Features of 3-D phase space attractors as descriptors of chaotic laws in voice signals for the AI-based detection of dysphonia

Pathological voice signals, especially when affected by artifacts such as hoarseness or wetness, can be hard to analyze with traditional transform-based methods. Thus, we employed 3D phase-space attractors stemming from non-linear chaos theory, to build a pipeline for Machine Learning (ML) based classification of dysphonic versus healthy voices. Spatial and analytical features were extracted from the attractors and paired with features related to the F0 of the signal, then two feature selection methodologies were employed and fed four different ML classifiers. Results show how even a small amount of features bring promising classification accuracies, in turn confirming the potential of attractors as descriptors of pathological voices. The best-performing model is a linear SVM fed with 10 features selected using Kruskal Wallis' test, and attractors related to pathological voices are more erratic, less regular, less prone to being watertight, and with higher volume and face area.