Data Augmentation for Pathological Speech Enhancement

A new study systematically tested data augmentation strategies to improve speech enhancement AI for pathological speakers. Researchers found noise augmentation delivered the largest gains, transformative methods provided moderate improvements, while generative augmentation often harmed performance. The research shows AI models still perform worse on pathological versus neurotypical speech, highlighting a persistent gap. The work evaluated both predictive and generative models across three DA categories using objective metrics.