Authors:
A. Shyam Sundhar, J. Angelin Jeba, B. Muthu Deepika, Vishwajeet Vivek Kharote, Utsav Kale, Jerrel S. Reyes, Randy A. Joco
Addresses:
Department of Information and Communication, Tampere University, Tampere, Pirkanmaa, Finland. Department of Electronics and Communication Engineering, S.A. Engineering College, Chennai, Tamil Nadu, India. Department of Mathematics, SRM Institute of Science and Technology, Ramapuram, Chennai, Tamil Nadu, India. Department of Computer Engineering, San Jose State University, San José, California, United States of America. Department of Cloud and Artificial Intelligence, Microsoft Corporation, Chicago, Illinois, United States of America. Department of Agricultural Engineering, Mindoro State University, Victoria, Oriental Mindoro, Philippines.
Abstract:
The proliferation of sophisticated deepfake audio generation technologies poses significant challenges to information security and media authenticity verification. This paper presents a comprehensive analysis of deep learning approaches for deepfake audio detection, examining the evolution from traditional machine learning methods to state-of-the-art transformer-based architectures. Researchers evaluate multiple detection frameworks across diverse datasets including ASVspoof 2019, ASVspoof 2021, ASVspoof 5, and FakeAVCeleb, demonstrating the critical importance of feature extraction methodologies and multi-dataset training strategies. Our results show that while traditional Mel-Frequency Cepstral Coefficients (MFCCs) with Support Vector Machines provide a baseline, self-supervised pre-trained models like Wav2Vec2 and Whisper significantly outperform handcrafted features. The proposed multimodal detection system achieves superior performance, with mAP scores reaching 0.85 at an IoU threshold of 0.5 and maintaining robust recall rates above 0.8. Researchers demonstrate that combining audio and visual analysis dramatically improves detection accuracy, particularly for sophisticated attacks. The study provides comprehensive benchmarking results and establishes a roadmap for transitioning from proof-of-concept systems to production-ready deepfake detection platforms capable of real-time analysis and batch processing integration.
Keywords: Machine Learning (ML); Deepfake Voice Detection; Support Vector Machines (SVMs); Multimodal Detection; Media Authenticity Verification; Deep Learning (DL).
Received on: 24/09/2025, Revised on: 13/11/2025, Accepted on: 22/11/2025, Published on: 07/06/2026
DOI: 10.64091/ATICR.2026.000344
AVE Trends in Intelligent Computing Research, 2026 Vol. 1 No. 2 , Pages: 125-137