Datasets:
Clarification on License (CC BY-SA 4.0 vs. Apache 2.0) and Source Video Mapping (Revision eee168f9)
#1
by dattruongthanh - opened
Hello @mueller91 ,
We are conducting an academic study on audio deepfake detection and plan to use a subset of this dataset strictly for local cross-dataset evaluation of our pre-trained models. Before downloading any .wav files from the pinned revision (eee168f92c367f8c82ff2cf42b6f61e362fd6211), could you kindly clarify three points?
- Applicable License: The Hugging Face dataset card specifies CC BY-SA 4.0, whereas the Fraunhofer AISEC project page lists Apache 2.0, and the ZIP archive does not contain a separate
LICENSEfile. Which license governs the WAV files and metadata in this release? - Scope of Use & Publication: Are we permitted to download a small subset (up to 512 clips) for local evaluation and publish aggregated metrics (e.g., ROC-AUC, error rates) in our paper, without redistributing any raw audio files? Are there specific attribution requirements for the underlying third-party sources?
- Source Provenance Mapping: Do you provide a mapping table linking each audio filename to its original source video, timestamps, or synthesis source so we can control for source-level dependencies during evaluation?
Thank you very much for your time and assistance!
Hi @dattruongthanh ,
thanks for checking first. Here are the answers:
- License: CC BY-SA 4.0 applies to both the audio and the metadata. The mismatch is on us. I’ll align the project page and the dataset card.
- Scope: Yes, that’s fine. You can download a subset, evaluate locally, and publish aggregated metrics. You can also use the full dataset if you like. Please don’t redistribute the raw audio, and please cite the original paper.
- The clips come from publicly available recordings of public figures. We don’t have per-clip source attribution, so citing the paper is sufficient.
- Provenance: No, there’s no mapping to source videos or timestamps, unfortunately.