Prototype networks improve and explain synthetic speech attribution
Synthetic Speech Attribution via Prototypical Networks
Sound
Summary
It can be hard to tell which computer made a fake speech clip because current methods act like black boxes, giving no clues about their decisions. The authors study a new kind of system that compares speech samples to example sounds it has seen before, making it easier to understand why it thinks a speech clip was made by a certain method. They tested their idea on a dataset with different languages and kinds of fake speech and found it works as well or better than existing methods. This shows you can both identify fake speech sources accurately and explain the reasons behind the choice.
What this means in practice
- •For security teams: Detect and attribute synthetic speech in security settings with interpretable reasoning to explain decisions.
- •For voice biometric developers: Integrate explainable speech attribution to better verify speaker and detect manipulated audio samples.
- •For media authentication services: Provide users with example-based explanations for identifying synthetic voices in digital content.$Commercial implications: Enables development of trustworthy voice authenticity verification tools for content platforms concerned with deepfake detection.
Authors
Viola Negroni, Paolo Bestagini, Stefano Tubaro
Abstract
Synthetic speech attribution aims to identify the generative system responsible for a speech signal, but current approaches typically rely on black-box neural networks that provide limited insight into their decisions. This work investigates prototype-based networks as an interpretable alternative, where predictions are grounded in comparisons with representative training examples. We adapt ProtoPNet to spectrogram-based speech representations and evaluate the proposed framework on the MLAAD dataset under closed-set, cross-lingual, and open-set conditions. Experiments show that prototype-based reasoning achieves competitive or improved attribution performance compared with the baseline while enabling example-based explanations. These results highlight that interpretability and performance can be jointly achieved in synthetic speech attribution through prototype-based modeling.