Abstract
There has been a surge of interest in the use of speech as a biomarker for a wide range of health conditions, based on the premise that neurological, mental or physical impairments affecting speech production can be objectively assessed by automated analysis. Recent advances in clinical speech artificial intelligence (AI) have applied supervised learning techniques, similar to those used in general-purpose speech technology applications, to support the diagnosis and monitoring of mental health, cognitive decline, and motor disorders. Although clinical speech AI offers significant promise as a scalable and low-burden tool for health assessment, it also presents unique challenges compared to other speech technology applications. These include the need for condition-specific speech elicitation tasks, limited and heterogeneous datasets, sensitive data collection protocols, diverse speech representation and modeling strategies, and uncertainty in ground-truth clinical labels. This overview synthesizes the emerging literature that addresses these challenges in the full clinical speech AI pipeline: from speech elicitation and recording to representation, model development, and deployment. We provide a practical review of the field, including the technical pipeline for clinical AI, the design of clinical speech tasks, data collection practices, traditional and clinically oriented speech representations, predictive modeling approaches, and ethical considerations. We also discuss open technical and translational challenges and outline future research directions. The goal is to equip researchers and algorithm developers with the knowledge needed to move clinical speech AI from experimental development to real-world clinical impact.