Open-Set Source Tracing as Compositional Factors via Structured Prototypes
2026-07-03 • Machine Learning
Machine Learning
AI summaryⓘ
The authors study how to tell where synthetic speech comes from by looking not just at the machine that made it, but also what data and training details were used. They create a new system that separates these parts into different sections so they can better recognize new combinations that weren't seen before. This method helps the system work well even with limited examples and when facing speech from completely unknown sources. Their tests show this approach does better than previous methods.
source tracingsynthetic speechgenerative architecturetraining datastructured orthonormal prototypessubspace partitioningcompositional generalizationopen-set identificationfew-shot learning
Authors
Santiago Rubio, Antonio Almudévar, Antonio Miguel, Eduardo Lleida, Alfonso Ortega
Abstract
Recent research expands beyond binary anti-spoofing with the emergence of Source Tracing, the task of identifying the specific generative origins of synthetic speech. However, current research often equates a "source" with its generative architecture. We propose redefining a source as a compositional tuple of Architecture, Training Data, and other training factors affecting the generated speech. We propose a framework using Structured Orthonormal Prototypes to minimize class overlap and intra-class variance. Our Subspace Partitioning strategy splits the embedding into architecture and data subspaces, while a residual subspace captures stochastic variability, enabling "compositional generalization" for novel factor combinations. This approach improves performance for partially seen sources and maintains robustness in fully open-set scenarios. MLAAD evaluations for Few-Shot open-set Identification show our approach significantly outperforms angular-margin baselines.