In this paper, we propose an approach for detecting and recognizing pathologies of vocal folds from high speed endoscopy videos to help the otorhinolaryngologist to draw up a diagnosis. Our method investigates the relevance of features extracted from high speed sequences of vocal folds. The first novelty of the work consists in considering different types of features supposed to reflect the shape and the dynamics of the vocal folds. The second originality relies on the analysis of the contribution of each feature to the recognition task. Finally, we have highlighted the impact of the segmentation of the region of interest from where the features are computed. Experiments have considered different pathologies such as polyps, paresis. Results show the benefit that can be drawn by resorting to computer vision tools coupled with explainable machine learning tools.