Swiss German speech recognition and translation to Standard German
Swiss German is a difficult language to transcribe. Our project has resulted in an easy, commercially available tool for transcribing and translating Swiss German into Standard German.
Recognising spoken language and transcribing it into text is a solved issue – if you’re dealing with one of the major global languages. So far, Swiss German speakers have had no access to decent transcription tools. With only five million native speakers, Swiss Germans haven’t been large enough of a data pool nor a commercially attractive market for international developers. Even those five million don’t really speak the same language: local dialects are numerous and sometimes wildly different. A native speaker from St. Gallen might not be able to understand what their compatriot from Valais is saying.
Yet the biggest challenge is that Swiss German is not a written language. When native Swiss German speakers write in a professional context, they typically do so in Standard German. This means there is no comprehensive data set that would index spoken words with written text in Swiss German. Despite their similarities, Swiss German and Standard German have a different vocabulary, different spelling, and even a different grammar. We can’t simply transcribe audio to text: we must translate between the two languages.
This is a practical problem that needs practical solutions, not just a theoretical method. We wanted our research project to result in a user-friendly, commercially available tool for both consumers and companies. That’s why we collaborated with the FHNW-spin-off Ateleris, a software company that specialises in transferring research knowledge into industrial and commercial use. Our institute’s role was to use our expertise in deep learning, machine learning, natural language processing, as well as speech-to-text and text-to-speech models. We also took advantage of the Swiss Parliaments Corpus data set, and we contributed by publishing the updated Version 2.0 of the data set after the project.
The project was a success with tangible results: Ateleris has embedded Swiss German into its multilingual transcription service Stenoris, for both cloud and on-premise use. Our institute also hosts a simple, browser-based demo site, available to everyone for no cost.
English might be the lingua franca of the global world, but local languages carry local cultures and are thus irreplaceable. When Swiss German speakers and learners have access to good digital tools, Swiss culture can also flourish online.
Information
School / Institute | FHNW School of Computer Science / FHNW Institute for Data Science |
Partner | |
Funding | |
Project timeline | June 2023 – August 2024 |
Project lead | Manfred Vogel (FHNW Institute for Data Science) |
Project team | Claudio Paonessa (FHNW Institute for Data Science) Yanick Schraner (FHNW Institute for Data Science) Vincenzo Timmel (FHNW Institute for Data Science) Laszlo Etesi (Ateleris) Silvan Laube (Ateleris) Andrea Zirn (Ateleris) Orell Bühler (Ateleris) Luca Schafroth (Ateleris) Simon Beck (Ateleris) Joel Blumer (Ateleris) Matthias Krebs (Ateleris) |
More information | Manfred Vogel (FHNW Institute for Data Science) |
Open resources |