequency- vs time-direction Huffman deltas, over NHigh / NLow envelope bands and NQ noise bands.
• Envelope / noise DPCM reconstruction (sbr_reconstruct) — §4.6.18.3.5: inverts the delta coding to the quantized scalefactors E_Q(k,l) / Q(k,l). Frequency deltas accumulate from the start value; time deltas add to the reference envelope (previous in-frame, or the prior frame's last envelope for l == 0) with the i(k) high↔low band remap when the reference resolution differs; the coupled second channel's δ = 0.5 is applied as an integer ×2 on the even transmitted values, threading cross-frame state.
• Element framing (sbr_element) — §4.4.2.8 Tables 4.65 / 4.66 / 4.74: SbrElement::parse_single / parse_pair decode a whole SBR data element in spec order — the optional bs_data_extra field, the per-channel grid / dtdf / invf / envelope / noise blocks (coupled shared-grid vs. independent-grid layouts, second coupled channel in balance mode), the sbr_sinusoidal_coding() add-harmonic flags, and the bs_extended_data block (id + raw body captured for a later PS pass). The single-envelope FIXFIX bs_amp_res override is applied before envelope decode.
• sbr_extension_data() (sbr_extension) — §4.4.2.8 Table 4.62: the top-level walker that ties the header + element framing into a whole SBR extension payload, in spec order — the optional 10-bit bs_sbr_crc_bits (for the EXT_SBR_DATA_CRC type), the bs_header_flag + sbr_header(), then sbr_data(id_aac, bs_amp_res) dispatching onto parse_single (ID_SCE) / parse_pair (ID_CPE) with the band tables derived from the active header at the SBR internal rate (FsSBR = 2·core), and the trailing bs_fill_bits alignment (num_align_bits = (8·cnt − 4 − num_sbr_bits) % 8). A clear bs_header_flag reuses the threaded previous header (the non-scalable core fixes sbr_layer == SBR_NOT_SCALABLE, so the flag is always present). Reachable from the natural FIL entry point via extension_payload::ExtensionPayload::parse_with_sbr, which routes the SBR extension types here (the default parse still rejects them, keeping the byte-exact AAC-LC corpus path untouched).

The SBR bitstream side info is decoded end to end — CRC field, header, element framing, band tables, and envelope / noise DPCM reconstruction — and the back-end DSP is now implemented too (see the next section).

SBR back-end (HE-AAC v1) — §4.6.18

The complete SBR reconstruction chain, from the core decoder's time signal to dual-rate PCM, validated 99.98% sample-exact (max error 1 LSB) against the staged HE-AAC v1 expected.wav:

• QMF filterbanks (sbr_qmf) — §4.6.18.4 / Figures 4.42–4.44: the Table 4.A.89 640-tap prototype window (transcribed digit-for-digit from the spec PDF), the 32-band complex analysis bank, the 64-band real-output synthesis bank (dual-rate), and the downsampled 32-channel synthesis variant. Pinned by near-perfect-reconstruction properties (< 1e-4 error ratios).
• Dequantization + stereo decoding (sbr_dequant) — §4.6.18.3.5: EOrig = 64·2^(E/a), QOrig = 2^(6 − Q), and the coupled-pair pan split with panOffset = [24, 12] (energy-sum-preserving).
• Time / frequency grid (sbr_time_grid) — §4.6.18.3.3: the tE / tQ border vectors for all four frame classes, the Table 4.174 middleBorder and the Table 4.176 lA.
• HF generation (sbr_hf_gen) — §4.6.18.6: the Figure 4.48 patch construction, the covariance-method second-order inverse filtering (εInv = 1e-6, |α| ≥ 4 reset), the Table 4.175 chirp-factor blend, and the patched XHigh generator.
• Limiter band table (sbr_limiter) — §4.6.18.3.2.3 / Figure 4.41, fed by the patch borders (closing the previously deferred limiter-table item).
• Envelope adjustment (sbr_env_adjust + sbr_noise_table) — §4.6.18.7: mapping, ECurr estimation (both bs_interpol_freq regimes), amplitude-domain gains (the spec PDF's typeset equations carry square roots the plain text layer drops), the limiter / boost compensation, hSmooth smoothing with cross-frame tails, the Table 4.A.91 noise table with the running fIndexNoise, and the sinusoid injection with the (−1)^(m+kx) alternation.
• Frame driver (sbr_decoder) — §4.6.18.5 / Figure 4.47: the tHFGen = 8-slot XLow history, header-reset handling, the lTemp splice of the previous frame's Y', the coupled-pair invf sharing, and the pure-upsampling path for SBR-less frames.
• Stream wiring (decode) — the ADTS StreamDecoder walks FIL extension payloads via extension_payload::parse_with_sbr, attaches each SBR payload to its preceding SCE / CPE, threads the sbr_header() reuse state per element slot, and (once SBR-active) emits every frame at the doubled rate — 2048 samples/channel — with SBR-less frames upsampled so the output rate never flaps. The runtime Decoder trait surfaces the dual-rate frames unchanged (pinned byte-identical to the raw StreamDecoder).
• Downsampled output mode (§4.6.18.4.3) — selectable end to end: SbrDecoder::set_downsampled / StreamDecoder::set_sbr_downsampled / the sbr_downsampled codec option run the 32-channel synthesis bank so an SBR-active stream is emitted at the core rate (1024 samples per channel per frame; the SBR range above the core Nyquist is discarded by construction, the bands below it are kept). The LATM driver selects the mode automatically when an explicitly signalled ASC carries extensionSamplingFrequency == samplingFrequency (the §4.6.18.2.6 in-band core-rate declaration), and PS composes (stereo through two downsampled banks). Validated on the HE-AAC v1 fixture at 1.8e-4 per-channel err/sig RMS against a band-limited 2:1 decimation of the reference decode (v2 PS at 1.95e-4), byte-identical between the LATM and forced-ADTS paths.
• Low power SBR tool (§4.6.18.8) — selectable end to end (SbrDecoder::set_low_power / StreamDecoder::set_sbr_low_power / the sbr_low_power codec option; composes with the downsampled output): the §4.6.18.8.2 real-valued filterbank trio (sbr_qmf), the §4.6.18.8.3 aliasing detection (sbr_lp + the reflection coefficients in sbr_hf_gen: the Figure 4.53 degree walk, the patch-carried degPatched, the Figure 4.54 gain groups), the §4.6.18.8.4 ×2 energy estimation, and the §4.6.18.8.5 aliasing reduction (GLimBoost → GA, exact group-energy restoration), no-smoothing rule, modified real-valued sinusoid injection (−0.00815 neighbour correction, first-16 rule, kx − 1 / kx + M spill) and modified X assembly. Validated on the HE-AAC v1 fixture: sub-crossover content at 9e-5 err/sig RMS against the reference with per-frame full-band energy within 0.05% (the real-valued HF path is energy-normative, not phase-normative). A PS payload in this mode is rejected (Error::SbrLowPowerPs — the subpart-8 tool needs the complex QMF domain). All four mode combinations survive a deterministic corruption battery (tests/sbr_mode_mutations.rs).

