LENS-DF Sample Data: A Dataset for Long-Form, Multi-Speaker, and Noisy Audio

This repository offers a sample dataset generated using the LENS-DF pipeline, a novel framework designed for creating long-form, multi-speaker, and noisy audio data. This work has been accepted to IEEE IJCB 2025. The arXiv link to the camera-ready paper will be added shortly. The provided sample data is derived from the ASVspoof 2019 LA development and evaluation sets, with partitions corresponding to the original dataset's structure. For detailed information regarding the file structure and usage of this sample data, please refer to the main LENS-DF GitHub repository [1]. You can also utilize the pipeline available in the main repository to generate your own custom datasets. NOTE: The 'Path' entries in ultra_deepfake.csv demonstrate solely as formatting examples within their respective clusters. Please use them for reference only. If you find this work valuable for your research, please consider citing our paper: @inproceedings{Liu2025LENSDF, author = {Liu, Xuechen and Ge, Wanying and Wang, Xin and Yamagishi, Junichi}, title = {LENS-DF: Deepfake Detection and Temporal Localization for Long-Form Noisy Speech}, booktitle = {IEEE International Joint Conference on Biometrics (IJCB)}, address = {Osaka, Japan}, year = {2025},} [1] https://github.com/nii-yamagishilab/LENS-DF-DataGen (not active yet) ------------------ Authors: Xuechen Liu, Wanying Ge, Xin Wang, and Junichi Yamagishi Affiliation: National Institute of Informatics, Japan Acknowledgements: This study is supported by the New Energy and Industrial Technology Development Organization (NEDO, JPNP22007), and JST AIP Acceleration Research (JPMJCR24U3). This study was partially carried out using the TSUBAME4.0 supercomputer at the Institute of Science Tokyo.