Related Experiment Videos
Automated Detection of Motor Speech Disorders and Subtype Classification
Fenqi Wang1, Rene L Utianski1, Leland R Barnard1
1Neurology, Mayo Clinic, Rochester, MN.
Medrxiv : the Preprint Server for Health Sciences
|July 30, 2026
Summary
Automated speech analysis can detect motor speech disorders (MSDs), early neurological disease signs. Advanced models like HuBERT show high accuracy in binary classification, offering a promising clinical tool.
Area of Science:
- Neurology
- Speech-Language Pathology
- Computational Linguistics
Background:
- Motor speech disorders (MSDs) are crucial early indicators of neurological conditions.
- Access to expert speech analysis is limited to specialized centers.
- Automated speech analysis presents a scalable solution for MSD detection.
Purpose of the Study:
- To compare different automated speech analysis models for MSD classification.
- To evaluate model performance on clinically relevant metrics using independent datasets.
- To assess the feasibility of automated MSD detection in clinical settings.
Main Methods:
- Compared static acoustic features, Phonet features, and self-supervised models (HuBERT, SSAST).
- Trained and evaluated models on 583 speech samples for binary and multi-label MSD classification.
- Assessed models using validation AUC and tested generalization on two independent datasets.
Main Results:
- Pretrained (HuBERT: AUC 0.95) and Phonet-based models significantly outperformed static acoustic features.
- Binary classification models generalized well to independent datasets (sensitivity 0.94, specificity 0.97).
- Multi-label classification showed strong performance (macro AUC 0.86) but limited threshold stability across datasets.
Conclusions:
- Automated detection of MSDs is feasible and shows clinical promise.
- Binary classification models demonstrate robust generalization for MSD detection.
- Multi-label classification for specific MSD subtypes requires further refinement for consistent performance across diverse datasets.