MLX conversion of openai/whisper-large-v3 with the LoRA adapter Flurin17/whisper-large-v3-peft-swiss-german merged into the weights. This is a standalone FP16 checkpoint for Apple Silicon, mlx-audio, oMLX and mlx-whisper.

Usage

oMLX

Use oMLX 0.3.0 or newer. Homebrew and DMG installations include the audio dependencies; source installations require pip install "omlx[audio]". Download this repository through the oMLX model browser or place it below the configured model directory. It is detected as an audio_stt model and served through POST /v1/audio/transcriptions.

bash

curl http://localhost:8000/v1/audio/transcriptions \
  -F "file=@audio.wav" \
  -F "model=whisper-large-v3-swiss-german-mlx-fp16" \
  -F "language=de"

mlx-whisper

pip install mlx-whisper

python

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="gcoli/whisper-large-v3-swiss-german-mlx-fp16",
    language="de",
    task="transcribe",
)
print(result["text"])

Format and provenance

Format: MLX safetensors, FP16
Base revision: 1ecca609f9a5ae2cd97a576a9725bc714c022a93
Adapter revision: 2ae117cf342bc57d6068066181a9d359e98a2961
MLX conversion mapping: Apple mlx-examples Whisper converter, revision 796f5b53cab69a3d48a44233ce21aae889e94a08

The conversion remaps the merged Hugging Face Whisper tensors to the native mlx-whisper layout and transposes convolution kernels as required. No quantization is applied.

oMLX compatibility

The repository includes the Hugging Face processor and tokenizer assets needed by oMLX's mlx-audio Whisper loader. oMLX auto-detects model_type: whisper as an audio_stt engine.

Limitations

The model can hallucinate or omit text, particularly with noise, silence, overlapping speakers, uncommon dialects, or specialized vocabulary. Do not use its output as the sole basis for high-impact decisions.

License and usage conditions

SwissDial CC BY-NC 4.0 with no-reidentification condition

The Swiss German adaptation was trained on the SwissDial dataset. Use of this model is subject to the following conditions inherited from the SwissDial access terms:

Non-commercial use only under the Creative Commons Attribution-NonCommercial 4.0 International License.
No re-identification: do not attempt to determine the identity of any speaker in the SwissDial dataset.
Attribute SwissDial and cite its publication when using this model for research.

This checkpoint is a modified and converted derivative: a SwissDial-trained PEFT adapter was merged into openai/whisper-large-v3 and converted to MLX FP16. The Whisper base model remains subject to its Apache-2.0 license. This repository grants no additional rights to upstream models, training data, or software and does not imply endorsement by ETH Zurich, OpenAI, or the adapter authors.