New benchmark detects early-stage Parkinson's disease from speech

Early detection of Parkinson's disease (PD) through voice analysis has long been promising but lacks standardization. Studies vary widely in datasets, languages, tasks, evaluation protocols, and even definitions of early PD, making results impossible to compare. To solve this, Terry Yi Zhong and colleagues from Radboud University and other institutions present the first dedicated benchmark for speech-based EarlyPD detection. Their framework uses a speaker-independent split to ensure fair, replicable cross-method evaluation across researcher-accessible datasets. The benchmark covers three common speech tasks—sustained vowels, running speech, and diadochokinetic tasks—and tests methods under different training-resource scenarios (e.g., few-shot vs. full data). This rigor allows researchers to isolate algorithmic improvements from dataset artifacts.

The benchmark also delivers multi-dimensional evaluation breakdowns by dataset, aggregation level, gender, and disease stage, offering granular insights for clinical adoption. Results show that current models achieve moderate accuracy but degrade on finer-grained metrics, highlighting room for improvement. The authors provide a replicable reference and actionable guidelines for future work. By open-sourcing the benchmark and evaluation pipeline, they aim to accelerate robust, clinically meaningful EarlyPD detection from speech—a non-invasive, scalable screening tool that could transform early diagnosis and monitoring. The paper has been submitted to Interspeech2026.