Weakly-Supervised Learning in Partially Spoofed Audio Localization

Partially Spoofed Audio Localization (PSAL) involves identifying segments of altered real speech, a task more intricate than verifying the authenticity of the entire utterance. Existing algorithms for temporal deepfake localization often depend on strong labels that provide millisecond-level annotations, which are challenging to obtain in practice. To address this challenge, we first propose a weakly-supervised framework for PSAL. This framework consists of a transformer encoder to extract expressive frame-level features, and a pooling aggregation network to obtain global predictions. Our approach allows for straightforward integration with other models. Experimental results demonstrate that our weakly-supervised method surpasses the ADD2023 Track2 fully-supervised LFCC-LCNN baseline.