Authors
Antonino Maniaci, Moreno La Quatra, Maria Stella Di Modica, Andy Bertolin, Maria Rosaria Barillari, Marco Lionello, Mario Lentini, Jérôme R Lechien, Stéphane Hans, Valerio Mario Salerno
Published in
Journal of voice : official journal of the Voice Foundation. Nov 06, 2026. Epub Nov 06, 2026.
Abstract
To compare, in a real-world multicentre clinical cohort, three analytic approaches for distinguishing healthy from pathological voices: classical Praat-derived acoustic analysis with conventional machine-learning classifiers; self-supervised learning (SSL) models used as frozen feature extractors; and SSL models fully adapted via end-to-end fine-tuning.
Prospective, multicentre, controlled observational study conducted at three Italian clinical centers, reported in accordance with STROBE guidelines.
200 Italian-speaking adults (100 healthy, 100 with confirmed laryngeal pathology) provided voice recordings of sustained vowel phonation, diadochokinetic repetition, and sentence reading. Three analytic pipelines were compared on a speaker-disjoint held-out test set (n = 40): (1) Praat-derived acoustic features with machine-learning classifiers; (2) frozen SSL embeddings (wav2vec2, HuBERT, voc2vec); and (3) the same SSL backbones fully fine-tuned end-to-end (E2E).
Patient-level accuracy ranged from 75.0% to 95.0% across 14 pipelines. The best-performing system was an E2E fine-tuned voc2vec/HuBERT model (sensitivity, specificity, PPV, NPV, and accuracy all 95.0%; AUC 0.992). The best frozen-SSL configuration (HuBERT-base, LibriSpeech) reached 90.0% accuracy (AUC 0.980), and the best classical pipeline (Praat + SVM) reached 85.0% accuracy (AUC 0.927), with high sensitivity but comparatively low specificity.
End-to-end fine-tuned SSL models achieved the highest point estimates among the pipelines compared in this pilot cohort, but the small test set and marked age/sex imbalance between groups mean these findings are preliminary and require confirmation in larger, independent cohorts.
PMID:
42995028
Bibliographic data and abstract were imported from PubMed on 07 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 16
- Comments 0