Reson Model Repository

This repository indexes machine-learning models used by the Reson forensic phonetics and linguistic research application. Reson downloads model data and runs supported analysis locally on the user's computer. User audio is not uploaded to this repository. This repository is not an inference API.

model_manifest.json is the canonical registry. Feature slots identify the purpose, while versioned entries identify the currently selected compatible model. The application includes a small fallback snapshot for first-run/offline metadata; explicit registry refreshes discover newer entries.

Slot	Current model	Weight source	Licence
transcription	Multilingual Whisper small, CTranslate2	SYSTRAN/OpenAI	MIT
speaker-analysis	ECAPA-TDNN VoxCeleb	SpeechBrain	Apache-2.0
bundled dependency	Silero VAD v6 in Faster-Whisper 1.2.1	Silero / Faster-Whisper	MIT

Public upstream weights are referenced at immutable revisions rather than copied unnecessarily. Full model licence texts are in licences/; upstream cards and attributions are retained in downloaded packs. No account/token is needed for public downloads. No paid inference service, backend server or GPU host is used.

Maintaining the registry

1. Choose a compatible model and review its licence and required inference files.
2. Register a pinned public Hugging Face source, or upload permitted clean assets under models/<feature>/<model-id>/ and register their pinned commit revision.
3. Calculate exact per-file SHA-256 and byte sizes; retain source attribution.
4. Update the feature entry, model_id, version, descriptor, source, totals and compatibility interval. Increment registry_version and the update date.
5. Validate the manifest using Reson's model_registry.validate_registry, then test local loading before publishing the update.

The application supports replacement within reviewed local engine adapters. New engines, changed package pins, new executable YAML or unknown feature slots require an application update. Remote registry data cannot install Python, execute scripts or rewrite Reson source. Whisper replacements must remain multilingual/Arabic-capable CTranslate2 packs. ECAPA configuration and dimensions must match the application's reviewed template. Model quality still requires independent evaluation; a successful hash check is not scientific validation.

Downloads use verified staging and atomic activation. Older installations remain usable offline and can be used while an optional update is available. Models can be installed individually or from Settings → Models & Offline Use. The large optional packs total 575,203,220 bytes; the 1,245,151-byte VAD dependency ships with the engine. Application/native runtime libraries are separate from model storage and may still be substantial, particularly Torch.