Discovering Hidden Vocal Subtypes: An Unsupervised Acoustic-Biomechanical Exploration of Voice Profiles.

OBJECTIVE

This study aims to explore latent acoustic-biomechanical patterns of voice production using an unsupervised multivariate approach, and to identify data-driven vocal profiles across individuals with amyotrophic lateral sclerosis (ALS) and nonneurological dysphonia.

METHODS

A cross-sectional sample of 100 individuals, including patients with ALS and individuals with nonneurological dysphonia, was analyzed. Sustained vowel phonation was recorded and characterized using 26 variables, including standard acoustic measures (fundamental frequency -fo-, jitter, shimmer, and harmonics-to-noise ratio (HNR)) and 22 biomechanical parameters. Principal component analysis was applied to investigate relationships among variables and reduce dimensionality. Unsupervised clustering was performed at both the variable level to identify functional groupings and the participant level to derive data-driven voice profiles. Cluster validity was assessed using internal indices. Post hoc statistical comparisons and chi-square tests were used descriptively to characterize between-cluster differences and their relationship with clinical categories.

RESULTS

The first five principal components explained 70.7% of the total variance, revealing structured relationships between acoustic and biomechanical features. Participant level clustering consistently supported a two-profile solution. Fifteen voice parameters differed significantly between profiles after false discovery rate correction, with the largest effects observed for shimmer, HNR, and the biomechanical parameter Pr11, reflecting differences in vocal stability and noise-related characteristics. The identified profiles were not significantly associated with clinical diagnostic categories.

CONCLUSIONS

An unsupervised multimodal analysis of sustained phonation revealed two coherent vocal profiles that transcend traditional diagnostic labels. These data-driven voice phenotypes may capture functional patterns of voice production and support future efforts toward more refined and personalized characterization of voice disorders.