Models v1

Speech models for WalkieTalkey, converted for transcribe.cpp. The app downloads them from this release and checks every file against SHA256SUMS.

Model	Files	Licence
Edda v0.2 by the Alexandra Institute (default)	edda-v0.2-Q8_0.gguf (886 MB)	Apache-2.0: edda-v0.2-LICENSE, edda-v0.2-NOTICE (base model Whisper, MIT)
Ekko v1 Tiny by RyeAI (fast model)	ekko-v1-tiny-Q8_0.gguf (134 MB)	NVIDIA Open Model License, CC BY 4.0 for RyeAI's changes, CoRal terms: ekko-v1-tiny-*
Ekko PnC v2 by RyeAI (punctuation)	pnc.int8.onnx, tokenizer.json, config.json	CC BY 4.0, CoRal terms: ekko-pnc-*

Model cards: edda-v0.2-README.md, ekko-v1-tiny-README.md, ekko-pnc-README.md (conversion, validation, accuracy, training data).

The use restrictions of the CoRal licence travel with all three models: no impersonating a person (including synthetic speech that resembles a specific person's voice), no detecting or inferring a person's identity or characteristics, and the other uses in Attachment A of that licence. Dictation and transcription are fine. All three were trained partly on FT Speech (Folketinget); whether its terms limit companies' use of models trained on it is an open question (decision 18.7 in docs/MODELS.md).

Converted and published by the WalkieTalkey project (Navigent.io ApS). The weights are the originals, only converted (and for Edda and Ekko v1 Tiny quantised to Q8_0).