SBR frequency band setup (HE-AAC)

• SBR frequency band tables (sbr_freq_bands) — §4.6.18.3.2 the static, header-only half of the Spectral Band Replication band setup, computed directly from the closed-form spec algorithm (no QMF back-end required):
  • k0 / k2 — §4.6.18.3.2.1 the low and high QMF subband boundaries. k0 = startMin + offset(bs_start_freq) with the per-FsSBR offset table and the startMin = NINT(c·128/FsSBR) thresholds; k2 covers the bs_stop_freq < 14 stopDkSort accumulation path and the bs_stop_freq == 14 / 15 min(64, 2·k0) / min(64, 3·k0) shortcuts.
  • master_table — §4.6.18.3.2.1 fMaster, both the Figure 4.39 linear path (bs_freq_scale == 0, the dk/vDk/k2Diff away-from-zero correction walk) and the Figure 4.40 warped path (bs_freq_scale > 0, the bands/warp log-spaced regions with the single-/two-region split at k2/k0 > 2.2449 and the min(vDk1) < max(vDk0) smoothing step).
  • HiLoTables::derive — §4.6.18.3.2.2 the derived fTableHigh, fTableLow (the i(k) = 2k − (1−(−1)^NHigh)/2 decimation), and fTableNoise (the NQ = max(1, NINT(bs_noise_bands·log2(k2/kx))) band count plus its i(k) recursion), along with the M and k_x outputs every later SBR stage keys off.
  • The §4.6.18.3.6 requirements are enforced (k2 > k0, numBands > 0, vDk > 0, bs_xover_band < NMaster), surfacing Error::SbrFreqBandInvalid on violation. The §4.6.18.3.2.3 limiter band table is out of scope here — its bs_limiter_bands > 0 path consumes the §4.6.18.6 patch borders that need the QMF patching back-end.

Parametric Stereo (HE-AAC v2) — subpart 8 / Annex 8.A

The complete §8.6.4 PS tool, reconstructing a stereo image from the mono SBR signal, validated 5e-5 per-channel error-to-signal RMS against the staged HE-AAC v2 MP4 fixture (filterbank-rounding level):

• Bitstream (ps_data, ps_huffman) — §8.4.2 Tables 8.9–8.14: the persistent enable_ps_header configuration, FIX/VAR framing (Table 8.29), per-envelope IID/ICC/IPD/OPD delta rows on all ten Annex 8.B codebooks (each verified a complete prefix code; the six IID/ICC books cross-checked leaf-for-leaf against the staged ps-huffbook-*.csv trees), and the §8.5.2 time/frequency DPCM resolution with range checks and modulo-8 phase wrap.
• Hybrid filterbank (ps_hybrid) — §8.6.4.3: both configurations (71 / 91 sub-subbands) on the Table 8.37/8.38 13-tap prototypes, with the Figure 8.20 merge/reorder, the odd-QMF-band inversion, and the Annex 8.A.3 zero-delay alignment (6 look-ahead XLow slots + 6 history slots). Analysis→synthesis reconstructs exactly.
• De-correlation (ps_decorr) — §8.6.4.5: the 3-link complex all-pass chain behind z⁻²·φ_fract, the Table 8.40/8.41 centre frequencies, the 14-/1-slot delays above NR_ALLPASS_BANDS, and the transient duck (peak decay / smoothing / γ = 1.5) per stereo band, with the Annex 8.A.3 partial + full resets.
• Stereo processing (ps_stereo, ps_map) — §8.6.4.6: Table 8.25/8.26/8.28 dequantization (cross-validated against the staged Q30 tables), mixing procedures Ra and Rb, IPD/OPD three-position smoothing, the Table 8.48/8.49 b(k) maps + conjugate channels, the Table 8.45/8.46 10↔20↔34 re-mappings, and the §8.6.4.6.4 border interpolation with hold semantics.
• Frame driver + wiring (ps_decoder, sbr_decoder) — Annex 8.A: inactive (mono) until the first header'd ps_data(), parameter hold over payload-less frames, band-count switches, and the per-frame de-correlator reset above k_x + M. A PS-carrying SCE renders stereo through two synthesis banks in both the SBR-processed and pure-upsampling paths, end to end through the ADTS / LATM / raw StreamDecoder entries and the runtime Decoder.

LATM / LOAS transport framing

The §1.7 low-overhead transport layer is now decoded from the LOAS sync frame down to the recovered MPEG-4 Audio access units — every field sourced from the ISO/IEC 14496-3 §1.7 syntax tables.

• StreamMuxConfig() (latm::StreamMuxConfig) — §1.7.3.1 Table 1.42 plus LatmGetValue() (Table 1.43). Decodes the whole multiplex configuration: the audioMuxVersion / audioMuxVersionA version flags (with the audioMuxVersion == 1 taraBufferFullness and per-ASC length-prefix + fillBits extensions), allStreamsSameTimeFraming, numSubFrames / numProgram / per-program numLayer, and the per-streamID[prog][lay] LayerConfig table — each layer carrying its inline AudioSpecificConfig() (parsed via the asc module's parse_bits / parse_bits_bounded entry points) or the resolved useSameConfig inheritance, the frameLengthType, and the type-0 latmBufferFullness / CELP-core coreFrameOffset or type-1 frameLength. The crcCheckSum is recomputed against the configuration prefix via the §1.8.4.5 CRC8 generator and validated. The reserved audioMuxVersionA == 1 branch and the CELP (3/4/5) / HVXC (6/7) frameLengthType values index frame-length tables for object types this AAC-focused crate does not decode, so they surface dedicated errors.

