Discriminating Pathological Voice From Healthy Voice Using Cepstral Peak Prominence Smoothed Distribution in Sustained Vowel

This paper deals with cepstral peak prominence smoothed (CPPS) distribution and its descriptive statistics as possible indicators of vocal health status. A total of 41 voluntary patients and 35 control subjects participated in the experiment: all of them followed the same protocol, which includes three repetitions of the sustained vowel /a/ simultaneously acquired with a microphone in air and a contact sensor, the perceptual assessment of voice quality, and the videolaringoscopy examination. The fifth percentile and the standard deviation of CPPS distribution were the parameters included in the best logistic regression models for the microphone in air and the contact sensor, respectively. The selected CPPS parameters had a strong to good discrimination power: an area under curve of 0.95 and 0.87 has been found for the microphone in air and for the contact sensor, respectively. For each CPPS parameter, the repeatability has been also estimated and the Monte Carlo method has been implemented for the uncertainty evaluation of the discrimination threshold. Furthermore, preliminary recommendations for better accuracy and repeatability of future studies are provided: analyses on the main CPPS influence quantities and on the effect of the frequency content of the signal spectrum on the CPPS parameters have been provided.