REIMU's heterogeneous model detects speech deepfakes 10.8% more efficiently
New AI model REIMU cuts speech deepfake detection parameters by 10.8% while maintaining accuracy...
Speech deepfake detection gets a new efficiency-focused study: REIMU. The controlled study systematically compares conventional single-pass backbones, weight-shared recurrence, homogeneous hierarchical reasoning, and heterogeneous hierarchical reasoning across four Base-scale self-supervised learning (SSL) frontends. Experiments on the ASVspoof 2019 and 2021 evaluation sets show that recurrence and hierarchical decomposition don't inherently improve detection, while heterogeneous operator assignment—combining self-attention with linear attention—offers a more competitive configuration. Notably, this heterogeneous design remains competitive while using 10.8% fewer downstream parameters than the matched baseline, demonstrating its potential for parameter-efficient speech deepfake detection as realistic text-to-speech and voice conversion continue to pose growing challenges.
- REIMU uses heterogeneous hierarchical reasoning to detect speech deepfakes from SSL representations (tested on ASVspoof 2019/2021 datasets)
- Model reduces downstream parameters by 10.8% vs baselines while maintaining accuracy, proving parameter efficiency is achievable
- Combines self-attention and linear attention in a heterogeneous design, outperforming homogeneous/hierarchical alternatives
Why It Matters
Offers a 10.8% parameter reduction in speech deepfake detection while maintaining accuracy, critical for scalable voice authentication systems.