End-to-end Audio Deepfake Detection from RAW Waveforms: a RawNet-Based Approach with Cross-Dataset Evaluation

TL;DR

This work tackles the problem of robust audio deepfake detection under open-world conditions by proposing RawNetLite, a lightweight end-to-end model that operates directly on raw waveforms. The authors integrate three robustness pillars—domain-mix training across FoR, AVSpoof2021, and CodecFake; Focal Loss to emphasize hard samples; and waveform-level augmentations—to improve generalization without relying on large pretrained models or handcrafted features. In-domain results on FakeOrReal are near-perfect (F1 ~99.2%, EER ~0.29%), while cross-domain performance improves substantially when employing the proposed strategies, achieving up to the triple-domain configuration with an F1 of ~83.4% on AVSpoof2021 + CodecFake and an EER as low as ~16.4%. The findings underscore the value of diverse training data and tailored loss/augmentation strategies for resilient audio forgery detectors, with practical implications for deployment in real-world, open-world scenarios.

Abstract

Audio deepfakes represent a growing threat to digital security and trust, leveraging advanced generative models to produce synthetic speech that closely mimics real human voices. Detecting such manipulations is especially challenging under open-world conditions, where spoofing methods encountered during testing may differ from those seen during training. In this work, we propose an end-to-end deep learning framework for audio deepfake detection that operates directly on raw waveforms. Our model, RawNetLite, is a lightweight convolutional-recurrent architecture designed to capture both spectral and temporal features without handcrafted preprocessing. To enhance robustness, we introduce a training strategy that combines data from multiple domains and adopts Focal Loss to emphasize difficult or ambiguous samples. We further demonstrate that incorporating codec-based manipulations and applying waveform-level audio augmentations (e.g., pitch shifting, noise, and time stretching) leads to significant generalization improvements under realistic acoustic conditions. The proposed model achieves over 99.7% F1 and 0.25% EER on in-domain data (FakeOrReal), and up to 83.4% F1 with 16.4% EER on a challenging out-of-distribution test set (AVSpoof2021 + CodecFake). These findings highlight the importance of diverse training data, tailored objective functions and audio augmentations in building resilient and generalizable audio forgery detectors. Code and pretrained models are available at https://iplab.dmi.unict.it/mfs/Deepfakes/PaperRawNet2025/.

Figures (3)

• Figure 1: Architecture of the proposed RawNetLite model for audio deepfake detection.
• Figure 2: Comparison of fake F1-scores across test sets for models trained with Cross and Triple-domain strategies.
• Figure 3: Comparison of EER across test sets for models trained with Cross and Triple-domain strategies.