ZHAW researchers receive this year’s SwissNLP Award
Every year, the Swiss Association for Natural Language Processing (SwissNLP) selects a team, organization or project that pushes the boundaries of Natural Language Processing in Switzerland. In 2023, a team of researchers from ZHAW, FHNW and UZH received this prestigious award for creating the largest Speech Translation Corpus for Swiss German to date.
NLP research is an interdisciplinary field and involves the automatic processing of spoken and written language by computers. It is a branch of artificial intelligence (AI) and makes it possible to analyse, understand, translate or generate texts or speech. In Switzerland, the association SwissNLP promotes NLP, computational linguistics, and text analysis by bringing together experts and users from industry and academia. Every year, SwissNLP organises the conference SwissText, where NLP practitioners and researchers discuss the latest trends and challenges in workshops, presentations and networking sessions.
The SwissNLP Award goes to…
The SwissNLP Award is an award presented annually by SwissNLP at the SwissText conference. It recognises outstanding initiatives that advance NLP in Switzerland. The projects can come from research, education, industry, or society and must have a clear benefit or added value for the Swiss NLP landscape.
The SwissNLP Award 2023 was presented to the creators of the STT4SG-350 corpus, the largest Speech Translation Corpus for Swiss German to date. The three universities behind the project, FHNW (University of Applied Sciences and Arts Northwestern Switzerland), ZHAW (Zurich University of Applied Sciences) and UZH (University of Zurich), collected a balanced corpus of nearly 350 hours of Swiss German audio recordings and the corresponding Standard German texts. This corpus was then used by its creators to train state-of-the-art Speech-to-Text systems (STT) for Swiss German dialects.
Largest Swiss German Speech Corpus
The STT4SG-350 corpus contains 247,527 audio recordings in Swiss German dialects. Each recording is accompanied with its transcript in Standard German and meta-data of its speaker (age, gender, dialect region), totalling around 42,000 words. The Swiss German speaking part of Switzerland is segmented into 7 dialect regions, and for each region there are approximately 45 speakers, totalling to 49 hours per region on average.
The corpus has been used already to train speech-to-text systems for each dialect, to generate synthetic Swiss German audio from written text (text-to-speech), for analyzing linguistic features of different dialects and to build a dialect identification system that determines where the dialect of a speaker originates from. STT4SG-350 is available for research and commercial purposes.
More information