Speech-to-text for Swiss German and 100 more languages, tailored to you
Every dialect region, straight into Standard German
Swiss German has no standard written form. Our model hears dialect and writes Standard German you can use right away. Trained on recordings from every dialect region: Basel, Bern, Grisons, Eastern Switzerland, Valais, Central Switzerland, Zurich.
On par with the state of the art
On the STT4SG-350 test set (35 hours of audio from every Swiss German dialect region), our current model reaches a word error rate of 12.4 (0 = identical to the reference, 100 = every word is wrong). This puts it on par with the current state of the art (Timmel et al.), ahead of the model in the STT4SG-350 paper and far ahead of the original Whisper model.
Performance of our current model per dialect region on the STT4SG-350 test set:
Whisper large-v3 and Ateleris model: our own measurement
STT4SG-350 paper (2023): Plüss et al. (2023)
Timmel et al. (2024): Timmel et al. (2024)
Transcribe 100 different languages
Besides Swiss German, we cover German, French, Italian, English, Spanish and many other languages from around the world – with automatic language detection.
Full list of supported languages: Languages
Please note: the transcription quality is not the same for all 100 languages. For some, there simply isn't enough training data to reach good quality. We're happy to help with specific questions, e.g. about quality for less common languages.
Performance of our model across languages on the FLEURS dataset:
Optimal transcription quality with custom models
If you have your own data, meaning audio and the matching transcripts, we can train a model on it that performs significantly better than any standard model. The custom model runs through the same API or on your own infrastructure.
Rare dialect or low-resource language?
Whatever the language: if you have transcribed recordings, we can train a model on them. Even 10 hours of data often bring a clear improvement.
Special requirements for the transcript?
Clean read instead of verbatim, your own formatting rules, domain vocabulary: with example data in your target format, the model learns exactly that style.
What we're working on
The service keeps getting better. New model versions ship through the same API, with no changes to your integration.
Pass your own terms and names via the API.
Improved quality for low-quality recordings and audio with background noise.
Pass context about the recording and instructions via the API; an LLM then improves or rewrites the transcript as instructed.