Duet model packages
Model packages and the signed model catalog for Duet, an Android app by Nyquist that runs speech models entirely on the phone. The app downloads from this repo; nothing is sent back.
Built by the Duet model lab from upstream releases: unchanged except for packaging, or (fastdiar) exported to ONNX from the released PyTorch checkpoints, checked against PyTorch. Layout:
catalog.json which package each phone tier gets (rollout percentages)
catalog.json.sig Ed25519 signature of catalog.json (base64); the app embeds the public key
packages/<id>/<version>/manifest.json files, sha256, license, runtime parameters
packages/<id>/<version>/... model files
The app trusts a package only if the signed catalog's manifest_sha256 matches its manifest and
every file matches the manifest's sha256.
Packages and licenses
| package | what | license |
|---|---|---|
| moonshine-v2-tiny-en-quantized@1.0.0 | Moonshine v2 tiny (English), quantized | MIT (Moonshine AI / Useful Sensors) |
| moonshine-v2-base-en-quantized@1.0.0 | Moonshine v2 base (English), quantized | MIT |
| moonshine-tiny-en-int8@1.0.0 | Moonshine tiny (English), int8 | MIT |
| moonshine-base-en-int8@1.0.0 | Moonshine base (English), int8 | MIT |
| silero-vad@1.0.0 | Silero VAD (voice activity detection) | MIT (snakers4/silero-vad) |
| fastdiar-small@1.0.0 | fastdiar speaker diarization, 2.4 M parameters, + Silero VAD v6 | MIT (Microsoft; Silero Team) |
| fastdiar-medium@1.0.0 | fastdiar speaker diarization, 4.6 M parameters, + Silero VAD v6 | MIT |
| fastdiar-large@1.0.0 | fastdiar speaker diarization, 10.4 M parameters, + Silero VAD v6 | MIT |
| redimnet2-b6@1.0.0 | ReDimNet2 B6 speaker embeddings (voiceprints), 12.3 M parameters | MIT (Palabra.ai) |
| redimnet2-b3@1.0.0 | ReDimNet2 B3 speaker embeddings (voiceprints), 4.1 M parameters | MIT (Palabra.ai) |
ONNX exports by sherpa-onnx (Apache-2.0). Each package directory carries its upstream LICENSE where one was shipped. fastdiar: ONNX export of the v0.1.0 checkpoints with the streaming state as explicit inputs/outputs, by the Duet model lab. Upstream models: Moonshine, Silero VAD, fastdiar, ReDimNet2.