open-standards/wai/score-conformance

Score Conformance Corpus

Vectors for the score extension (wai.audio.score). The criterion is mixdown-equivalence: given the same score, a conforming sink MUST render the identical canonical i16 PCM via the deterministic reference synth, hence the identical BLAKE3 — on every machine, no tolerance parameter. The music counterpart of spatial audio's mixdown-equivalence, with the recorded sound objects replaced by note events.

Provenance. Generated here: wai_score_vectors gen writes every score from notes authored in code. No recording.

The reference synth is trig-free: a phase accumulator drives an oscillator (the piecewise-linear saw / square / triangle / pulse; the same saw, square and pulse band-limited by a polynomial step correction; a sine by a refined parabola; noise) shaped by an envelope, all in the wai.det.fixed64 floor. A voiced score (score §2.4) gives each note a pan between two channels, curved envelope segments, low-pass sections that follow the envelope, a glide, a noise seed and echoes. Pitch is carried as a per-sample phase increment authored off-line, so the floor never evaluates an exponential. A sink with a soundfont or neural instrument renders the same score richly — that is presentation, not conformance.

Layout

• vectors/<name>.wscr — the WSCR container (contract + note events, and voices in a voiced score).
• notes/<name>.json — the same score in the authoring form (score §2.3): the contract and every note field as an integer, each voice member too, the input of wai score build.
• expected_hashes.json — per score: notes, duration (sample frames), channels, sample_rate, score_hash (identity), mixdown_hash (the conformance quantity), and pcm_sha256, the SHA-256 of the same i16le samples, two channels interleaved (SPEC §7 Digests).

Vectors at 22 050 Hz in one channel, of the first revision's two sections:

• scale — a C-major scale ascending (saw).
• chord — a C-major triad held together (triangle, polyphony).
• arp — a four-step arpeggio that changes timbre across the phrase.

Vectors at 48 000 Hz:

• bandlimited — one channel, no voices: a high arpeggio (C6 to C7) of the band-limited saw, square and pulse and the sine, then noise.
• stereo — two channels, no voices: a triad of the band-limited saw and square and the sine, each centred.
• voiced — one channel, voiced: a plucked bass line through low-pass sections that close as each decay falls, a note that glides up a fifth, and one echoed three times.
• sting — two channels, voiced, two seconds: a kick that glides down, a hi-hat of seeded noise, a detuned pad spread left and right through low-pass sections that open with the envelope, a square bass, and a bell melody whose echoes cross from side to side.
• scale48 — two channels, voiced: a C-major scale panned left to right, each note echoed once from the other side; the composing sink's reference package plays its notes up and down.

Running

cd ../wai-rs
cargo run --features score --bin wai_score_vectors -- verify ../score-conformance
cargo run --features score --bin wai_score_vectors -- gen    ../score-conformance   # regenerate

Pass = mixdown-hash, PCM SHA-256 and score-hash equality, each notes/<name>.json exactly the vector's notes in the authoring form, the contract's channels and rate, and every note ending (echoes included) within the score.

With the wai tool (a build with score):

wai score build  ../score-conformance/notes/scale.json -o scale.wscr   # the committed WSCR, byte for byte
wai score render scale.wscr --pcm scale.s16 --json                     # samples, channels, pcm_sha256, mixdown_blake3, score_hash

tests/cli_generate_score.rs runs both over every vector.