Audio you can put in a receipt.
The same samples, everywhere.
A neural audio codec (the SoundStream / EnCodec family) decodes through a
float network, so a ~1e-5 round-off makes the reconstructed waveform — and its
hash — diverge between machines: never receiptable. wai.audio.int_codec is a
learned codec built the opposite way: the synthesis is a stack of integer
1-D transposed convolutions + integer IGDN, run over a factorized-rANS latent, the whole
decode the same pure-i64 Rust the native sink runs, compiled to wasm32 — no float,
no ML runtime. So the int16 PCM is byte-identical on every machine, and this page
re-checks its hash against the native reference before playing it. That is
sample-equivalence — a portable PCM hash, a waveform you can put in a receipt. It is
the audio sibling of the learned image codec (wai.neural.int_hyper).
Why this is the moat
The waveform you hear was reconstructed in your browser from a ~— KB learned model + a tiny rANS bitstream, and before it played the page hashed the decoded int16 PCM and checked it against the hash the native reference produced — they match, byte-for-byte. A float neural codec can't make that promise: its samples drift between machines, so you can't pin them. This decode is integer all the way down — the transposed convolutions are integer accumulation + an arithmetic-shift requant, IGDN is integer isqrt, the latent comes off an integer rANS — so the PCM hash is portable. Reference engine + conformance corpus in the open-standards repo, Apache-2.0.