Swiss German Transcription
Swiss German isn't just "German with an accent"
That's why we built something different.
The problem with generic tools
Generic transcription tools are trained on standard German. They hear "Grüezi" and guess. They encounter Bernese and struggle. They hit Walliser dialect and give up.
Swiss German has no standard written form, massive regional variation, and vocabulary that simply doesn't exist in High German. This makes it one of the hardest challenges in speech recognition.
Our approach
We didn't try to fine-tune a generic model. We built specialized Swiss German speech recognition from the ground up — in formal collaboration with FHNW's Institute for Data Science.
This was a funded research project, not a marketing exercise. Real researchers. Real methodology. Published results.
Born from research
This capability came from a funded research project with FHNW's Institute for Data Science.
Research project
Schweizerdeutsche Spracherkennung und Übersetzung in Standarddeutsch
- Duration: June 2023 – August 2024
- Institution: FHNW Hochschule für Informatik
- Institute: Institut für Data Science
- Project Lead: Prof. Manfred Vogel
- Funding: Forschungsfonds Aargau
What we built
- Specialized Swiss German speech recognition models
- Translation pipeline from Swiss German to Standard German
- Swiss Parliament Corpus 2.0 (published as open data)
- Standardized test corpus for benchmarking
The underlying technology is also available as a free demo at stt4sg.fhnw.ch
How we measure quality
We don't throw around percentage accuracy claims — they're meaningless without context.
Word Error Rate (WER)
The percentage of words that need correction. Lower is better. We test against a standardized Swiss German corpus to measure real performance.
BLEU Score
For Swiss German to Standard German translation quality. Measures how well the output matches reference transcripts. Industry-standard metric.
We're honest about limitations
Audio quality matters. A clear recording in a quiet room will always transcribe better than a noisy street interview. No AI can fix bad audio.
Dialect variation is real. Heavy dialect mixing or very localized expressions may still need manual correction. Swiss German is diverse — that's what makes it beautiful, and challenging.
We're good, not magic. Try it yourself with your own recordings to see how it performs on your specific content.
Best results when
- Clear audio with minimal background noise
- Single speaker or clear turn-taking
- Standard recording equipment
- Common Swiss German dialects
Try it yourself
Upload a Swiss German recording and see how it performs. The best test is your own content.