FlowFake: Liquid Networks for Audio Deepfake Detection

A new audio deepfake detector called FlowFake uses a liquid time-constant architecture to tackle cross-dataset generalization, a persistent weakness in current detection systems, according to research posted to arXiv on June 17, 2026 [1][2]. The model, described in a preprint, departs from conventional detectors that aggregate fixed-window frame statistics. Its hidden state evolves through a learned ordinary differential equation, with per-neuron adaptive time constants that simultaneously resolve spectral cues at 10 milliseconds and prosodic cues at two seconds [2]. The architecture achieves formal bounded-input bounded-output stability and an integration error of O(dt^4) [2]. FlowFake operates with 34,000 parameters, roughly 0.01 percent of the parameter count of the self-supervised model Wav2vec2, yet matches its performance on cross-domain benchmarks [2]. On a four-dataset evaluation spanning ASVspoof2019, FakeOrReal, InTheWild, and MLAAD, FlowFake reached 75.29 percent accuracy on ASVspoof2019 when trained only on FakeOrReal, and 79.97 percent accuracy when trained only on MLAAD [2]. It outperformed RawGAT-ST and Whisper-DF on every evaluated pair [2]. The work addresses a structural problem in audio deepfake detection. Neural text-to-speech and voice-cloning systems produce synthetic speech artifacts that manifest as multi-timescale trajectory anomalies, and detectors trained on one synthesis pipeline routinely fail on unseen forgeries [2]. The researchers argue that fixed-window approaches misalign the architecture with the signal, a limitation FlowFake’s continuous-time dynamics are designed to overcome [2]. Deep learning architectures have been applied to speech recognition and natural language processing for years, with models ranging from recurrent neural networks to transformers producing results that in some cases surpass human expert performance [3]. The FlowFake preprint appears on arXiv, an open-access repository that hosts electronic preprints across mathematics, physics, computer science, and related fields [7]. As of November 2024, the repository was receiving about 24,000 submissions per month [7]. The source code for FlowFake has been made available on GitHub [2].