Speaker-Independent Parkinson's Disease Detection from Speech Using Compact Acoustic Features and a Lightweight MLP

parkinson's disease (PD) is a progressive neurodegenerative disorder in which speech impairment can appear early and provides a practical target for non-invasive digital screening. This paper presents a compact and deployable framework for speaker-independent PD detection from smartphone speech, designed to better reflect real-world use where models must generalize to unseen individuals. Using the MDVR-KCL corpus, we focus on a single-speaker read-speech task and extract a lightweight set of acoustic descriptors capturing phonatory stability and spectral characteristics, including FO statistics, jitter, shimmer, harmonics-to-noise ratio, and MFCC-based summaries. A regularized multilayer perceptron (MLP) is trained with offline waveform augmentation applied only to the training split. Performance is assessed under group-aware, subject-disjoint evaluation across five random seeds. The proposed pipeline achieves consistent discrimination between PD and healthy controls, with Accuracy = ۰.۹۱۰ ± ۰.۰۵۸, ROC-AUC = ۰.۹۷۳ ± ۰.۰۲۶, and PR-AUC = ۰.۹۸۱ ± ۰.۰۱۹ (mean ± SD), while the most challenging split still maintains ROC-AUC = ۰.۹۲۷. These results suggest that compact, interpretable acoustic features combined with a lightweight classifier can provide strong speaker-independent performance on smartphone speech. The framework is intended as a reproducible baseline for scalable screening and motivates a future journal extension with calibration, broader baselines, and cross-condition validation.

کلیدواژه ها:

Parkinson's disease ، speech biomarkers ، smartphone speech ، speaker-independent evaluation ، acoustic features ، multilayer perceptron