Audio Transformer for Synthetic Speech Detection via Benford's Law Distribution Analysis

This paper introduces a novel approach that enhances synthetic speech detection by applying Benford's law analysis to the encoder embeddings. By leveraging Benford's Law as supplementary information alongside the existing embeddings, the model gains a detailed understanding of both content-related features and numerical distribution patterns. Our approach demonstrates superior performance on the ASVspoof 2019 LA dataset, achieving an AUC score of 0.921, while providing enhanced interpretability.