FlowFake: Liquid Networks for Audio Deepfake Detection

FlowFake is a new Liquid Time-Constant (LTC) architecture designed for audio deepfake detection, addressing the challenge of cross-dataset generalization in speaker verification. With only 34K parameters, FlowFake utilizes a learned ODE to adaptively manage time constants, effectively capturing both spectral and prosodic features. It achieves benchmark results of 75.29% and 79.97% on the ASVspoof2019 dataset when trained on different datasets, outperforming existing models like RawGAT-ST and Whisper-DF while being significantly smaller, which is crucial for practitioners aiming to develop efficient and effective deepfake detection systems.