• AudioMuxElement() (latm::AudioMuxElement) — §1.7.3.1 Tables 1.41 / 1.44 / 1.45. Recovers a whole multiplexed element: the muxConfigPresent useSameStreamMux branch (inline StreamMuxConfig() vs. inherited previous config), the per-subframe PayloadLengthInfo() + PayloadMux() loop over numSubFrames + 1 frames (both the allStreamsSameTimeFraming program/layer walk and the numChunk chunk layout with its streamIndx + AuEndFlag), the frameLengthType-0 MuxSlotLengthBytes 8-bit-escape byte count and the frameLengthType-1 fixed (frameLength + 20) * 8 bits, the otherData skip, and the trailing ByteAlign(). Each access unit is returned as a MuxPayload carrying the raw §4.4.2.1 raw_data_block() bytes.

• AudioSyncStream() / EPAudioSyncStream() (latm::AudioSyncStream, latm::EpAudioSyncHeader) — §1.7.2.1 Tables 1.36 / 1.37. AudioSyncStream scans a LOAS byte buffer for the 11-bit 0x2B7 syncword, reads the 13-bit audioMuxLengthBytes, and decodes the byte-aligned AudioMuxElement(1) body, exposing an Iterator of LoasFrames with the StreamMuxConfig threaded across frames for useSameStreamMux inheritance. EpAudioSyncHeader decodes the EPAudioSyncStream FEC header (0x4DE1 syncword, futureUse, audioMuxLengthBytes, frameCounter, headerParity) and reports the byte-aligned EPMuxElement body offset.

• LoasDecoder (latm::LoasDecoder) — the end-to-end LATM/LOAS → PCM driver. decode_all walks the AudioSyncStream, and for every recovered MuxPayload drives the payload's §4.4.2.1 raw_data_block() through the shared decode::StreamDecoder::decode_raw_data_block core, configuring the decode from the layer's AudioSpecificConfig (AOT / samplingFrequencyIndex / resolved sample rate). One StreamDecoder is held per streamID[prog][lay] so each multiplexed stream's §4.6.11 overlap / §4.6.7 LTP / §4.6.6 predictor state threads independently. An SBR-signalling ASC (explicit AOT-5 wrapper or implicit AAC-LC-only) rides the same §4.6.18 auto-detect the ADTS path uses and emits dual-rate output, and a PS payload synthesizes stereo through the Annex 8.A tool. Pinned against the aac-latm-stream fixture (stereo, 44.1 kHz) to a §8 PCM-RMS error ratio of 0.0004, proven bit-identical to a hand-fed decode_raw_data_block pass, and — for a re-multiplexed HE-AAC v1 stream (both signalling modes) — byte-identical to the ADTS decode.

The runtime Decoder (codec_decoder::AacDecoder) auto-detects its carrier on the first packet and routes LOAS packets through LoasDecoder (see "Runtime Decoder registration" below). The EPMuxElement() EP-tool payload de-interleave decodes via LoasDecoder::decode_all_ep (see the EP section below).

Stream decode + PCM output

• Integer-PCM rendering (pcm) — §4.6.11 filterbank f64 time signal → 16-bit signed PCM: nint (the §1.3 NINT() round-half- away-from-zero operator), to_s16 (round + saturate), channel_to_s16, and interleave_s16 (element-order interleave). The conversion is the only output-rendering step (no resampler / dither), so it is fully spec-determined. The canonical channel reorder for default channelConfiguration layouts (Table 1.19, see channel_map below) is applied to the per-channel buffers before this interleave.
• Default-config channel reorder (channel_map) — ISO/IEC 14496-3 §1.6.3.5 / Table 1.19. A raw_data_block() lists its channel elements in bitstream order, so the decoder produces channels in element order (e.g. a 5.1 stream as C, L, R, Ls, Rs, LFE for SCE, CPE, CPE, LFE); channel_map::reorder_channels permutes them into the canonical interleaved order that oxideav_core::ChannelLayout adopts (the WAVE_FORMAT_EXTENSIBLE / BS.775 convention — 5.1 becomes L, R, C, LFE, Ls, Rs). The driver threads the signalled channelConfiguration through decode_raw_data_block and applies the reorder for every default config 1–7 — mono / stereo are identity permutations and config 7 (the Table 1.19 7.1 arrangement: centre + inner Lc/Rc centre-front pair + outer L/R front pair + surround pair
  • LFE) rank-sorts to L, R, C, LFE, Lc, Rc, Ls, Rs. Config 0 (PCE-defined) layouts are also mapped: channel_map::pce_speaker_assignment implements the ISO/IEC 13818-7 §8.5.2.2 rules — the front list center-outward (lone SCE = center, SCE pairs L-then-R, two front pairs = the Table 42 inner Lc/Rc + outer L/R arrangement), the side list front to back, the back list outside-in (outer pair = side surround, inner = rear; a final unpaired SCE = rear center), one LFE — keyed by (element kind, instance tag) so the block's element order is irrelevant; unmappable shapes fall back to element order. The decode driver captures an in-band PCE (§8.5.2.2 persistence), StreamDecoder::set_program_config installs an out-of-band (ASC-inline) one, and the LATM driver does so automatically. The whole path is validated end to end in tests/multichannel_mp4.rs (a minimal ISO 14496-12 sample-table + esds walk): the 5.1 config-6 fixture at 2.4e-4–8.9e-4 per-channel err/sig RMS, the 7.1 PCE fixture at 2e-5–2.9e-4, and the hexagonal custom 6.0 PCE fixture at 2.2e-4–7.8e-4 — every speaker carries a distinct source tone, so the per-channel ratios pin the mapping (silent LFEs reproduced exactly).
• Stream-level ADTS decode driver (decode) — StreamDecoder walks the §4.4.2.1 raw_data_block() of each ADTS frame above the per-element driver, keying one ElementDecoder per (syntactic-element-id, element_instance_tag) slot so every element's §4.6.11 overlap / §4.6.7 LTP / §4.6.6 predictor state persists across frames, and renders to element-order interleaved s16 PCM. decode_all walks a whole raw-ADTS buffer (ID3v2-skip + aac_frame_length framing). The decoded PCM is validated against the staged expected.wav corpus: the two PNS-free ADTS fixtures (aac-lc-mono-8000-16kbps-adts, aac-lc-intensity-stereo) are 99.9% byte-exact to the reference s16 output with a max error of 1 LSB — the residual is purely the difference between this crate's f64 direct-sum IMDCT and a float32 fast transform. The PNS-bearing fixtures are compared in the PCM RMS domain (per the fixtures-doc §8), where the error-to-signal RMS ratio stays below 0.1%; full byte-exactness on those is precluded by the §4.6.13.3 spec-undefined noise-phase RNG (energy is normative, phase is not). A coupling_channel_element() (CCE) carried in the block is decoded and applied: the block walk is two-pass, so the §4.6.8.3.3 coupling contribution lands on the addressed SCE / CPE channels at the signalled cc_domain stage whether the CCE precedes or follows its targets (see the cce bullet above). Multi-raw_data_block frames decode as consecutive 1024-sample blocks (per-block channel render + time concatenation), and decode_adts_frame verifies the whole §8.1.1 error_check() CRC layer when protection_absent == 0 (see adts_crc above).
• Runtime Decoder registration (codec_decoder) — AacDecoder adapts the persistent StreamDecoder / LoasDecoder into the framework's packet-in / frame-out oxideav_core::Decoder trait. It auto-detects the carrier on the first packet — the 0xFFF ADTS syncword vs. the 0x2B7 LOAS AudioSyncStream syncword — and then routes every later packet the same way: ADTS frames through StreamDecoder::decode_frame (ID3v2-skip + aac_frame_length framing; one or many frames per packet), LOAS packets through LoasDecoder::decode_all (one or many sync frames per packet, with the StreamMuxConfig and per-stream state threaded across packets). receive_frame returns one interleaved-S16 AudioFrame (1024 samples/channel) per decoded access unit, flush drains to Eof, and reset drops both backends and re-arms carrier detection for a clean post-seek restart. register() installs it under id "aac", claiming the MP4 object-type 0x40, WAVEFORMATEX 0x00FF / 0x1601, the mp4a / aac FourCCs, and the Matroska A_AAC CodecID; the probe scores a structurally-confirmed ADTS header at 1.0 and a bare LOAS syncword at 0.9 to win shared tags. Both carrier outputs are pinned byte-identical to their underlying StreamDecoder / LoasDecoder.

