Partial Editing Spoofing (PES) Database for Partial Deepfake speech Detection

Anonymous release for double-blind review. This repository provides the datasets used in our study on partial deepfake speech detection, covering multiple partial-forgery paradigms and fine-grained temporal annotations. The dataset includes audio samples and frame-level labels designed to support research on detecting localized manipulations in speech.This dataset covers three forgery mechanisms: Fully synthesized (Ful) forgery: the entire utterance is generated by modern TTS systems Cut-and-Paste (CaP) forgery: produced by splicing segments from a synthesized one into a genuine speech End-to-end Editing (Edi) forgery: produced by modern TTS systems supporting seamless neural editing Each audio file is accompanied by frame-level labels (20 ms per frame), indicating whether each frame is genuine or manipulated.