Skip to content

MUSIC.md

Latest commit

 

History

History
189 lines (165 loc) · 10.1 KB

MUSIC.md

File metadata and controls

189 lines (165 loc) · 10.1 KB
Raw

The music of Xenosaga Episode I — how it works, and what the disc told us

Xenosaga Episode I does not stream its music. The orchestral score by Yasunori Mitsuda is sequenced: the disc carries note data (.SMD) and instrument banks of sampled orchestra (.SWD), and a driver on the PS2's sound processor plays them live, the way a MIDI file plays against a SoundFont. This kit turns those files back into MIDI, SoundFont 2 and rendered WAV/FLAC — and because Monolith shipped the driver with its debug symbols intact, everything below is read from the game's own code, not guessed.

python cli.py extract --iso GAME.iso --out OUTDIR         # once
python cli.py music-export --dump OUTDIR/dump --out OUTDIR/browse/music
python cli.py browse --out OUTDIR --kinds banks,battle_audio,audio
Output What it is
browse/music/*.mid each sequence as a standard MIDI file (96 PPQN, tempo map, loop markers)
browse/music/*.sf2 each bank as a SoundFont 2 — the game's own orchestra samples, ADSR and key splits
browse/music/*.wav / .flac rendered by the kit's SPU2-faithful sampler (needs numpy; FLAC needs ffmpeg)
browse/soundbanks/ every bank's samples as WAV + smd_catalog.csv (title, composer, notes per sequence)
browse/battle_audio/ the battle voice / sound-effect banks, 1,502 sample WAVs
browse/audio/ the streamed stereo voice/ambience/cutscene streams as WAV

The MIDI, SoundFonts and renders are derived from the disc's own data. They are for your own use; regenerate them from your own copy of the game rather than sharing them.

Three audio systems on one disc

  1. Sequenced BGM — 115 .SMD sequences and 289 .SWD banks under sound/smd, sound/sed and the battle folder yamamoto/snd. Played by SSD.IRX, Procyon Studio's sound driver running on the IOP.
  2. Streamed audio — 93 .vds files (92 under chain1/sound/vda, plus yajima/gameover.vds): headerless stereo SPU ADPCM at 48 kHz. Spoken lines, ambience beds, and the big pre-mixed music+voice tracks of in-engine cutscenes (up to 20 MB each).
  3. Movie audio — inside the .pss cutscene movies, a Sony ADS stream tucked into MPEG private stream 1. The kit demuxes it itself because ffmpeg sees phantom "mp2, 0 channels" streams there.

Only the first is music in the composer's sense. Of the 115 sequences, ten are real pieces: Battle1 (two copies), LastBattle, Escape! (three copies), U.M.N. Mode (two copies) and Jingle2 (two copies). The other 105 are 196-byte ambience stubs. That is the famous quiet of the field maps, visible in the data.

How a sequence plays

An .SMD is a stack of tracks, each an event stream the driver walks with a dispatch table (SsdSeqFuncTrap) of 128 opcode handlers.

  • Note-ons are tiny. Any byte below 0x80 is a note-on and the byte is the velocity. The next byte packs three things: how many duration bytes follow (or "reuse the last duration"), an octave step of −1 to +2 relative to a running octave, and the semitone. A melody costs about two bytes per note. This is why a full battle theme is 17 KB.
  • The clock is fixed at 96 ticks per quarter note. The driver arms a 2 ms interrupt; each firing subtracts a rate from a 16.16 fixed-point accumulator and advances one tick on underflow. The rate is tempo × 53687 >> 8, which works out to exactly 1.6 × tempo ticks per second — tempo × 96 / 60. So the tempo opcode's operand is literal BPM.
  • Looping is two opcodes. Repeat (0x91) bookmarks the stream position; Stop (0x90) jumps back to it forever. All 115 sequences loop this way; the loop length is simply the distance between them. The opcode named Jump is a forward skip, not the loop.
  • Everything else is what you would expect of a 2002 sequencer: rests and ties, nested count loops with a 12-byte stack frame per track, program change, absolute and relative volume/pan/expression, pitch bend and fine tune, key transpose, and a set of SMPTE/label opcodes the driver ignores at runtime.

How the instruments sound

An .SWD bank is a sample table plus programs.

  • Samples are SPU ADPCM, the PS2's native 16-byte-frame codec, with loop points carried in the stream's own frame flags. Each sample stores a base pitch in 1/256-semitone units relative to key 60 at 48 kHz, so its native rate is 48000 × 2^(pitch/3072). Retail banks decode to 22.05, 32, 44.1 and 48 kHz — plus deliberate detunes of a few cents (32,007 Hz-style rates are intentional, not errors).
  • Programs map key ranges to samples (up to 16 splits each) with a root key, volume, pan and two raw SPU2 ADSR registers. The SoundFont export carries all of it.
  • Pitch is exact equal temperament. The driver's pitch table (SsdAllPitchTable) matches 0x1000 × 2^((note − 60)/12) to the register value, so the SoundFont plays in tune with the game.
  • The driver squares the volume. The combined velocity × channel × expression × split × sample level is squared before it reaches the SPU volume register (SsdDeviceVoiceEvent). A renderer that applies it linearly sounds flat and crowded; the kit squares it.

