aalst/nb-whisper-large-mlx

NB-Whisper Large — MLX

This model was converted to MLX format from NbAiLab/nb-whisper-large.

Refer to the original model card for detailed information about training data, performance benchmarks, and intended use.

About the Original Model

NB-Whisper Large is a Norwegian speech recognition model developed by the National Library of Norway AI Lab (NB AI-Lab). It was fine-tuned from openai/whisper-large-v3 on 66,000 hours of Norwegian speech data (8 million aligned audio clips of 30 seconds each, 250,000 training steps). It supports Norwegian Bokmal, Nynorsk, and English.

Performance (from the paper):

Dataset	OpenAI Large-v3 WER	NB-Whisper Large WER
Fleurs (Bokmal)	10.4%	6.6%
NST (Bokmal)	6.8%	2.2%

See the paper for additional benchmarks including Nynorsk and dialect evaluation.

Use with mlx-whisper

pip install mlx-whisper

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.mp3",
    path_or_hf_repo="aalst/nb-whisper-large-mlx",
    language="no",
)
print(result["text"])

Conversion Details

• Converted from: NbAiLab/nb-whisper-large (PyTorch/Safetensors)
• Converted to: MLX format (optimized for Apple Silicon)
• Precision: float16
• Conversion tool: mlx-examples/whisper/convert.py
• Changes: Format conversion only. No fine-tuning or weight modifications.

License

This model inherits the Apache 2.0 license from the original NB-Whisper model.

Citation

@misc{kummervold2024whisperingnorwegian,
    title={Whispering in Norwegian: Navigating Orthographic and Dialectic Challenges},
    author={Per Egil Kummervold and Javier de la Rosa and Freddy Wetjen and Rolv-Arild Braaten and Per Erik Solberg},
    year={2024},
    eprint={2402.01917},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}