The Grade of Breathiness Index (GBI): A Fuzzy Logic Model for Vocal Assessment

To develop and validate a multiparametric acoustic model, based on fuzzy logic, to predict the degree of breathiness in dysphonic and nondysphonic voices of Brazilian Portuguese speakers.

This is a cross-sectional, retrospective study that included vocal samples of the sustained vowel [a] and connected speech from 300 participants (235 women, 65 men; mean age 36.47 ± 12.07 years), including both dysphonic and nondysphonic individuals. Five speech-language pathologists specializing in voice performed auditory-perceptual judgment of grade of breathiness (GB) on a visual analog scale from 0 to 100 points, using the VoxMore APJ plug-in. Based on these judgments, a Fuzzy Visual Analog Scale (Fuzzy VAS) was developed, which integrated the experts' scores, weighted by their reliability (intraclass correlation coefficient), and modeled perceptual uncertainty using fuzzy logic and a k-means algorithm. Subsequently, 47 acoustic measures were extracted from the vocal samples. A multiple linear regression model was constructed using stepwise variable selection, with the Fuzzy VAS as the dependent variable, to identify the acoustic predictors of GB. Model validation included the evaluation of statistical assumptions (normality, homoscedasticity, independence, and multicollinearity) and the calculation of the Mean Absolute Error (MAE) relative to Fuzzy VAS.

The developed acoustic model (Grade of Breathiness Index-GBI) demonstrated high predictive ability, explaining 80.49% of the variability in perceived GB. Four acoustic measures were identified as significant predictors: median smoothed cepstral peak prominence, jitterddp, GNE_3000 Hz, and Hfno_6000 Hz. The model presented an MAE of 8.04 points (on a scale of 0-100) between GBI predictions and Fuzzy VAS scores.

The GBI demonstrates robust predictive ability and high adherence to perceptual judgments of breathiness, with an adjusted R² of 80.49% and an MAE of 8.04 points. This index emerges as a reliable and objective tool for quantifying GB. It provides valuable insights into the underlying vocal physiology, aids in longitudinal monitoring, and serves as a teaching resource for training new clinicians in vocal assessment.