AP-GRPO: Anchor-Gated Phonetic Alignment with Policy Optimization for Pathological Speech Reconstruction
Researchers have developed a new framework called AP-GRPO, designed to reconstruct pathological speech by leveraging audible anchors within recordings. This method uses an anchor-gated reward to match reliable speech segments and an inter-anchor phonetic alignment reward to ensure recovered content is phonetically consistent with the corrupted speech. AP-GRPO demonstrates improved speech reconstruction across various disease conditions and adapts its anchor enforcement based on the severity of articulatory or linguistic impairment. AI