Towards Frequency Band Explainability in Synthetic Speech Detection

Abstract

Recent advancements in deep learning techniques have brought remarkable developments in synthetic media generation, leading to the creation of forged contents that are almost indistinguishable from real data. This phenomenon poses a new challenge for the multimedia forensics community, as the misuse of synthetic media can potentially cause adverse consequences. Regarding the audio field, several methods have been proposed to detect synthetic speech, but due to their data-driven nature, their results are often little interpretable. To overcome this limitation, the scientific community is focusing on Explainable AI (XAI) aimed at understanding the critical elements in a speech track that drive the predictions of the detectors. In this work, we address the task of XAI in synthetic speech detection and explore the critical factors that allow us to detect forged tracks generated by unseen techniques. Our results suggest that the artifacts of synthetic speech are contained in specific frequency bands and show how we can make the detection process more accurate by focusing on single spectral bands. We also generalize our findings to other detectors, showing how these can benefit them and improve their final classification performances.