A0361
Title: Generalized additive models for multiclass detection of voice disorders by using acoustic features
Authors: Lizbeth Naranjo - University of Extremadura (Spain)
Carlos Javier Perez Sanchez - University of Extremadura (Spain) [presenting]
Victor Miranda - University of Auckland (New Zealand)
Josefa Hernandez - Universidad Politecnica de Madrid (Spain)
Abstract: Computer-aided diagnosis systems for detecting voice-related diseases from speech recordings require developing and using reliable statistical models. In a binary classification context, several approaches have addressed the problem of discriminating between healthy and pathological subjects based on acoustic features extracted from voice recordings. In these cases, the binary classification was used to distinguish between healthy and single pathology subjects or between healthy and a group of subjects with different pathologies. However, the multiclass problem has not been sufficiently addressed, as there is difficulty in discriminating between two voice pathologies. A generalized additive model with a variable selection procedure has been implemented and applied to classify subjects with vocal fold nodules, Reinkes edema, and without any pathology. An in-house built database of voice recording based on the sustained phonation of the vowel was used. A total of 30 acoustic features were extracted using various speech processing algorithms, covering a wide variety of feature types, including nonlinear ones. The approach achieved an accuracy of 93\%, compared to 71\% accuracy obtained with logistic regression. This highlights the potential value of the nonlinear nature of the generalized additive model in addressing multiclass problems in this context.
