R Schönweiler1, P Wübbelt, M Hess, M Ptok. 1. Klinik für Phoniatrie und Pädaudiologie, Medizinische Hochschule Hannover. Schoenweiler.Rainer@MH-Hannover.de
Abstract
BACKGROUND: The purpose of the study was to analyze if perceptual voice quality ratings of the well-known RBH rating procedure (a 4-point scale of roughness, breathiness, and hoarseness) covary with acoustical voice parameters. METHODS: 120 voice samples from subjects with healthy and hoarse voices were rated on the RBH-index in a multicenter study with 31 raters. Multivariate regression tree analysis classified the perceptual ratings as "gold standard". Voice samples were acoustically analyzed with a feature extraction method. Feedforward-networks were trained to selected acoustical parameters having highest "relative importance" in the regression trees. Based on the best classifier, a computer program consisting of 50 simultaneous working networks was developed. RESULTS: Mean probabilities for correct classifications were found at 0.65-0.85, implying a significance level over chance (0.25). Classifications of the program matched in 40% with a priori values in the categories roughness combined with breathiness, and in 65% in at least one domain. CONCLUSIONS: The new method described here provides a psychoacoustically based "objective" classification of hoarse voices, which seems to enable future analysis of new parameters (like GNE), which may even improve the present results.
BACKGROUND: The purpose of the study was to analyze if perceptual voice quality ratings of the well-known RBH rating procedure (a 4-point scale of roughness, breathiness, and hoarseness) covary with acoustical voice parameters. METHODS: 120 voice samples from subjects with healthy and hoarse voices were rated on the RBH-index in a multicenter study with 31 raters. Multivariate regression tree analysis classified the perceptual ratings as "gold standard". Voice samples were acoustically analyzed with a feature extraction method. Feedforward-networks were trained to selected acoustical parameters having highest "relative importance" in the regression trees. Based on the best classifier, a computer program consisting of 50 simultaneous working networks was developed. RESULTS: Mean probabilities for correct classifications were found at 0.65-0.85, implying a significance level over chance (0.25). Classifications of the program matched in 40% with a priori values in the categories roughness combined with breathiness, and in 65% in at least one domain. CONCLUSIONS: The new method described here provides a psychoacoustically based "objective" classification of hoarse voices, which seems to enable future analysis of new parameters (like GNE), which may even improve the present results.
Authors: Matthias Seipelt; Andreas Möller; Tadeus Nawka; Ute Gonnermann; Felix Caffier; Philipp P Caffier Journal: Biomed Res Int Date: 2020-01-23 Impact factor: 3.411