Open-Set Source Tracing as Compositional Factors via Structured Prototypes

2026-07-03

AI Summary

Existing voice source attribution methods oversimplify the generation source as merely the model architecture, neglecting critical factors such as training data and thereby limiting generalization to novel, open-set composite sources. This work proposes modeling the voice generation source as a structured tuple comprising architecture, training data, and other training-related factors. To achieve disentangled representations, the authors introduce structured orthogonal prototypes and a subspace partitioning mechanism that separates embeddings into architecture-, data-, and residual-specific subspaces, trained with an angular margin loss. The approach demonstrates strong generalization under both partially visible and fully open-set scenarios, significantly outperforming baseline methods on few-shot open-set source identification tasks using the MLAAD dataset, thereby substantially improving source identification accuracy and robustness in open-set conditions.

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.

Problem

• Source Tracing
• Open-Set Recognition
• Compositional Generalization
• Synthetic Speech
• Generative Origins

Innovation

• Compositional Generalization
• Structured Prototypes
• Source Tracing
• Open-Set Recognition
• Subspace Partitioning