The authors collected a Malayalam speech corpus and used multidomain acoustic features with mRMR and grouped permutation importance to train 12 classifiers, achieving very high subject-dependent accuracy (~0.96) and moderate subject-independent generalization (accuracy up to ~0.75, macro-F1…

Offers a language-specific, speaker-robust speech biomarker and feature-selection pipeline useful for remote staging and monitoring in PD clinical assessments and trials, but it provides little mechanistic or therapeutic discovery insight.

OBJECTIVE: Speech provides a lightweight window into articulatory and phonatory impairment in Parkinson's Disease (PD), yet clinically reliable severity staging has to extend across speakers. This study develops a PD severity classification framework which compares Subject Dependent (SD) and subject-independent (SI) settings on clinically calibrated labels such as mild, severe, and Healthy Controls, using a newly collected Indian conversational linguistic phrase in Malayalam language. METHODS: The key acoustic features were extracted, and include prosodic, phonatory, perturbation, noise, and spectral-cepstral descriptors. The feature selection approach on SD is minimum redundancy maximum relevance (mRMR) and for SI, both mRMR and grouped permutation importance (GPI) applied separately with grouped inner cross validation for leakage-safe feature selection. Further evaluation carried out using trained and tuned 12 classical classifiers-K-Nearest Neighbors (KNN), Support Vector Machine-Radial Basis Function kernel (SVM-RBF), logistic regression (LR), Ridge, GaussianNB, quadratic discriminant analysis (QDA), Extra Tree, Random Forest, Gradient Boosting, AdaBoost, SVM-Linear, and Multi-Layer Perceptron (MLP). RESULTS: In SD evaluation, the scores were inflated in model SVM-RBF as Acc 0.961, micro-AUC ≈ 0.999 using mRMR approach. Under SI (unseen speakers) evaluation, GPI's top-20 features generated the best result from the SVM-Linear classifier with Acc 0.736 (macro-F1 as 0.682, micro-AUC 0.891), and the classifier LR produced the best accuracy of 0.747 (macro-F1 0.732, micro-AUC 0.883). In comparison to mRMR, GPI matched the macro-F1 in SI evaluation but improved accuracy/AUC (Area Under the Curve) on nonlinear models. CONCLUSION: This study shows the potential of the multidomain acoustic features that effectively classify the PD severity on speech task in both SD and SI settings. While SD evaluation achieved higher accuracy (∼0.961), the SI setup reflects the real-world generalization and showed strong performance with macro-F1 ranges from 0.55 to 0.73, accuracy up to ∼0.747. These results show relative performance to prior works and highlight the need for speaker-robust modeling, balanced tasks, and multimodal integration to bridge the SD-SI performance gap.