Advancing stuttering detection: A systematic review and meta-analysis of artificial intelligence-based models

Artificial intelligence (AI) has been increasingly applied to the automatic detection of stuttering, but the literature remains methodologically and conceptually heterogeneous. This review synthesized research on AI-based stuttering detection, examined what these systems are designed to measure, and considered how those targets relate to lived-experience understandings of stuttering.

Following PRISMA guidelines, 44 studies were included. A narrative synthesis examined study characteristics, datasets, model approaches, task formulations, and evaluation practices. Random-effects meta-analyses pooled reported accuracy (22 studies) and F1 score (13 studies).

The literature showed a clear shift from handcrafted acoustic features and traditional machine learning classifiers toward deep learning, self-supervised speech representations, and end-to-end architectures. Task formulations expanded from binary detection to subtype classification, temporal localization, and severity-related estimation, but most systems operationalized stuttering through overt, listener-detectable speech behaviors rather than broader speaker-experienced dimensions of stuttering. Pooled estimates suggested strong reported performance under study-specific conditions (accuracy = 86.9%; F1 = 73.2%), but heterogeneity was very high (I² > 85% for both), indicating substantial variation in datasets, task definitions, and evaluation approaches.

Current AI systems primarily measure overt, operationally defined speech behaviors and therefore do not fully capture stuttering as lived and experienced by speakers, including dimensions such as anticipation, effort, and loss of control. Future progress requires speaker-informed target definitions, more transparent reference standards, and multimodal, externally validated models that better align AI outputs with experientially meaningful constructs.