AAC-LC encoder

• encoder — the §4.5/§4.6-written-forward AAC-LC encode chain. StreamEncoder consumes interleaved S16 PCM hop by hop (encode_frame / finish / one-shot encode_all) and emits one complete ADTS frame per 1024-sample hop with a 1024-sample encoder delay. Per hop: an energy-jump transient detector over the [hist | cur] subblock grid drives the §4.6.11.3.2 ONLY_LONG → LONG_START → EIGHT_SHORT → LONG_STOP state machine; the §4.6.11.3.1 forward MDCT runs under the same composite windows the decoder synthesizes with (one 2048-point transform for long sequences, eight 256-point transforms at 448 + j·128 for short); EIGHT_SHORT frames merge envelope-alike adjacent windows into §4.5.2.3.4 window groups (one scalefactor/section track per group, §4.5.2.3.5 interleaved transmission order; the emitted 7-bit mask is pinned as the exact inverse of the decoder-side derivation); per-band scalefactors follow a masking-spread rule (band target magnitude 42·(peak_b/peak_frame)^½, sub-step bands culled to ZERO_HCB) with the DPCM ±60 track threaded across window groups; a bidirectional rate loop (±4 scalefactor ladder + ±1..3 fine pass) fits each frame to the bitrate-derived byte budget; codebooks and sections come from a measured-bit-cost dynamic program (every candidate same-class run priced at its section_data() header overhead plus the cheapest Table 4.95 book, actual codeword + sign + escape bits measured with the real tuple writer — pinned never-larger than the classic smallest-LAV + merge rule), and long frames additionally price a §4.4.6.3 pulse_data() variant (band outliers reduced to the rest-of-band floor, restored bit-exactly by the decoder's §4.6.3.3 fix-up; kept only when the measured channel stream is smaller). Stereo pairs code per-band §4.6.8.1 M/S (m=(l+r)/2, s=(l−r)/2 where the transform concentrates band energy, emitted as ms_mask_present 2 / 1+mask / 0) inside a common_window CPE — long frames per sfb, short frames per (window group, sfb) under the pair's joint grouping (decided once on the pair envelope; independent decisions would desync the shared ics_info). Every Table 1.19 default channel layout encodes — 1–6 and 8 (7.1) channels: the element-plan raw_data_block() assembly (SCE / common_window CPE / LFE with per-kind instance tags), canonical-order input permuted by the exact inverse of the decoder's §1.6.3.5 reorder, and §4.5.2.1.3-conforming LFE elements (always ONLY_LONG / sine through the frame's block switching, no TNS, only the lowest 12 spectral lines transmitted). Round-trips through the crate's own decoder: multitone 128 kbps at 0.016 err/sig RMS, staged-fixture transcodes at 0.0008–0.003, identical-channel stereo at 1.02× the mono stream size (short-run identical channels decode L exactly equal to R), multichannel layouts pinned with one distinct tone per speaker, and the wire-level window-sequence walk pins the exact LongStart → EightShort → LongStop pattern around a percussive burst; a deterministic bit-flip/truncation battery covers the multichannel and grouped-short streams.
• codec_encoder — the frame-in / packet-out oxideav_core::Encoder adaptor (make_encoder, honouring sample_rate / channels (1–6, 8) / bit_rate, default 64 kbps/channel); registered alongside the decoder under id "aac", and re-exported as encoder::make_encoder per the workspace dual-API convention.

PCE-described layouts (channelConfiguration = 0) in the encoder

• StreamEncoder::with_layout — any distinct speaker set on the canonical WAVE/BS.775 rank list becomes a §8.5.2.2 program_config_element() (front list centre-outwards with the inner Lc/Rc pair before L/R; Ls/Rs in the side list when a back pair or Cs is present, else as the sole back pair — the Table 1.19 5.1 reading; Lb/Rb then Cs in the back list; one LFE), instance tags counting per element kind in bitstream order, written at the head of every raw_data_block() under an ADTS channelConfiguration = 0. The input stays in canonical order — the order the decoder's PCE reorder emits — so encode ∘ decode is channel-identity: 2.1 / quad / 4.1 / 6.0 / 6.1 / 7.0 / side+back 7.1 / an 11-speaker cinema set all round-trip at ≈ 0.01 err/sig per channel through the crate's decoder, and the reference decoder binary accepts the streams at their channel counts. Sets the lists cannot express (an unpaired L, Lc/Rc without L/R, Lb/Rb alone — a lone back pair reads as the side pair) are rejected.
• Registry: make_encoder sends a named non-default CodecParameters::channel_layout (and the bare 7-channel count, as 6.1) through the PCE path; default sets keep their Table 1.19 configuration.

