Reproduce “SDS-200: A Swiss German Speech to Standard German Text Corpus” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Michel Plüss, Manuela Hürlimann, Marc Cuny, Alla Stöckli, Nikolaos Kapotis, Julia Hartmann, Malgorzata Anna Ulasik, Christian Scheller and 5 more
Researchers from FHNW, ZHAW, and SpinningBytes AG created SDS-200, a 200-hour multi-dialectal Swiss German speech-to-Standard German text corpus. This publicly available dataset, collected from nearly 4000 speakers via crowd-sourcing, addresses the data scarcity for Swiss German speech technology and achieves a baseline WER of 21.6 and BLEU of 64.0 for end-to-end speech translation using a fine-tuned XLS-R model.
Key insights
- Crowd-sourcing combined with a two-step validation process effectively collects a large-scale, high-quality, multi-dialectal speech corpus for low-resource languages.
- The SDS-200 corpus is foundational for end-to-end Swiss German speech-to-Standard German text translation, enabling the application of modern deep learning models.
- Combining SDS-200 with existing dialect-specific corpora like SPC, or fine-tuning pre-trained multilingual models like XLS-R, significantly improves speech translation performance, demonstrating the value of data aggregation and transfer learning for low-resource dialects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.