Comparative study of linear prediction parameters for quantification and monitoring of dysphonic voices

Linear prediction parameters have been used to study the vocal changes accompanying a case of laryngitis, the aim being to quantify the vocal degeneration so as to monitor the progress of the laryngeal disfunction. The effectiveness for this purpose of a number of parameters (predictor coefficients, autocorrelation function, area coefficients, long-area ratios, PARCOR coefficients and cepstrum) and metrics (the Euclidean and Itakura distances and the crosscorrelation function) are compared and discussed´in the paper. Of the parameters used the most intrinsically sensitive (as measured by the Kullback divergence) are the PARCOR coefficients. However, when a metric is introduced to measure the deviation from normal vocal performance, the combinations producing the best results seem to be the Itakura distance applied to the predictor and autocorrelation coefficients and the crosscorrelation of the cepstral coefficients, both of which appear capable of registering changes in the patient´s vocal condition which might pass unnoticed in a conventional noninvasive examination. The way seems to be open towards effective quantification of voice deformation for the diagnosis and treatment of laryngeal disorders.