Røst v3 Whisper 1.5B — WhisperKit CoreML
CoreML conversion of
CoRal-project/roest-v3-whisper-1.5b
for on-device Danish speech recognition with
WhisperKit.
The source model reports 11.6% CER on conversational CoRal-v3 and 4.5% CER on read-aloud CoRal-v3. This repository changes the runtime format, not the training data or intended use.
Reproduce the conversion
Generated on Apple Silicon with whisperkittools commit
84f77a83c8f530022ae55fbb1a64b3351ef63c7a:
whisperkit-generate-model \
--model-version CoRal-project/roest-v3-whisper-1.5b \
--output-dir ./roest-v3-whisper-1.5b-coreml \
--generate-quantized-variants \
--allowed-nbits 8 \
--force-recipe-nbits \
--disable-default-tests
The published variant contains WhisperKit's AudioEncoder,
MelSpectrogram, and TextDecoder compiled model bundles plus the original
model configuration.
License and restrictions
The source model and these converted weights use the source model's custom
OpenRAIL-M-derived license. Commercial use is allowed, but speech synthesis,
voice cloning, and biometric identification are prohibited. See LICENSE
for the complete terms.
Citation:
Nielsen, D. S., et al. CoRal: A Danish Corpus and Benchmark for Automatic Speech Recognition. Alexandra Institute, 2024.
- Downloads last month
- -
Model tree for kramerthomas/roest-v3-whisper-1.5b-coreml
Base model
openai/whisper-large-v3