A Lightweight End-to-End Model for Detecting Audio Deepfakes Using Raw Waveforms

With the rise of generative artificial intelligence, synthetic voices—commonly known as audio deepfakes—have become increasingly convincing and difficult to distinguish from genuine recordings. These technologies pose major threats to privacy, media integrity, and digital trust. This study replicates and analyzes an existing RawNetLite-based approach for end-to-end audio deepfake detection using raw waveforms. The model eliminates manual feature extraction, directly processing 1D audio signals through convolutional and recurrent neural layers. Cross-dataset experiments were conducted using the FakeOrReal, AVSpoof2021, and CodecFake datasets to assess the model’s robustness. The results show strong performance on in-domain data and notable generalization improvements when combining diverse datasets and applying audio augmentation. This paper provides both a technical evaluation and a broader reflection on the ethical and social implications of audio deepfake technologies