Half-Truth Audio Detection for Marathi (Low-Resource language)

Adapted a partially-fake audio (HAD-style) detection framework to Marathi — a low-resource language — enabling word-level fake speech detection and localization of manipulated audio regions. Built a custom dataset using state-of-the-art TTS with selective word-level synthesis; re-implemented missing HAD components to ensure full reproducibility and end-to-end evaluation.