AI Ensemble of Vision Transformers Detects Voice Disorders With Record Accuracy

A team of researchers in Kerala, India, has built an artificial intelligence system that can listen to a person’s voice and flag some of the most elusive disorders of the larynx and the nervous system with an accuracy of 86.84 percent, a result that significantly outperforms any single model in its arsenal. The study, published in Neural Computing and Applications, describes how three of the most advanced vision transformer architectures available today were strengthened, combined, and calibrated into a single voting committee that reads the acoustic fingerprints of disease hidden inside mel-spectrograms, the visual representations of sound that have become the workhorse of modern audio analysis.

The conditions targeted by the framework are as varied as they are debilitating. Dysarthria, laryngitis, laryngozele, vox senilis, Parkinson’s disease, and spasmodic dysphonia all disrupt vocal patterns in characteristic ways, leaving distinct signatures in the frequency content, stability, and timing of speech. Because these signatures often emerge before patients or even clinicians notice overt symptoms, automated detection of vocal pathology has long been viewed as a promising avenue for early intervention. Yet the field has struggled with a persistent problem: individual machine learning models tend to excel at some disorders while failing badly on others, particularly acoustically subtle conditions such as laryngitis, where inflammation of the vocal folds produces changes that are easy to miss against the natural variability of human voices.

The researchers, led by R. Sreehari and M. S. Chinchu of Sree Buddha College of Engineering together with Rajeev Rajan of Government Engineering College Idukki, attacked this weakness with a strategy drawn from one of the oldest ideas in machine learning: ensembles. Rather than trusting a single neural network, the team trained three state-of-the-art vision transformers, DinoV3, EVA-02, and MaxViT, and then systematically evaluated more than a dozen ensemble techniques to determine how best to merge their opinions. The winning configuration, a Boosted Weighted Voting ensemble enhanced with Temperature Scaling for model calibration, learns the optimal voting weight for each constituent model based on its performance on a validation set, so that more reliable classifiers exert proportionally greater influence over the final diagnosis.

The technical pipeline begins with annotated speech recordings drawn from two openly accessible resources: the Saarbrücken Voice Database, a widely used German corpus of pathological and healthy voices, and the Italian Parkinson’s Voice and Speech dataset. From these recordings the team extracted mel-spectrogram features, two-dimensional images in which the horizontal axis represents time, the vertical axis represents frequency on a perceptually motivated mel scale, and pixel intensity encodes acoustic energy. This transformation is what allows vision transformers, which were originally designed to process photographs, to be applied to audio: the spectrogram is treated as an image, and the transformer’s self-attention mechanism learns which regions of time and frequency matter most for distinguishing healthy tissue from diseased.

Vision transformers themselves represent a relatively recent shift in deep learning architecture. Unlike convolutional neural networks, which scan images through local filters and build up understanding hierarchically, transformers divide an image into patches and let every patch attend directly to every other patch, capturing long-range dependencies in a single operation. DinoV3, the successor to the influential DINOv2 self-supervised learning framework, produces remarkably robust visual features without requiring dense labels. EVA-02 is a highly optimized transformer trained on massive image corpora, while MaxViT introduces a hybrid attention scheme that combines local windowed attention with global grid attention, making it efficient enough to process high-resolution inputs. Each of these architectures brings a different inductive bias to the task, which is precisely why combining them proved so powerful.

The boosting component of the final ensemble traces its lineage to the foundational work of Freund and Schapire, whose 1997 decision-theoretic generalization of online learning established that weak learners can be converted into a single strong learner by iteratively focusing on examples that previous models got wrong. In the present framework, boosting manifests as the adaptive assignment of voting weights: models that demonstrate superior validation performance on the difficult classes receive heavier votes, effectively concentrating the ensemble’s decision-making power where it is most needed. The authors report that this sophisticated weighting scheme was decisive in improving the classification of hard pathologies such as laryngitis, which individual models handled poorly.

Equally important, and less glamorous, is the calibration step. Modern neural networks are notorious for being overconfident, producing probability estimates that do not match their true likelihood of being correct. Temperature Scaling, introduced by Guo and colleagues in 2017, addresses this by learning a single scalar parameter that softens or sharpens the network’s output logits until the stated confidence aligns with actual accuracy. In a clinical context, calibration is not a nicety but a necessity: a system that claims 95 percent certainty when it is right only 70 percent of the time is dangerous, whereas a well-calibrated system can be trusted to escalate uncertain cases to human specialists. By applying Temperature Scaling to the ensemble, the researchers ensured that the weighted votes were cast on comparable, trustworthy probability scales.

The headline result, 86.84 percent accuracy across the full set of laryngeal and neurological disorders, represents a marked improvement over any individual backbone model, and the authors emphasize the system’s markedly improved ability to classify difficult pathologies like laryngitis. The work builds on a growing body of evidence that spectrogram-based transformer approaches are well suited to medical acoustics. Recent studies have applied audio spectrogram transformers to Parkinson’s disease classification from multilingual sustained vowel recordings, to depression detection from speech, and to the detection of voice and lung pathologies, while other groups have explored fisher vector representations of cepstral features for neurogenic voice disorders. The new study distinguishes itself by rigorously comparing ensemble strategies rather than simply stacking models, and by demonstrating that intelligent combination, not architectural novelty alone, drives the gains.

The datasets used in the study are openly accessible, which strengthens reproducibility and lowers the barrier for other groups to verify and extend the results. The authors also note that the work received no external funding and that the authors declare no competing interests. The research team spans Sree Buddha College of Engineering in Alappuzha, Government Engineering College in Idukki, and APJ Abdul Kalam Technological University in Thiruvananthapuram, reflecting a collaborative effort between experimental contributors who ran the experiments and drafted the manuscript and senior supervisors who shaped the concepts and edited the final text.

The clinical implications are considerable. Voice disorders affecting the larynx and the neurological control of speech are frequently diagnosed late, after patients have endured months of hoarseness, breathiness, or slurred articulation, and early detection is crucial for timely intervention and effective patient care. A system that can screen routine voice recordings and flag suspicious acoustic signatures could serve as a triage tool in telemedicine, particularly in regions where access to otolaryngologists and neurologists is limited. Vocal biomarkers for Parkinson’s disease, in particular, have attracted intense interest because speech changes often precede motor symptoms by years. The authors caution, implicitly, that their 86.84 percent figure, while impressive, still leaves room for error, and the calibrated confidence outputs of the ensemble are what make it plausible to deploy such a system safely alongside human judgment. What the study ultimately demonstrates is that the future of automated medical diagnosis may belong less to any single heroic model than to carefully engineered committees of models, each compensated for its weaknesses by the strengths of its peers, and each taught, through calibration, to know exactly how much it knows.

Subject of Research: Deep learning ensemble detection of laryngeal and neurological voice disorders from speech spectrograms

Article Title: Boosted vision transformer ensembles for automatic detection of laryngeal and neurological voice disorders

Article References: Sreehari, R., Abhinav, R., Kalidas, V. S., Sukumaran, N., Rajan, R., & Chinchu, M. S. (2026). Boosted vision transformer ensembles for automatic detection of laryngeal and neurological voice disorders. Neural Computing and Applications, 38(17), Article 707. https://doi.org/10.1007/s00521-026-12393-5

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12393-5

Keywords: voice disorders, vision transformer, deep learning, ensemble learning, mel-spectrogram, model calibration, DinoV3, EVA-02, MaxViT, Parkinson's disease, laryngitis, boosting