Stuttering Speech Disfluency Prediction using Explainable Attribution Vectors of Facial Muscle Movements

Non-invasive detection is possible: The study demonstrates that stuttering can be predicted from facial muscle movements captured on video alone, without requiring controlled environments or invasive wearable sensors, making the technology more accessible and practical for real-world applications.

Explainability matters for medical AI: Rather than using a "black box" algorithm, the researchers employed explainable AI (XAI) methods to identify which specific facial muscles (cheek and lip muscles, p<0.005) are most important for prediction, allowing clinicians and patients to understand why the system makes its decisions.

Observable patterns precede stuttering: The research found that specific facial muscle movements occur before stuttering happens, suggesting these pre-speech behaviors reflect "arousal and anticipation to speak"—providing a window for early intervention rather than only detecting stuttering after it occurs.

Responsible AI requires transparency in healthcare: By making the model's reasoning transparent through attribution vectors, this work demonstrates how AI systems used for speech disorder detection should be designed to be interpretable to both healthcare providers and patients, rather than relying on unexplainable predictions.

Potential for personalized support: Early detection of stuttering moments through facial cues could enable real-time interventions or therapeutic support, though the paper suggests further research is needed to translate these findings into clinical practice.