SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Toward Cutting-Edge Speech Generation Methods
Keywords: dataset, deepfake detection, anti-spoofing, speech generation
Abstract: As speech generation technology continues to evolve, the risk of misuse through deepfake audio has become a pressing concern, which underscores the critical need for robust detection methods. However, many existing speech deepfake datasets fall short in terms of size, diversity, and linguistic coverage, limiting the ability of models to generalize effectively to unseen deepfakes. To address these limitations, we present SpeechFake, a large-scale dataset specifically designed for speech deepfake detection. With over 3 million deepfakes totaling more than 3,000 hours of audio, SpeechFake was generated using 40 different speech generation tools, including cutting-edge techniques, and spans 46 languages. This paper provides a detailed overview of the dataset’s composition and statistics, emphasizing its scale and diversity. Additionally, we establish baseline results for SpeechFake and explore how factors such as generation methods, language diversity, and speaker variation influence detection performance. We believe SpeechFake will be a valuable resource for advancing speech deepfake detection research, offering opportunities to explore new detection strategies and improve model robustness across diverse and evolving generation techniques. The dataset will be publicly available soon.
Primary Area: datasets and benchmarks
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 14000
Loading