<p>Laryngeal motility assessment is essential for diagnosing and managing laryngeal disorders. However, paralysis evaluations suffer from high inter-rater variability, necessitating a more objective and quantitative approach. This study introduces a novel AI-driven pipeline that leverages computer vision techniques to classify 155 video-laryngoscopies into unilateral paralysis (n = 68), bilateral paralysis (n = 50), and healthy laryngeal function (n = 37). Our approach includes several advancements over existing literature. We extract the vocal fold positions from each video and automatically identify the most informative, noise-cleaned video segments for classification. We define novel movement-based features that quantitatively capture the restricted mobility characteristics of laryngeal paralysis. These features are used to train two classification models using a 5-fold cross-validation strategy: one model for binary classification (healthy vs. paralyzed) and the other for multi-class classification (healthy vs. unilateral paralysis vs. bilateral paralysis). To assess the importance of these features, we conduct an ablation study using Shapley values. Our method achieves a precision of 0.83, sensitivity (recall) of 0.85, F1-score of 0.84, and balanced accuracy of 0.85 for distinguishing between healthy and paralyzed individuals. For multi-class classification (healthy vs unilateral paralysis vs bilateral paralysis), our model achieves a precision of 0.80, sensitivity of 0.83, F1-score of 0.81, and a balanced accuracy of 0.83. These results highlight the effectiveness of our method and underscore the relevance of our features, further validated by the ablation study. Our AI-grounded approach enhances the accuracy and reliability of automatic laryngeal motility assessment. By introducing novel metrics to quantify paralysis severity, we provide a more objective, reproducible, and clinically valuable evaluation tool.</p>