# External model assets

`models/` is a deployment directory, not a source-controlled package and not
a wheel payload. Set `AX_MODEL_DIR` to this directory (or an identically laid
out copy) at runtime.

The authoritative inventory is
[`src/ax_meeting_cpp/resources/model_assets.json`](../src/ax_meeting_cpp/resources/model_assets.json).
It records every required file's relative path, byte size and SHA-256 for the
two ASR choices:

* `sensevoice`: standard ASR plus word timestamps.
* `firered_punc`: FireRed AED and CT-Transformer punctuation (no word timings).

The required layout is deliberately grouped by model family:

```text
models/
├── vad/          # vad.axmodel, CMVN, config
├── cam++/        # CAM++ speaker embedding model
├── sensevoice/   # model, CMVN/config, BPE and language queries
├── fireredasr/   # FireRed encoder/decoder, CMVN, dictionary, BPE, PE
└── punc/         # CT-Transformer model and vocabulary
```

Before deploying a copied bundle, verify its integrity without loading any
model:

```bash
python3 tools/verify_assets.py \
  --manifest src/ax_meeting_cpp/resources/model_assets.json \
  --root /opt/ax-meeting/models --backend firered_punc
```

The default check only requires the declared files to exist. Add
`--verify-size --verify-sha256` when an exact supplied asset set must be
audited; hashes are informative rather than a runtime or build prerequisite.

Model files retain their upstream terms. The source-code MIT license neither
grants model redistribution rights nor asserts a license that was not supplied
with the assets.