HE-AAC v1 (SBR) encoder — Annex 4.B.18 over the §4.6.18 decoder

• sbr_qmf::EncoderAnalysisQmf — the Figure 4.B.16 64-band complex encoder analysis bank (full-tap prototype, 128-element polyphase sum), pinned scale- and slot-aligned against the decoder's 32-band bank on decimated twins.
• sbr_writer — the forward SBR bitstream: sbr_header(), all four sbr_grid() frame classes, dtdf / invf / envelope / noise with the Annex 4.A.6.1 codebook lookup (inverse of sbr_huff_dec), sinusoidal coding, both sbr_channel_pair_element coupling layouts, and the whole extension_payload() body as EXT_SBR_DATA or EXT_SBR_DATA_CRC (10-bit G10 over the padded §4.5.2.8.1 region). Every writer round-trips bit-exactly through the crate's decode-side parsers.
• sbr_encoder — Annex 4.B.18 written forward: transient-driven grid election (FIXFIX 1/2/4; FIXVAR / VARFIX attack borders with bs_pointer = lA and cross-frame border continuity; VARVAR for a late onset after a variable lead — the trailing border pushed past a full two-slot attack envelope — and for two onsets in one frame, the §4.B.18.3 five-envelope grid with an envelope on each, lA on the first, both bs_num_rel_* sides within their 2-bit counts; three onsets go through a general VARVAR solver that searches every split of the wanted borders between the leading- and trailing-relative sides, every trailing border and every one-slot border adjustment — onsets only earlier — for the cheapest even 2..=8 segmentation within the §4.6.18.3.6 five-envelope limit, weighing a two-slot first-attack envelope against border movement), per-envelope bs_freq_res election (the high table for envelopes of four or more slots whose low-band children differ by more than 6 dB — the two-slot attack envelopes and flat spectra stay low), §4.B.18.4 envelope energies, §4.B.18.6 quantisers, §4.B.18.7 time/frequency delta election closed-loop over the decoder's own §4.6.18.3.5 reconstruction (LAV-clamped, header frames freq-forced for random access), and coupled level/balance coding. The noise floor and bs_invf_mode come from a forward-prediction tonality contrast between each band and its Figure 4.48 patch source, plus an analysis-by-synthesis of the decoder's §4.6.18.7.5 limiter: energy the capped gains cannot deliver through the copy is re-routed to the noise floor, peaked tonal bands the patch cannot supply become bs_add_harmonic sinusoids (default-on detector), and envelopes pre-compensate the injected-noise path's measured half-energy synthesis loss.
• he_aac_encoder — the Figure 4.B.15 chain: crossover-cut FIR downsampler → half-rate LC core, group-delay-aligned continuous analysis ring sliced per frame (the decoder's tHFGen/tHFAdj column alignment), SBR payload charged inside the frame budget and emitted as the fill element after its SCE/CPE — implicit §4.5.2.8.1 ADTS signalling. Explicit signalling ships too: asc_writer (the backward-compatible syncExtensionType 0x2b7 trailer and the hierarchical AOT-5 form, both reparsed by the crate's ASC parser) and latm_writer (LOAS AudioSyncStream frames at audioMuxVersion = 1 so the length-prefixed ASC keeps the trailer probe unambiguous).
• Validation: encode → the crate's own SBR decoder measures 2.8–3.4 dB mean per-envelope-band energy error on an adversarial multitone+noise synthetic (HF total within 0.4 dB), 0.9 dB HF ratio on staged-fixture PCM transcodes, 22–43 dB low-band SNR; the reference decoder binary (skip-if-absent) decodes every stream without diagnostics at the doubled rate and its PCM agrees with this crate's decoder to 0.00 dB mean per-QMF-band energy; LOAS-wrapped output decodes identically to the ADTS path.
• Registry: codec_encoder::make_he_aac_encoder, and make_encoder dispatches on an SBR-signalling extradata AudioSpecificConfig (output params advertise the backward-compatible ASC).

