| 講演抄録/キーワード |
| 講演名 |
2026-01-19 13:25
Multilingual Audio Deepfake detection: A Robust and Generalizable Hybrid Model Approach Candy Olivia Mawalim・Yutong Wang,・Aulia Adila・Shogo Okada・○Masashi Unoki(JAIST) EMM2025-102 |
| 抄録 |
(和) |
(まだ登録されていません) |
| (英) |
This paper introduces a robust approach to multilingual audio deepfake detection, addressing the critical challenge posed by increasingly sophisticated AI-generated voices within the context of the SAFE Challenge evaluation. To build a system capable of generalizing across diverse linguistic and acoustic environments, the study utilized a comprehensive evaluation corpus covering 17 languages and a broad spectrum of synthesis methods. The core contribution is a hybrid detection model that synergistically combines the established RawNet and AASIST architectures with language-agnostic representations derived from a multilingual self-supervised learning model. Furthermore, the efficacy of RawBoost data augmentation was explored to enhance robustness against real-world noise. Experimental evaluation confirms the system’s promising generalization capabilities, achieving approximately 73% balanced accuracy across the diverse multilingual data and on previously unseen synthesis algorithms. |
| キーワード |
(和) |
/ / / / / / / |
| (英) |
multilingual / deepfake / speech synthesis / hybrid model / / / / |
| 文献情報 |
信学技報, vol. 125, no. 317, EMM2025-102, pp. 7-12, 2026年1月. |
| 資料番号 |
EMM2025-102 |
| 発行日 |
2026-01-12 (EMM) |
| ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
| PDFダウンロード |
EMM2025-102 |