New framework traces synthetic speech sources by architecture and training data

A team led by Santiago Rubio at Universidad de Zaragoza has introduced a new framework for source tracing in synthetic speech that goes beyond traditional binary anti-spoofing. Instead of identifying a source solely by its generative architecture, the paper defines a source as a compositional tuple of Architecture, Training Data, and other training factors (e.g., hyperparameters). To handle the combinatorial explosion of possible sources, the authors propose Structured Orthonormal Prototypes (SOP) that minimize class overlap and intra-class variance. Their key innovation is a Subspace Partitioning strategy: the embedding is split into separate subspaces for architecture and data, while a third residual subspace captures stochastic variability. This design enables "compositional generalization" — the model can handle novel combinations of factors not seen during training, which is critical for real-world open-set scenarios.

Evaluated on the MLAAD dataset for few-shot open-set identification, the SOP approach significantly outperforms baseline methods that rely on angular-margin losses (like ArcFace). The paper demonstrates robustness both for partially seen sources (combinations missing one factor) and fully open-set settings (unseen architectures or new data). This work, submitted to the IEEE Spoken Language Technology Workshop 2026, offers a practical path to trace the origin of deepfake audio even when the exact model or dataset wasn't in the training set. It has implications for forensic audio analysis, deepfake detection, and attribution of generated content across evolving generative models.