# Neural audio route vectors

One vector for each capability whose payload SPEC §5 pins (*Payloads of the neural audio
codecs*): `wai.neural.encodec24`, `wai.neural.encodec48`, `wai.neural.dac16` and
`wai.neural.dac24`. They hold the route of the reference sink (`neural/audio.rs`) to the
pinned layout and decode; `tests/neural_audio_vectors.rs` decodes them.

**Provenance.** Generated by `tools/wai_neural_audio_vectors.py` from fixed dyadic
parameters and SplitMix64 codes (seeds 1 to 4). No audio and no model weights: each decoder
has its model's geometry and contract, not its network. Apache-2.0, as the repository.

Each directory holds:

- `decoder.onnx` — a small decoder with the model's codebook count `Q`, codebooks of 1024,
  a hop of 320 and its channels: inputs `audio_codes` (int64 `[1, 1, Q, T]`) and
  `audio_scales` (float `[1]`), output `audio_values` (float `[1, C, T·320 − d]`, `d` = 8
  for the two DAC models, else 0). A code `c` of codebook `i` adds `c·A[i] + B[i]` to its
  frame's latent `z`, a row holding 1024 adds nothing, and sample `k` of the frame on
  channel `ch` is `tanh(z·W[ch][k])` times the scale. `A`, `B` and `W` are dyadic, so every
  value before the `tanh` is exact in binary32.
- `vector.wai` — a `WAI1` envelope of the capability, pinned to `decoder.onnx`, whose payload
  carries fewer codebooks than the model holds (all 32 for `dac24`), so the route must mark
  the others unused; `encodec48` carries two chunks with scales 0.75 and 0.5, its last chunk
  six frames.
- `expected.f32` — the decode at the samples `meta.json` selects (every one, or for
  `encodec48` every sixteenth and every one in the chunks' overlap), each sample's channels
  in order, as little-endian binary32. The generator computes it itself, in Python, from the
  payload and the decoder's parameters: the second implementation of the route.
- `meta.json` — the capability, its geometry, the payload's shape, the selection rule and the
  decoder's SHA-256.

The reference's decode is within 10⁻⁵ of `expected.f32` (the decoder's `tanh` is the
runtime's). `python3 -I tools/wai_neural_audio_vectors.py --check` rebuilds every file and
compares: the decoders and envelopes byte for byte, `expected.f32` within 10⁻⁶.