HE-AAC v2 (parametric stereo) encoder — Annex 8.C.6 over the §8.6.4 decoder

• ps_analysis — the stereo front end: both input channels through the decoder's own §8.6.4.3 hybrid filterbank (71 / 91 sub-subbands) on the Annex 8.A.3 Xinput framing, and the Annex 8.C.6.2 band excitations e_l / e_r / e_R / e_O over the Table 8.C.2 / 8.C.3 summation ranges (10-band ranges folded through Table 8.45) — pinned against the decoder's Table 8.48/8.49 b(k) maps (every range inside one band, no conjugate channels).
• ps_encoder — parameter positions (§8.6.4.4): FIX_BORDERS with 1 / 2 / 4 envelopes elected by quarter-frame cue variation beyond its own estimation noise (6.1/√N dB, 1/√N coherence, three sigma), VAR_BORDERS [t−1, t, 31] on an attack whose stereo cues change, the four-bit num_env = 0 hold for a repeated set, and cross-frame smoothing of narrow-band excitations. IID / ICC / IPD / OPD from the excitations (real coherence for mixing procedure Ra — the cross-correlation the decoder's output reproduces — magnitude coherence for the phase layer's Rb, phases only where coherence ≥ 0.6), quantised on Tables 8.25 / 8.26, 8.28 and 8.31; per-row time-versus-frequency differential coding by measured Annex 8.B codebook bits (header frames frequency-forced on envelope 0 for random access, phase deltas modulo 8), with the decoder's own PsData::resolve run against a mirrored index state to assert every frame lands on the intended indices. Header-level tools are elected from the measured image (Election::Auto, the default for both): the fine IID grid (iid_mode + 3) when the energy-weighted error of the coarse Table 8.25 grid against the measured level differences exceeds 0.75 dB over the frames since the last header (hard pans sit on the coarse grid's 3–7 dB steps, near-centre images do not), and the phase layer (enable_ext, ICC as magnitude coherence, mixing procedure Rb) when at least 10 % of the IPD/OPD bands' energy sits in bands coherent above 0.6 with an inter-channel phase beyond half a Table 8.31 step; a decision taken over silence (the encoder's priming frame) is provisional and the first frame with signal re-elects with its own header. Inside the phase layer enable_ipdopd is sent per frame only when some band carrying energy has a non-zero quantised IPD or OPD (a hold element repeats the previous decision so the decoder keeps the same H vectors); the IPD/OPD band count is Table 8.24's nr_ipdopd_par_tab[iid_mode] (5 / 11 / 17), not a separate election.
• ps_writer — the forward ps_data() (Tables 8.9–8.14): header block, Table 8.29 inversion, VAR borders, Huffman-coded rows on the ten codebooks (inverse of ps_huff_dec), and the byte-counted extension layer sized so fewer than eight fill bits follow it. Every element reparses bit-exactly through the crate's parser (all IID × ICC modes, both frame classes, the phase layer, the hold element).
• he_aac_encoder (HE-AAC v2 mode) — a stereo input coded as one SCE plus ps_data() under EXTENSION_ID_PS in the SBR payload: the Annex 8.C.6.1 downmix formed in the QMF domain with a per-band √((e_l + e_r)/(2·e_m)) gain (capped +12 dB, interpolated across the hop) restoring the pair energy the decoder's energy-preserving mixing expects, re-synthesised through the §4.6.18.4.2 bank into the mono v1 path; the 576-sample QMF pair delay + FIR delay + a 32-sample pad make the stereo→mono lag exactly 10 columns so the PS encoder sees the decoder's Xinput slots for each frame's core block. Implicit ADTS signalling (channel_configuration = 1); asc_writer::he_aac_v2_asc writes both §1.6.6 forms (the 0x2b7 + 0x548 psPresentFlag trailer, 49 bits, and hierarchical AOT 29, 25 bits) and latm_writer carries either.
• Validation (own subpart-8 decoder as oracle, band-wise multi-sine stereo with per-region pan and coherence): per-frame per-band IID within 1.5 dB mean and ICC within 0.11 on the default configuration (20 bands, coarse grid, Ra) at ~1.7 kbps of PS side info; every band count × grid × phase-layer combination under 2 dB / 0.12; a hard left→right switch re-appears within 4 slots of the audio; the phase layer reproduces an eight-sample inter-channel delay's per-band IPD to 0.13 rad mean (0.83 rad on the phase-less Ra control) — and the Auto election reaches the same 0.13 rad on that pair while staying on Ra for the phase-less panned synthetic; LOAS wrapping with either v2 ASC decodes identically to ADTS as stereo. The reference decoder binary (skip-if-absent) decodes every stream without diagnostics as stereo at the doubled rate, its long-term per-band IID agreeing with this crate's decoder to 0.0–0.5 dB mean, and on the auto-elected phase layer its per-band IPD agrees with this crate's decode to 0.000 rad (both 0.127 rad from the source). Pure tones expose the parametric model's own limit (an all-pass de-correlator turns a sinusoid into a phase-shifted copy), not the encoder's.
• Registry: codec_encoder::make_he_aac_v2_encoder, and make_encoder dispatches on a PS-signalling extradata ASC (AOT 29 or psPresentFlag).
• Fuzz: fuzz/ carries ps_data_roundtrip (parse → write → reparse identity + build-first recipes), ps_encoder_frame (arbitrary, non-finite QMF input through the encoder and back through the parser; the first run found and fixed a parser quirk — a repeated ps_extension_id = 0 block left stale phase rows) and sbr_encoder_frame (every SBR encoder configuration — both core families, mono / coupled pairs, every header option — over arbitrary analysis frames, each payload reparsed to the encoder's element with a decoder-derived grid; its first run found the coupled balance quantiser adding panOffset to a saturated integer cast of a non-finite ratio — fixed).

ER BSAC (AOT 22) — noiseless-coder bring-up (§4.4.2.6 / §4.5.2.6 / §4.6.4)

The Bit-Sliced Arithmetic Coding decoder roster is implemented and its front half is conformance-pinned against the ISO/IEC 14496-26 er_bs* corpus; the spectral bit-slice probability selection of the deployed encoder diverges from the printed spec and is the component still open (see below):

• Numeric tables (bsac_tables) — Tables 4.A.31–4.A.77 transcribed from the staged spec PDF: the cband_si_type parameter matrix, the scalefactor / cband_si / stereo / PNS cumulative-frequency models, the Table 4.A.34 context-position map, the Table 4.A.35/36 min_p0/max_p0 budget clamps, and the 22 spectral probability tables with the printed alias scheme resolved. (The 2001 and 2009 editions print different alias schemes — tables 11–22 onto 9/10 alternating with sub-MSB zero rows from 7/8 in 2009, everything onto 10 with zero rows from 8 in 2001 — plus one conflicting cell in table 7; both were transcribed and tested.)
• Arithmetic decoder (bsac_arith) — the §4.5.2.6.2.7.4 procedure exactly as listed (decode_symbol over 14-bit cumfreq models, binary decode_bit, the half[] renorm schedule, 30-bit init, zero-stuffing segment reader), round-tripped against a spec-inverse test encoder over every model.
• Layer geometry (bsac_layer) — the §4.5.2.6.2.4/5 roster: base sub-layer split, per-layer coding-band / spectral / sfb coverage (the literal end_sfb = sfb + 1 one-band lookahead, corpus-confirmed), layer_si_maxlen, the rate-anchored layer_bit_offset derivation with the overflow/underflow redistribution, and the SBA terminal_layer marks.
• Block decode + reconstruction (bsac_decode) — the bsac_header() / general_header() raw-bit parse, the full layer walk (side info, first-pass spectra, the bsac_lower_spectra() refinement, budget carry between layers), bit-slice reassembly with interleaved sign decode, and the AAC back end (§4.6.2 dequant, group de-interleave, §4.6.8 stereo hooks, §4.6.9 TNS, §4.6.11 filterbank) behind a persistent BsacDecoder.
• What the corpus pins (tests/bsac_bringup.rs, corpus-gated): on er_bs01_48_ep0 the silent access units decode sample-exact against the reference waveform (headers, layer roster, arithmetic side-info decode all in sync), and on content frames a TDAC oracle (the §4.6.11 perfect-reconstruction property recovers each frame's exact transmitted spectrum from the reference PCM) confirms the decoded cband_si MSB plane and scalefactor gains match the deployed encoder precisely.
• The open divergence: the §4.6.4.2.3 spectral-bit probability selection as printed decodes the wrong sliced bits partway into the first coding band (both editions' alias readings, several structural variants, and every printed row under every position mapping tried were tested against the oracle truth). A constraint solver over the real stream proves a consistent context→p0 dictionary exists — the symbol order, sign interleave and context classes are right — but its values match no printed row (the all-zero-context position demands p0 ∈ {0x3700..=0x3a00}; every plausible row prints 0x3b00+ there). A clean-room behavioural trace of the deployed p0 selection (or the corrigendum text) is the standing docs ask; tests/bsac_bringup.rs carries the divergence locator and the solver instrument, and tests/iso_14496_26_conformance.rs reports the structural decode rate (618/703 AUs on er_bs01_48_ep0) without asserting PCM until the rule lands.