How we cracked it

SSD.IRX ships unstripped. Every sequencer opcode handler has its symbol name, the dispatch table and the operand-length table are in .data, and the pitch and timer maths are a few dozen MIPS instructions each. The kit's ssd.py is a transcription of that driver's behaviour, not a statistical guess about the file format. extract --code drops the driver into browse/code/IOP/ if you want to read it yourself.

Two lessons cost real time and are worth passing on:

  • The header byte that is not a timebase. Byte 0x20 of every .SMD holds 100, 120, 123 or 127 — plausible PPQN values — and the first tooling used it as one. Battle themes came out 25 % fast. U.M.N. Mode, whose byte is 100, was only 4 % off, which masked the bug for weeks. The driver never reads that byte; the clock is fixed.
  • Advance by what the handler returns. The operand-length table disagrees with a few handlers about how many bytes they consume; the dispatcher trusts the handler. Trusting the table injected two phantom notes at tick 0 of every track (SMPTEOffset is 5 bytes, not 1).

The conversion was then put through a six-dimension adversarial audit (timing, pitch, envelopes, loop structure, bank carving, levels) — 21 findings, none refuted after checking against the driver — and an independent re-render through FluidSynth agreed with the kit's sampler.

Fun facts

  • The music credits itself. Every real piece embeds its metadata in ASCII: title, game, Yasunori Mitsuda, PROCYON STUDIO, and a notes field. BATTLE1.SMD's note reads "IBENT BATTLE" — a romaji slip for "EVENT BATTLE"; BATTLE2.SMD spells it correctly. Both battle themes are titled "Battle1", both jingles "Jingle2", and the game name alternates between XENOSAGA and Xenosaga from file to file. Nobody updated the metadata, and some ambience stubs carry the credit line too.
  • The orchestra is in the bank names. The Battle1 bank has 60 samples and 27 programs named like a session library: TimpC#2 FF, TimpG1 FF, SnareFF / SnareMF / SnarePP, F.HornSTCD3, F.HornCRSG#3, TrbFFC4LP.vag, CelloBassSTCC#3, ViolinSTCB5, Anvil, RollCymbal. The suffixes are dynamics and articulations (FF/MF/PP, STC staccato, CRS crescendo, LP looped) — multisampled orchestra with dynamic layers, played back live on a 2002 console.
  • Sixty seconds of orchestra in 17 KB. The note data for Battle1 is 17,168 bytes; LastBattle, the longest piece on the disc, is 23,960 bytes for a five-minute loop. The samples are where the megabytes go.
  • Rendered lengths (one loop plus a fade): Battle1/Battle2 162.7 s, LastBattle 305.5 s, Escape! 175.2 s, U.M.N. Mode 139.0 s, the jingles 34.9 s.
  • The voice streams hid their layout in plain sight. The .vds streams have no header. Decoded as mono they sounded pitch-correct but half-speed, "echoey", and tapped at block boundaries — and no sample rate fixed it, because the problem was a stereo interleave every 0x400 bytes. The confirmation came from the movies: their embedded ADS header spells out 48000 Hz, 2 channels, interleave 0x400 — the game's own spec for the format we had derived by correlation.
  • The battle voices are 535 little banks. yamamoto/snd/sed holds per-character attack, boost, special, bed and weapon banks (km_, so_, jr_, mo_, cs_, z8_ — KOS-MOS, Shion, Jr., MOMO, chaos and Ziggy, "Ziggurat 8"), 125 boss banks and enemy-tech banks. 238 samples ship as stereo halves named .aif.L / .aif.R, straight from the studio's file names.
  • Not every "voice" stream is a voice. The 92 streams under sound/vda mix spoken lines with ambience and machinery; a 12-second engine drone sits next to a 35-second monologue. A voiced-frame measure (fraction of frames with a strong pitch period) separates them reliably.
  • One stream lives alone. yajima/gameover.vds is the only streamed audio in chain 0 — the game-over music, parked in a developer's personal folder.

Still open

  • Reverb. The sequences use the SPU2 reverb send, but the reverb depth is per-scene engine state, not in the files, so renders are dry.
  • Envelopes are an approximation. The SoundFont carries the raw SPU2 ADSR registers; the kit's sampler models them piecewise- exponentially. Close, not bit-exact.
  • The kit's sampler is deliberately simple (linear-interpolation resampling, no per-voice filtering). For the most faithful listen, play the MIDI against the SoundFont in a real synth and compare.

Where the details are

  • docs/FORMATS.md § "Sequenced BGM" — the byte-level spec with the driver offsets it was read from, and the streamed-audio and movie sections.
  • ssd.py (SMD → MIDI, SWD → SoundFont 2) and ssd_render.py (the sampler and batch exporter).
  • docs/FINDS.md and docs/HISTORY.md — the metadata quirks and the recording history (London Philharmonic Orchestra, Metro Voices, Joanne Hogg) in context.