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

Area of Science:

• Speech processing
• Computational linguistics
• Artificial intelligence in healthcare

Background:

• Artificial intelligence (AI) is increasingly used for automatic stuttering detection.
• Existing research is methodologically and conceptually diverse, posing challenges for consistent application.
• A comprehensive synthesis is needed to understand current AI approaches and their alignment with lived experiences of stuttering.

Purpose of the Study:

• To review and synthesize research on AI-based stuttering detection systems.
• To examine the specific aspects of stuttering these AI systems are designed to measure.
• To evaluate how AI-measured targets align with the subjective, lived experiences of individuals who stutter.

Main Methods:

• A systematic literature review following PRISMA guidelines, including 44 studies.
• Narrative synthesis to analyze study characteristics, datasets, AI models, task formulations, and evaluation metrics.
• Random-effects meta-analyses to pool reported accuracy and F1 scores from relevant studies.

Main Results:

• A significant trend towards deep learning, self-supervised speech representations, and end-to-end architectures in AI models.
• Expanded task formulations beyond binary detection to include subtype classification, temporal localization, and severity estimation.
• High heterogeneity (I² > 85%) in pooled accuracy (86.9%) and F1 score (73.2%), indicating significant variations in methodology and definitions across studies.

Conclusions:

• Current AI systems primarily detect overt speech behaviors, not the broader experiential dimensions of stuttering (e.g., anticipation, effort, control).
• There is a need for AI models that better capture the lived experience of stuttering.
• Future advancements require speaker-informed target definitions, transparent standards, and validated multimodal models.