In voice assessment, the evaluation of voice quality is a major component in which roughness has received wide acceptance as a major subtype of abnormal voice quality. The aim of the present study was to develop a new multivariate acoustic model for the evaluation of roughness.
In total, 970 participants with dysphonia and 88 participants with normal voice were included. Concatenated voice samples of continuous speech and sustained vowel [a:] were perceptually judged on roughness severity. Acoustic analyses were conducted on the voiced segments of the continuous speech sample plus sustained vowel as well. A stepwise multiple linear regression analysis was applied to construct an acoustic model of the best acoustic predictors. Concurrent validity, diagnostic accuracy, and cross-validation were verified on the basis of Spearman correlation coefficient (rs), several estimates of the receiver operating characteristics plus the likelihood ratio, and iterated internal cross-correlations.
Six experts were included for perceptual analysis based on acceptable rater reliability. Stepwise multiple regression analysis yielded a 12-variable acoustic model. A marked correlation was identified between the model and the perceptual judgment (rs = 0.731, P = 0.000). The cross-correlations confirmed a high comparable degree of association. However, the receiver operating characteristics and likelihood ratio results showed the best diagnostic outcome at a threshold of 2.92, with a sensitivity of 51.9% and a specificity of 94.9%.
Currently, the newly developed roughness model is not recommended for clinical practice. Further research is needed to detect the acoustic complexity of roughness (eg, multiplophonia, irregularity, chaotic structure, glottal fry, etc).