Explaining Speaker and Spoof Embeddings via Probing

Publication
Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2025)

This paper presents a probing-based framework for analyzing the internal representations of speaker and spoof embeddings used in deepfake detection. The study investigates which dimensions of neural embeddings encode identity, spoof cues, and acoustic artifacts—offering insights into model interpretability and robustness across diverse spoofing conditions.