Dual-gating mixture of experts improves speech deepfake detection generalization
Domain-Adaptive Dual-Gating Mixture of Experts for Generalizable Speech Deepfake Detection
Sound
Summary
Detecting fake speech is hard when the attacks or recording conditions change. The authors designed a new method that uses two different ways to analyze sounds—the raw audio and higher-level speech features. They use special filters and learned patterns to decide which analysis experts to ask for help. This approach works better than previous ones, especially when encountering new types of fakes or different environments.
What this means in practice
- •For security software developers: Build better speech authentication tools that detect unknown fake voices in diverse acoustic settings by routing audio through adaptive expert modules.
- •For voice assistant engineers: Enhance voice assistants’ ability to recognize and reject deepfake speech commands even when encountering new attack types or noisy environments.
Authors
Siqing Qin, Zhe Li, Kong Aik Lee, Man-Wai Mak
Abstract
Recent advances in speech deepfake detection (SDD) have leveraged the Mixture of Experts (MoE) to enhance generalization capacity. However, existing gating networks often overlook the acoustic and temporal cues of deepfakes. In this work, we propose a novel domain-adaptive dual-gating MoE (DADGMoE) framework for SDD under unseen attack types and acoustic conditions. Our innovative dual-gating mechanism leverages Sinc-layer-based filters to process both low-level acoustic signals (raw waveforms) and high-level speech representations from a large self-supervised learning (SSL) model. It further incorporates domain prototypes to guide expert routing based on implicit deepfake patterns. The lightweight affine experts process the routed inputs. Experiments show that our DADGMoE significantly outperforms the baseline, achieving up to a 40.8% relative EER reduction on challenging out-of-dataset benchmarks. This framework demonstrates superior generalization capabilities and efficient design.