Not yet supported

• The deployed ER AAC LD tns_data() filter record. The ISO/IEC 14496-26 LD conformance bitstreams transmit an extra-spec TNS record: the corpus-resolved 1-bit-n_filt reading (docs/audio/aac/er-ld-tns-divergence.md §0, implemented here) reconciles the structure — every AU parses to its boundary — but the reference waveforms show the record carries a real variable-length filter (per-AU record lengths ≈ 19–61 bits, in 3-bit increments) whose layout matches no Table 4.54/4.155 reading (a grammar search over length 4/6 × order 3/5 × 1–2 filters × optional direction/compress fields, decoded and applied, reconciles none of it). TNS-bearing LD AUs (~6 % of the er_ad1103* family) therefore decode with wrong PCM until a behavioural trace of the deployed record lands; the conformance harness masks them (and bounds the LTP-setup vectors coarsely, since LD LTP history includes those AUs).
• Encoder-side tool remainders — the end-to-end AAC-LC encoder (see encoder below) covers block switching with §4.5.2.3.4 short-frame grouping, M/S on both frame shapes, the scalefactor/quantizer rate loop with measured-bit-cost codebook/section choice, a rate-derived coded bandwidth (EncoderConfig::default_bandwidth_hz = 0.26 Hz per bit/s of per-channel rate, 4–20 kHz; StreamEncoder::set_bandwidth), every Table 1.19 default channel layout, and opt-in §4.6.13 PNS emission (StreamEncoder::set_pns — off by default because a single-frame spectral statistic cannot tell true noise from noise-shaped deterministic content such as sweeps; default-on awaits a cross-frame tonality measure) and default-on §4.6.9 TNS emission (encoder_tns: per-window Levinson-Durbin prediction-gain decision under a time-domain temporal-envelope gate, PARCOR quantised on the §4.6.9.3 4-bit arcsine grid, and the §4.6.7.4.1 all-zero analysis pass derived from the wire coefficients so the decoder's all-pole synthesis is its exact inverse) and opt-in §4.6.8.2 intensity-stereo emission (StreamEncoder::set_intensity_stereo — correlated high bands transmitted once with codebook 15/14 + is_pos on the §4.6.8.1.4 track; off by default because intensity coding discards the pair's side information) and measured §4.4.6.3 pulse_data() emission (long-frame outlier-over-floor bands, kept only when the whole channel stream prices smaller; the decode-side fix-up restores the identical quantized spectrum), keeps PNS and IS long-frame-only (CPE PNS emits the §4.6.13.3 ms_used correlated-noise signalling — shared random vector — for both-channels-noise bands correlating above 0.5). The HE-AAC v1 (SBR) encoder and the HE-AAC v2 (parametric stereo) encoder are implemented end to end (see their sections above: analysis banks, grid/envelope/noise estimation, coupling, CRC payloads, the Annex 8.C.6 stereo parameter estimation and ps_data() writer, implicit + explicit signalling, LOAS carriage, registry profiles). Custom speaker sets encode through a §8.5.2.2 program_config_element() (StreamEncoder::with_layout, channelConfiguration = 0: 2.1, quad, 4.1, 6.x, 7.0, side+back 7.1, cinema sets with both front pairs and a rear centre — every set the PCE lists and the decoder's speaker assignment can express; a lone Lb/Rb pair, an unpaired L, or Lc/Rc without L/R are rejected up front), the PCE opening every raw_data_block(); the registry maps a named non-default channel_layout (and the bare 7-channel count, as 6.1) onto it. Beyond the 11 positions of that set (the 10.1 "cinema" layout: C, Lc/Rc, L/R, Ls/Rs, Lb/Rb, Cs, LFE — black-box decoded by the reference binary) the staged 2009 edition has no height layer in program_config_element() and Table 1.19 ends at channelConfiguration = 7 — 22.2 and the top-layer speaker sets are out of its reach (see the docs asks). Equal-rate harness (tests/encoder_psy_harness.rs, black-box reference encoders for AAC-LC at 64/96/128 kbps stereo and HE-AAC at 32/48 kbps): every stream — ours and the reference encoder's — decoded by this crate and by the reference decoder binary (per-band parity 0.00 dB on ours, 0.03–0.06 dB mean on theirs), the coding noise per 64-band QMF band against the input after lag alignment (NSR, and a noise-to-mask ratio under a 10 dB / 12 dB- per-band spreading masker). The measurement's largest lever was bandwidth: at 32 kbps per channel the square-root masking spread bought ~0 dB NSR in every band above 8 kHz while the bands below 4 kHz sat 8 dB noisier than the reference's, which zeroes the top of the spectrum; the rate-derived cull (slope swept over 0.20/0.26/0.32 Hz per bit/s — 0.26 matches the reference's cutoffs and wins the 64 kbps case) now puts the mean NMR at −0.2 / −0.9 / −2.0 dB against the reference on the mixed material at 64/96/128 kbps (from +1.1 / +0.1 / −2.1) and −0.1 / +1.6 / +1.6 dB on the transient material (from +0.8 / +3.0 / +2.6). The harness's second finding was the HE-AAC rate: the core's rate loop subtracted the SBR fill bytes from the frame budget and measured them inside the assembled block, charging them twice, so HE-AAC streams ran 16–27 % under target (26.8 / 43.9 kbps for 32 / 48); priced once, they land on it (31.9 / 48.4 kbps). Its third: the SBR crossover — the reference HE encoder still waveform-codes up to ≈ 5.9 / 8.3 kHz at 16 / 24 kbps per channel (−16 dB NSR there) where the old 0.115·fs·√(bps/24 000) rule (4.1 / 5.1 kHz) had already handed the bands to SBR (+3 dB); the crossover now follows 0.35 Hz per bit/s of core-channel rate (5.6 / 8.4 kHz, clamped to [0.09, 0.22]·fs), taking the mean NMR gap to +1.9 / +2.4 dB (from +2.8 / +4.1 at the round's start; the remaining bands are SBR-regenerated on both sides, where a waveform measure no longer discriminates). The SBR grid election spans all four frame classes (VARVAR up to the five-envelope limit for late, double and triple onsets — a fourth onset in one frame falls into the last envelope) with per-envelope bs_freq_res; the noise-floor count and split are not elections (Table 4.69 derives bs_num_noise from the envelope count, and Table 4.174 pins the split at lA for a transient frame).
