Whisper Large-v3 Swiss German – WhisperKit/CoreML
WhisperKit-compatible CoreML conversion of openai/whisper-large-v3 with the
LoRA adapter Flurin17/whisper-large-v3-peft-swiss-german merged into the model
weights before conversion.
Intended use
On-device transcription of Swiss German speech into German text on Apple
Silicon through WhisperKit. Use de as the language and transcribe as the
task. Accuracy and runtime characteristics of this CoreML conversion have not
yet been independently benchmarked.
Bundle size: approximately 2.9 GB (uncompressed float16 conversion).
Sources
Conversion
- Load
openai/whisper-large-v3 in float16.
- Load the PEFT LoRA adapter and merge it with
merge_and_unload.
- Save the standalone Transformers checkpoint.
- Convert that checkpoint using
whisperkit-generate-model.
Pinned revisions:
- Base:
1ecca609f9a5ae2cd97a576a9725bc714c022a93
- Adapter:
2ae117cf342bc57d6068066181a9d359e98a2961
- whisperkittools:
84f77a83c8f530022ae55fbb1a64b3351ef63c7a
Conversion verification
The default whisperkittools PyTorch-to-CoreML correctness tests completed
successfully on Apple Silicon:
| Component | PyTorch-to-CoreML PSNR | Static ANE support coverage |
|---|
| Text decoder | 55.2 dB | 99.86% |
| Audio encoder | 55.6 dB | 100.00% |
| Mel spectrogram | 67.0 dB | 80.00% |
Decoder token argmax accuracy was 100% in the converter test. These values
measure conversion fidelity and static operator support; they are not an ASR
accuracy or real-device speed benchmark.
Application testing
The model was tested successfully in Blitztegschter, a fork of the
Blitztext app.
Qualitative observations on Apple Silicon M5:
- better recognition accuracy than the tested Whisper Large-v3 Turbo variant;
- faster transcription than the tested Turbo variant;
- subjectively comfortable waiting time during normal application use.
A similar Large-v3 versus Turbo result had previously been observed on an
Apple Silicon M1 Max even without Swiss German fine-tuning. Testing this
fine-tuned CoreML bundle on the M1 Max is still pending.
These are practical, qualitative observations rather than a controlled
benchmark. Results can vary with hardware, audio quality, dialect, recording
length, decoding settings, and application integration.
Limitations
Whisper can hallucinate or omit text, especially with noise, silence, rare
dialects, overlapping speakers, or domain-specific vocabulary. Do not use the
output as the sole basis for high-impact decisions. Obtain consent before
transcribing people.
License
This model is provided under the SwissDial CC BY-NC 4.0 with
no-reidentification condition. The applicable license information and access
to the license terms are available here:
The repository is not gated. Downloading the files does not remove or replace
the user's obligation to comply with the applicable license terms.
This repository does not grant additional rights to other upstream models,
software, or dependencies. Each user is responsible for reviewing and
complying with their applicable licenses, usage conditions, attribution
requirements, and distribution restrictions.
In particular, users must independently verify the applicable terms for:
openai/whisper-large-v3;Flurin17/whisper-large-v3-peft-swiss-german;- WhisperKit, whisperkittools, Core ML, and their dependencies.
Availability of files on Hugging Face should not be interpreted as permission
for uses outside the linked license terms.
