XLSR-MamBo: Scaling the Hybrid Mamba-Attention Backbone for Audio Deepfake Detection

A research team has introduced XLSR-MamBo, a novel modular framework designed to tackle the growing security threat posed by highly realistic AI-generated speech. The system addresses a key limitation in audio deepfake detection (ADD): while pure causal state space models (SSMs) offer efficient linear complexity, they often struggle with the content-based retrieval needed to spot subtle, global frequency-domain artifacts in synthetic audio. XLSR-MamBo's hybrid architecture combines an XLSR front-end for feature extraction with a synergistic backbone that merges the efficiency of SSMs with the global context capture of attention mechanisms.

The researchers systematically evaluated four topological designs using advanced SSM variants: Mamba, Mamba2, Hydra, and Gated DeltaNet. Their experiments revealed that the MamBo-3-Hydra-N3 configuration delivered competitive performance on the challenging ASVspoof 2021 LA, DF, and In-the-Wild benchmarks. A critical finding was that Hydra's native bidirectional modeling proved more efficient at capturing holistic temporal dependencies than the heuristic dual-branch strategies used in prior work. Furthermore, the model demonstrated strong generalization on the DFADD dataset, effectively detecting audio created by unseen diffusion- and flow-matching-based synthesis methods.

Crucially, the team's scaling analysis provided valuable architectural insights. They found that increasing the depth of the model's backbone effectively mitigated the performance variance and instability observed in shallower architectures. This scaling property is essential for building reliable, production-grade detection systems. The research, accepted by ACL 2026 Findings, validates the hybrid Mamba-Attention approach as a powerful method for capturing the telltale artifacts left in spoofed speech signals, offering a more robust defense against evolving audio forgery techniques.