Transcribe Swiss German locally: what AI still cannot do

I wanted to transcribe my voice memos locally — on the Mac, no US server, data stays on the device. I did my research, found the favorite: Whisper Notes with Whisper Large V3 Turbo. Downloaded the model, started recording, spoke in Swiss German the way I always do when taking notes for myself. Then came the transcription.

Readable. Roughly. Every sixth or seventh word was wrong.

The problem goes deeper than poor optimization

Swiss German is not a dialect that can be described as a slight variant of Standard German. There is no standardized written language. An ASR model therefore has to do two things at once: recognize sounds and simultaneously translate into Standard German. Bavarian or Austrian German are much simpler in this regard — the same task, but with a standard form as the target that most models already know.

That explains the numbers: 80–87% accuracy for Swiss dialects, compared to 91–93% for Austrian German. The difference sounds small, but it isn’t. Every sixth or seventh word is wrong — for a meeting transcript that’s borderline, for an interview with affected individuals it’s insufficient, for legally relevant content it’s a problem.

What exists — and what that means

On HuggingFace there are fine-tuned Whisper variants for Swiss German. One of them was retrained in March 2026 with 301 hours of curated Swiss training data — primarily parliamentary speech and semi-formal recordings. The best checkpoint achieves a WER of 37.96%, the base model 45.71%. A measurable improvement. Whether such models run convincingly locally on a Mac is another question — the path from a HuggingFace checkpoint to a functioning local tool is undocumented and has its own hurdles.

Töggl from recapp IT AG (Visp) is another approach: a Swiss provider, trained with SRF data, with speaker diarization. Good dialect recognition, according to the developer. But not a local model — the audio data goes to a Swiss cloud. That’s better than a US cloud, but not the same as no cloud at all.

Why this matters to us

We speak Swiss German internally. That’s non-negotiable. Meetings, brainstorms, interviews with affected individuals — everything in dialect. For campaigners and NGO teams who need transcripts, evaluate interviews, or want to subtitle videos, this is a real limitation.

On top of that, there’s data protection. Anyone transcribing sensitive interviews or internal strategy discussions has good reasons not to send the audio data to a third-party server — whether it’s in the US or in Switzerland. A local model that is good enough would be almost ideal for this use case.

That model doesn’t exist yet for Swiss German. Or I haven’t found it yet.

What you can do

Right now there are two pragmatic paths: either speak Standard German when transcription matters — unsatisfying, but it works. Or use a Swiss cloud service like Töggl and consciously weigh the data protection question.

Those who are technically savvy can experiment with the HuggingFace models. The models are public, the retraining setup is documented.

What I’m looking for: a local model on the Mac or iPhone that runs well enough in practice with Züridütsch, Bärndütsch, or other common dialects. If you’ve managed to make that work — write to me.

This post was created with partial use of AI tools (including fact-checking) such as Gemini, Perplexity, and Manus. All content was carefully reviewed, revised, and approved by me.