• SSR remainders — the §4.6.12 gain-control tool is now implemented and wired end to end (front-half filterbank, gain reconstruction, IPQF — see the "SSR gain control" section above), and a writer-assembled AOT-3 fixture driving non-unity gain ladders through all four window sequences (with the §4.6.12.3.3 variable 1024/1472/576 frame lengths) is staged in the docs corpus (aac-ssr-gain-control-adts; no encoder for AOT 3 exists anywhere, so a captured conformance stream remains welcome — a black-box validator binary reports SSR gain control unimplemented, so there is no external oracle for the ladders). Still open: the 13818-7 SSR-profile bandwidth-scalable output modes (decoding only 1–3 PQF bands at a reduced rate) are not selectable — the decoder always reconstructs the full-rate signal. (The Main frequency-domain predictor, §4.6.6, is now fully wired into element_decode for the AAC Main object type on long windows — see predictor above. LTP, §4.6.7, is likewise wired in with the §4.6.7.4.1 / Figure 4.30 TNS-analysis-in-loop ordering. The ISO/IEC 14496-3:2001 Table 4.55 short-window LTP syntax — the per-short-window ltp_short_used / ltp_short_lag_present / ltp_short_lag loop that the 2009 edition removed ("LTP is restricted to long windows only", §4.6.7.1 2009) — is parsed and re-encoded under the explicit LtpEdition::Iso2001 selector (parse_ltp_data_edition / write_ltp_data_edition; the two editions are wire-incompatible there and nothing in-band signals which one a stream follows). The short-window synthesis stays unimplemented: the 2001 §4.6.7.3 text defines the x_rec buffer arrangement once but never fixes the per-subframe index origin for the eight windows, and no LTP fixture exists to disambiguate. The ER AAC LD long-window LTP — 10-bit lag, ltp_lag_update repeat, M = N/2 — is implemented and exercised by the staged LD fixtures.)
• SBR/PS remainders — the §4.6.18 SBR tool and the subpart-8 PS tool are implemented end to end (see the sections above) and wired into the ADTS / LATM StreamDecoder paths and the runtime Decoder, validated against the HE-AAC v1 (99.98% sample-exact) and HE-AAC v2 (5e-5 RMS) fixtures, with the 10-bit bs_sbr_crc_bits of every EXT_SBR_DATA_CRC payload now verified (§4.4.2.8.1 G10, zero init, over the Table 4.62 region — see adts_crc; a derived type-14 fixture is staged as he-aac-v1-sbrcrc-adts). The §4.6.18.4.3 downsampled-output mode and the §4.6.18.8 low-power variant are now both selectable end to end (see the SBR back-end section above), and the decoder runs over both the 1024-line (16-slot) and 960-line (15-slot) cores. Still open: an SBR payload on an LD stream is rejected (Error::SbrUnsupportedFrameFamily; the §4.6.19 LD SBR tool is ELD's and stays out of scope) — and low-power PS is undefined by design (the subpart-8 tool needs the complex QMF domain, so LP + PS is rejected). The coupling-channel (CCE) tool is decoded and applied end to end (cce + the two-pass stream walk; see the tool-chain section above), validated against the filterbank-linearity identity on writer-assembled CCE streams; a third-party CCE-bearing conformance fixture would still be a welcome external cross-check.
• ER BSAC (AOT 22) PCM — the noiseless-coder front half (headers, layer geometry, arithmetic side-info decode) is conformance-pinned, but the deployed encoder's spectral bit-slice probability selection diverges from every reading of the printed §4.6.4.2.3 / Table 4.A.34 selection (see the BSAC section above), so reconstructed PCM does not yet match the reference waveforms. Also out of scope until then: SBA-mode segment scheduling (sba_mode == 1), BSAC LTP, BSAC PNS (the noise-energy PCM conventions need a working spectral decode to pin), the zero_code extended part (channel / SBR / MPEG-Surround extensions), and the §4.5.2.6.1 multi-ES bsac_payload() large-step-layer reassembly (the er_bs02-style carriage).
• Error-resilience remainders — the ER story is now wired end to end for ER AAC LC (AOT 17), ER AAC LTP (AOT 19) and ER AAC LD (AOT 23): the ER channel-element body (ics_body::IcsBody::parse_er) selects all three §4.4.6 resilience branches, the reordered_spectral_data() payload is decoded and encoded (hcr_decode), and the §4.4.2.3 Table 4.19 er_raw_data_block() driver (StreamDecoder::decode_er_raw_data_block) walks the fixed per-channelConfiguration element sequence for all three AOTs — reachable from LATM (the LOAS driver routes AOT-17/19/23 layers there with the ASC's resilience triplet and the ASC-resolved §4.5.1.1 frame family). AOT 17 is pinned bit-identical to the equivalent non-resilient decode of the same spectra (aac-er-hcr-loas); AOT 19 — the §4.6.7 LTP tool over the Table 4.19 walk (11-bit lag, M = 0, per-element x_rec history across frames) — is pinned bit-identical to the equivalent AOT-4 decode with LTP active (SCE and pair-LTP CPE, plain and HCR spectra); AOT 23 is pinned by the staged LD fixtures (see the frame-length families section above). ER AAC scalable (AOT 20) now decodes end to end (see the scalable section above), and the §1.8 EP tool — ErrorProtectionSpecificConfig(), SRCPC / RS / interleaving, the ep_frame() codec and the LOAS EPMuxElement / EPAudioSyncStream carrier — is implemented (see the EP section above). Still open: the §4.5.2.4 Table 4.148/4.149 per-element category split of the codec payloads themselves (reassembling an er_raw_data_block whose bits arrive as separate error-sensitivity-category instances under epConfig == 1 / directMapping — the ep_frame class concatenation covers the in-order case), and an encoder-produced HCR conformance stream as an external cross-check.
• LATM/LOAS transport framing (§1.7) — the StreamMuxConfig(), AudioMuxElement(), PayloadLengthInfo(), PayloadMux(), LatmGetValue(), AudioSyncStream() and EPAudioSyncStream() bitstream walkers are now implemented and tested (see the LATM / LOAS transport section above), with the crcCheckSum recomputed against the §1.8.4.5 CRC8 generator in the crc module. The runtime Decoder LOAS entry point that routes the recovered MuxPayload raw-data-blocks into a StreamDecoder is now wired (latm::LoasDecoder + the codec_decoder carrier auto-detection above). The EPMuxElement() EP-tool payload de-interleave is now implemented (LoasDecoder::decode_all_ep; see the EP section above). (ADTS adts_error_check() CRC validation — the 192/128-bit region selection with the double-protection edge cases plus the ISO/IEC 11172-3 §2.4.3.1 code — landed in the dedicated adts_crc module; see the bitstream-parsing section.)

License

MIT — see LICENSE.