Betthupferl: A multi-dialectal dataset for German dialect ASR and dialect-to-standard speech translation.

This repository contains supplementary material for the paper

“A multi-dialectal dataset for German dialect ASR and dialect-to-standard speech translation” (Verena Blaschke, Miriam Winkler, Constantin Förster, Gabriele Wenger-Glemser, and Barbara Plank). Accepted to Interspeech 2025. Preprint: https://arxiv.org/abs/2506.02894

Please cite the paper if you use any of this data/code.

This repository does not contain any audio data. Access to the audio data must be granted by Bayerischer Rundfunk on a case-by-case basis due to copyright restrictions (contact Gabriele Wenger-Glemser). If you obtain the audio files, create a folder called audio and place them inside that folder.

All subfolders containing transcriptions are in zip archives with the password MaiNLP so as to prevent potential inclusion in web-scraped datasets (cf. Jacovi et al., 2023). Unzip them to get the subfolders with the same name. Please do not re-distribute the transcriptions.

You can find the following contents:

• Reference transcriptions:
  • transcriptions_dialect: Dialectal and Standard German references for dialectal audios (converted from FOLKER annotations)
  • transcriptions_standard_german: Standard German references for Standard German audios (converted from FOLKER annotations)
  • data_processed/transcriptions: Converted versions of the files in transcriptions_*: one TSV file per story & language variety combination, also contains the corresponding audio filepaths
  • data_processed/pure_text: Converted versions of the files in transcriptions_*: one file per language variety (or group thereof) and transcription type (dialectal _dial or standardized _deu), one sentence per line. The files with _gold_ in the name contain the original punctuation and capitalization, the files with _goldnorm_ are lowercased and without punctuation.
• Word-level and sentence-level annotations of the references and/or ASR hypotheses are in the analysis folder (more information in the readme file there)/
• The code for ASR predictions and analyses is in code (more details in the readme file there). The scripts should be run from within that folder.
• The model predictions are in predictions (one subfolder per model). The original predictions are in the TSV files; the TXT files contain the pure text so as to be directly comparable with the files in data_processed/pure_text. Important: The TSV files contain older versions of the reference transcriptions. These older references are ignored when calculating the ASR scores.
• The ASR scores (WER, CER, BLEU) are in scores. The files ending with zeroshot.tsv contain the results tables (results averaged over sentences), and the files ending with zeroshot_detailed.tsv list the scores for each sentence.
• The annotation guidelines/descriptions and the data statement are in the PDF data-statement_annotation-guidelines.pdf.