MOSAIC: A Unified Multilingual Deep Learning Framework for Detecting Manipulated Speech Audio Using Geometry-Aware Foundation Model Representations

Audio deepfakes (AD) are increasingly used to spread misinformation, yet many detectors are evaluated in monolingual or narrowly multilingual settings and break under language mismatch. In this study, we take a step toward truly multilingual forensic audio deepfake detection (ADD). We conduct a systematic evaluation of pretrained speech encoders in multilingual settings. We observe that different pretrained encoders capture complementary spoofing cues arising from their distinct pretraining objectives. To leverage this complementarity, we propose MOSAIC, which fuses pretrained speech representations across multiple geometric spaces. MOSAIC performs gated additive fusion in hyperbolic, spherical, and Euclidean spaces, followed by self-attention for feature refinement. Experiments on large-scale multilingual benchmarks demonstrate strong crosslanguage generalization of the proposed framework. MOSAIC achieves SOTA performance, with MMS+PaSST attaining 0.02% EER on SpeechFake-MD and 0.07% on MLAAD. Through MOSAIC and complementary pretrained FMs, we attain top performance over individual FMs and standard fusion baselines. Such manipulated speech constitutes a form of audio misinformation that can mislead listeners across languages and platforms.