Swiss German Whisper — GGML build (unofficial)
Ready-to-use whisper.cpp / GGML build of a Swiss German speech-to-text model. Transcribe Swiss German audio into written Standard German locally on a Mac — no cloud, no API key.
This is a format conversion, not a new model. I did not train it. The weights are Flurin17's fine-tune of OpenAI Whisper
large-v3-turbo; this repo only converts them to the GGML format that Mac apps like MacWhisper can load. Full credit below.
⚠️ License: non-commercial (CC-BY-NC-4.0)
Inherited from the source model. Free for personal / non-commercial use with attribution. No commercial use. Commercial use requires separate permission from the original author (Flurin17).
Files
| File | Size | Use it if… | sha256 |
|---|---|---|---|
ggml-model.bin | 1.5 GB | best quality (fp16) | e147fd5c5979a260b5ce3e0bff493918b14af58101f12cee9620f7f067e347c4 |
ggml-model-q5_0.bin | 547 MB | older Mac / less RAM (quantized) | 2d56e773724a247360067b527417842b81d25ff891fed014341a6844f15ea612 |
Verify after download:
shasum -a 256 ggml-model.bin
Use it — no terminal needed (MacWhisper)
- Download one of the
.binfiles. - MacWhisper → Settings → Models → Add custom model → pick the file.
- Set the language to German and transcribe.
Outputs written Standard German (Hochdeutsch), not dialect spelling. superwhisper and Whisper Notes load custom GGML models the same way.
Use it from the command line (whisper.cpp)
./build/bin/whisper-cli -m ggml-model.bin -f your-audio.wav -l de
Audio must be 16 kHz mono WAV:
ffmpeg -i input.m4a -ar 16000 -ac 1 -c:a pcm_s16le your-audio.wav
Limitations
- Trained mainly on parliamentary / read speech — spontaneous everyday dialect and heavy accents are harder.
- Outputs Standard German, not Swiss German spelling; use
-l de, not--translate. - Like all Whisper models it can hallucinate on silence or noise.
- Accuracy varies by dialect. Measure it on your own audio.
Model details
- Base: OpenAI Whisper
large-v3-turbo(n_mels 128, 32 audio / 4 text layers, vocab 51866). - Converted with whisper.cpp
v1.9.1-75-g0ae02cdbviaconvert-h5-to-ggml.py. - Integrity verified: model loads and reports the expected turbo dimensions.
Credits
- Fine-tuned model: Flurin17/whisper-large-v3-turbo-swiss-german — CC-BY-NC-4.0
- Base model: OpenAI Whisper
large-v3-turbo— MIT - Inference & GGML format: whisper.cpp by Georgi Gerganov & contributors — MIT
- Datasets referenced by the source model: SwissDial (ETH Zürich), Swiss Parliaments Corpus V2 (FHNW), All Swiss German Dialects Test Set (FHNW), ArchiMob (University of Zürich)
If you are one of the authors above and want a change to how this is credited or distributed, please open a discussion — happy to adjust.
License
Distributed under CC-BY-NC-4.0, inherited from the source model.