Does Content Matter? Introducing SWC-LibDeep for Cross-Lingual Audio Deepfake Detection in English and German

SWC-LibDeep: A Cross-Lingual Audio Deepfake Dataset for English and German SWC-LibDeep is a dataset for research on audio deepfake detection. It contains 22,500 synthetically generated speech samples in English and German, created with six state-of-the-art voice conversion (VC) and text-to-speech (TTS) methods. The dataset comprises three equally sized subsets: LibriSpeech (English): 7,500 samples Spoken Wikipedia Corpora English: 7,500 samples Spoken Wikipedia Corpora German: 7,500 samples The included deepfakes were generated using Seed-VC, Diff-HierVC, OpenVoice v2, kNN-VC, Zonos v0.1 Transformer, and XTTS v2. SWC-LibDeep supports research on cross-lingual and cross-dataset audio deepfake detection, method generalisation, language and accent effects, and synthetic-speech artefact analysis. This release contains synthetic deepfake audio only. The original bona-fide recordings are not redistributed and must be obtained from the respective source corpora (LibriSpeech and Spoken Wikipedia Corpora). The dataset was introduced in: Does Content Matter? Introducing SWC-LibDeep for Cross-Lingual Audio Deepfake Detection in English and German, accepted at the IEEE International Conference on Big Data 2026. Please consult the included README.md file for dataset structure, usage conditions, source-corpus attribution, and citation instructions. This restricted-access dataset is provided exclusively for non-commercial academic research purposes; access requests are subject to review and approval by the dataset authors.