<p>The classification of vocal pathologies is challenging due to the variability and nonlinear nature of acoustic signals. Traditional and recent methods rely on limited acoustic features, restricting their generalization in complex clinical scenarios. To address the limitation, the aim of this study is to develop a hybrid methodology that combines recurrence graphs and deep learning to improve the detection and manage uncertainty in acoustic signals. To achieve this, the Saarbrücken Voice Database (SVD), a public database containing recordings of both healthy and pathological voices, was used. The proposed methodology involves transforming each voice signal into a recurrence graph using intuitionistic fuzzy clustering, a technique that effectively captures the nonlinear dynamics inherent in acoustic signals and manages the uncertainty they entail. These graphs are then used to train a deep learning model, which is subsequently employed to classify voices as healthy or pathological. The experimental results indicate that this methodology enhances the model’s precision and its capacity for generalization, thereby providing an efficient solution for the classification of vocal pathologies that does not necessitate the implementation of complex preprocessing or feature selection techniques. Among the deep learning models that were evaluated, the Inception-v3 architecture emerged as the most effective, attaining an accuracy of 85% and an F1-score, suggesting that it can adequately generalize on data derived from recurrence plots.</p>