Spectral envelope and periodic component in classification trees for pathological voice diagnostic

This work investigates the effectiveness of features from the spectral envelope such as the frequency and bandwidth of the first peak obtained from a 30^th order Linear Predictive Coefficients (LPC) to identify pathological voices. Other spectral features are also investigated and tested to improve the recognition rate. The value of the Relative Power of the Periodic Component is combined with spectral features, to diagnose pathological voices. Healthy voices and five vocal folds pathologies are tested. Decision Tree classifiers are used to evaluate which features have pathological voice information. Based on those results a simple Decision Tree was implemented and 94% of all the subjects in the database are correctly diagnosed.