cfl4g/attune-whisper-coreml

Røst v3 Whisper 1.5B — WhisperKit Core ML

This repository contains a modified version of CoRal-project/roest-v3-whisper-1.5b, the Danish Whisper-large model trained by the Alexandra Institute as part of the CoRal project.

Changes: the original PyTorch weights were converted to Core ML with argmaxinc/whisperkittools (whisperkit-generate-model --generate-decoder-context-prefill-data) so the model can run with WhisperKit on Apple silicon. No retraining or fine-tuning was performed. Numerical outputs may differ slightly from the original because of the format conversion. All files in CoRal-project_roest-v3-whisper-1.5b/ were produced by this conversion.

It is used for the final Danish transcription pass in Attune.

License and use restrictions

The model and this derivative are distributed under the license in LICENSE, a custom license by Alvenir ApS adapted from the AI PUBS OpenRAIL-M license. All copyright and attribution notices of the original model are retained.

By using this model you agree to the use-based restrictions in that license (Section 4 of the preamble and Attachment A), including not to use the model or its derivatives to impersonate people, to detect or infer identity features of natural persons, or to disseminate machine-generated content without disclosing that it is machine-generated. These restrictions apply to everyone who receives the model from this repository.

Usage

let whisper = try await WhisperKit(WhisperKitConfig(
    model: "CoRal-project_roest-v3-whisper-1.5b",
    modelRepo: "cfl4g/attune-whisper-coreml"
))
let result = try await whisper.transcribe(
    audioPath: "meeting.wav",
    decodeOptions: DecodingOptions(language: "da", wordTimestamps: true, chunkingStrategy: .vad)
)

Credits

• Original model: CoRal-project/roest-v3-whisper-1.5b (Alexandra Institute, CoRal project)
• License: Alvenir ApS (Alvenir/coral-1-whisper-large)
• Base architecture: openai/whisper-large-v3