XA-ADPCM streams Confirmed

Sector layout

Each on-disc XA sector is the standard PSX 2352-byte raw layout:

+0x000  12 B  sync (00 + 10x FF + 00)
+0x00C   4 B  header (MM SS FF mode)
+0x010   8 B  CD-XA subheader (4 fields + duplicated copy):
                  file_no, ch_no, submode, coding_info
+0x018  2304 B  user data (18 sound groups × 128 B audio)
              - 0x14 trailing bytes are padding
+0x92C   4 B  EDC

Submode bits relevant for audio detection:

Coding-info bits:

The 18 sound groups inside the user data are 128-byte CD-XA ADPCM blocks; see the lib doc-comment for the per-block layout (8 sound units, 28 lines × 4 bytes).

Sound-group decode (4-bit)

Each 128-byte sound group holds 8 sound units of 28 samples. The decode is bit-exact against an external lossless reference decode of a real cutscene track (every interleaved sample matches), so the layout below is confirmed, not inferred.

Parameter bytes (0..16). The redundant copy is interleaved within each half, not appended:

byte:  0  1  2  3   4  5  6  7   8  9 10 11  12 13 14 15
unit: p0 p1 p2 p3  p0 p1 p2 p3  p4 p5 p6 p7  p4 p5 p6 p7

So unit u's parameter byte is at u + (if u < 4 { 0 } else { 4 }). Each byte is (filter << 4) | range, filter ∈ 0..=3, range ∈ 0..=12. (Reading bytes 0..8 as eight sequential params - the "appended mirror" reading - mis-assigns the parameters of units 4..7 and is the classic CD-XA decode trap.)

Sample nibbles (16..128). 28 lines of 4 bytes. Unit u reads byte u / 2 of each line, taking the low nibble when u is even and the high nibble when u is odd:

line byte:  0           1           2           3
nibble:   lo=unit0     lo=unit2    lo=unit4    lo=unit6
          hi=unit1     hi=unit3    hi=unit5    hi=unit7

Per-sample reconstruction (filter coefficients in 1/64 units, k0 = {0, 0.9375, 1.796875, 1.53125}, k1 = {0, 0, -0.8125, -0.859375}):

shifted = (sign_extend(nibble, 4) << 12) >> range
value   = shifted + k0 * prev1 + k1 * prev2
output  = clip16(round_half_away_from_zero(value))
prev2   = prev1;  prev1 = value     // history is the UNCLAMPED, UNROUNDED value

The predictor history is the full-precision reconstructed value, not the rounded+clamped 16-bit output. Re-feeding the clamped output instead is audible only at high volume - the prediction drifts on loud sound-units and rails to the opposite extreme - which is exactly the symptom that bit-exact history feedback removes.

Stereo de-interleave. The LEFT channel is the even units (0,2,4,6) and the RIGHT channel is the odd units (1,3,5,7); output is L,R interleaved, pairing (0,1),(2,3),(4,5),(6,7). Each channel keeps its own (prev1, prev2) history.

"Non-standard interleave" - what it is and isn't

The earliest extracted-XA tooling truncated each on-disc sector to 2048 bytes (Form 1 mode), which silently:

1. Dropped 276 bytes per sector of audio (Form 2 user data is 2324 B vs Form 1's 2048 B).
2. Collapsed every channel of the multiplexed stream into a single shuffled byte sequence, because the per-sector (file_no, ch_no) subheader was discarded.

The result was a stream where only ~10 % of 128-byte sound groups passed the standard XA validation rule bytes 8..16 == bytes 0..8. That's the "non-standard interleave" shorthand from earlier doc revisions - not a bespoke Legaia muxing scheme, just Form-1-truncation damage.

The fix lives in crates/xa/src/demux.rs (function demux_disc_range). It reads raw 2352-byte sectors, parses each subheader, filters to AUDIO + FORM2, and splits the audio data into one buffer per (file_no, ch_no) tuple. After that step the per-channel buffer is a clean concatenation of standard 128-byte sound groups that the 4-bit ADPCM decoder handles directly.

The xa demux-disc-all subcommand drives this across the whole disc - it walks the ISO9660 tree, finds every *.XA, and demuxes each at its own per-sector sample rate / channel mode read from the subheaders (no guessed global rate):

./target/release/xa demux-disc-all \
    "/path/to/Legend of Legaia (USA).bin" \
    --out extracted/xa_demux

One WAV lands per (file_no, ch_no) channel under extracted/xa_demux/, named <xa-stem>_fileN_chM.wav. The single-file xa demux-disc --lba --size variant remains for targeting one entry by directory offset.

Pacing is therefore data-driven per track - the whole point. A track that varies channel mode (the NA disc has 16-channel mono voice files like XA4/XA6 alongside 8-channel stereo music like XA5/XA7/XA8/XA9) decodes each channel at its real width: the Form-1 convert path read a stereo track as mono and played it at 2× speed; the demux path reads coding_info and gets it right. Channels reporting a non-4-bit width are skipped with a warning rather than mis-decoded.

What the older extracted/XA/*.XA files contain

If the tree on disk has a populated extracted/XA/ directory from a pre-fix extraction, those files are the Form-1-truncated bytes - usable for byte-stable hashing only, not for decoding. They should be deleted and re-extracted via xa demux-disc-all (or via the top-level legaia-extract once it integrates the new path).

What's still open

• Top-level extract integration. legaia-extract doesn't yet call xa demux-disc-all. Cutscene audio works end-to-end via the standalone xa binary; the unified pipeline still emits the legacy Form-1 bytes.
• 8-bit ADPCM mode. The NA corpus is entirely 4-bit, 37.8 kHz (demux-disc-all reports bits_per_sample = 4 for all 316 channels across 34 *.XA files), so the unimplemented 8-bit group decoder doesn't block anything. The demuxer surfaces bits_per_sample and the CLI skips-and-warns on any non-4-bit channel rather than mis-decoding it.
• Per-cutscene file-no / ch-no map. demux-disc emits one WAV per channel keyed by (file_no, ch_no). The mapping from cutscene name → expected channel pair lives inside the cutscene-overlay's mode driver, which is not yet captured. Until that's reversed, the WAV → cutscene assignment is manual.

See also