SDS-200: A Swiss German Speech to Standard German Text Corpus

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3250–3256 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3250 SDS-200: A Swiss German Speech to Standard German Text Corpus Michel Pl¨ussa, Manuela H¨urlimannb, Marc Cunyc, Alla St¨ocklic, Nikolaos Kapotisb, Julia Hartmanna, Malgorzata Anna Ulasikb, Christian Schellera, Yanick Schranera, Amit Jain, Jan Deriub, Mark Cieliebakbc, Manfred Vogela aUniversity of Applied Sciences and Arts Northwestern Switzerland, Windisch bZurich University of Applied Sciences, Winterthur cSpinningBytes AG, Winterthur michel.pluess@fhnw.ch Abstract We present SDS-200, a corpus of Swiss German dialectal speech with Standard German text translations, annotated with dialect, age, and gender information of the speakers. The dataset allows for training speech translation, dialect recognition, and speech synthesis systems, among others. The data was collected using a web recording tool that is open to the public. Each participant was given a text in Standard German and asked to translate it to their Swiss German dialect before recording it. To increase the corpus quality, recordings were validated by other participants. The data consists of 200 hours of speech by around 4000 different speakers and covers a large part of the Swiss German dialect landscape. We release SDS-200 alongside a baseline speech translation model, which achieves a word error rate (WER) of 30.3 and a BLEU score of 53.1 on the SDS-200 test set. Furthermore, we use SDS-200 to fine-tune a pre-trained XLS-R model, achieving 21.6 WER and 64.0 BLEU. Keywords: Corpus, Less-Resourced/Endangered Languages, Speech Recognition/Understanding, Speech Resource/Database, Statistical and Machine Learning Methods 1. Introduction We present Schweizer Dialektsammlung (SDS-200), a corpus of Swiss German dialectal speech with the cor- responding Standard German text. The data consists of 200 hours of speech. We make the corpus publicly available 1. Swiss German is a family of German dialects spoken by around five million people in Switzerland. It differs from Standard German regarding phonetics, vocabu- lary, morphology, and syntax and is primarily a spoken language. While it is also used in writing, particularly in informal text messages, it lacks a standardized or- thography. This leads to difficulties for automated text processing due to spelling ambiguities and huge vo- cabulary size. Therefore, it is often preferable to work with Standard German text, for which automated pro- cessing tools exist in abundance. The main challenge is that Swiss German is not a unified language but a collection of dialects, which sometimes differ signif- icantly in phonetics, grammar, and vocabulary. The immense vocabulary makes it hard to create a Swiss German Automatic Speech Recognition (ASR) system. Due to these reasons, Swiss German is a low-resource language. One way to tackle Swiss German ASR is an end-to-end Swiss German speech to Standard German text approach. This can be viewed as a speech transla- tion (ST) task with similar source and target languages. Training a model for this task requires a substantial amount of data. Unfortunately, not enough public data is available for Swiss German. The largest available corpus, the Swiss Parliaments Corpus (SPC) (Pl¨uss …

Content truncated to 500 words