[Paper Review] Parkinsonian Chinese Speech Analysis towards Automatic Classification of Parkinson's Disease

Hao Fang, Gong Chen+3 more|arXiv (Cornell University)|May 31, 2021

Voice and Speech Disorders18 references4 citations

TL;DR

This study proposes a novel Mandarin Chinese speech corpus and evaluates multiple machine learning and deep learning methods for classifying Parkinson’s disease (PD) from speech. Using free-form image description tasks, the approach achieves 94.0% accuracy—surpassing state-of-the-art results—demonstrating that natural speech outperforms structured tasks for early PD detection.

ABSTRACT

Speech disorders often occur at the early stage of Parkinson's disease (PD). The speech impairments could be indicators of the disorder for early diagnosis, while motor symptoms are not obvious. In this study, we constructed a new speech corpus of Mandarin Chinese and addressed classification of patients with PD. We implemented classical machine learning methods with ranking algorithms for feature selection, convolutional and recurrent deep networks, and an end to end system. Our classification accuracy significantly surpassed state-of-the-art studies. The result suggests that free talk has stronger classification power than standard speech tasks, which could help the design of future speech tasks for efficient early diagnosis of the disease. Based on existing classification methods and our natural speech study, the automatic detection of PD from daily conversation could be accessible to the majority of the clinical population.

Motivation & Objective

• To develop a new Mandarin Chinese speech corpus for Parkinson’s disease (PD) research, including diverse speech tasks such as free talk, DDK, and poem reading.
• To investigate the effectiveness of various machine learning and deep learning methods in distinguishing PD patients from healthy controls using speech signals.
• To evaluate whether natural, unstructured speech (e.g., image description) offers superior classification power compared to standardized speech tasks.
• To identify acoustic features most discriminative for PD-related speech impairments in Chinese speakers.
• To enable accessible, mobile-friendly early diagnosis of PD through automatic analysis of daily-life speech.

Proposed method

• Constructed a new Mandarin speech corpus with 34 PD patients and 34 healthy controls, recorded under uncontrolled real-world conditions (e.g., mobile phones, home environments).
• Extracted Mel-Frequency Cepstral Coefficients (MFCCs) as primary acoustic features, followed by feature selection using ranking algorithms to reduce dimensionality.
• Applied classical machine learning models (SVM with RBF kernel, KNN, Random Forest) and deep learning architectures (CNN, RNN, end-to-end systems) for classification.
• Used t-SNE visualization on PCA-reduced features (50D) to assess class separability between PD and healthy groups.
• Compared performance across three speech tasks: free talk (image description), diadochokinetic tasks (DDK), and structured reading (poems).
• Optimized model training with cross-validation and evaluated performance using classification accuracy (ACC).

Experimental results

Research questions

• RQ1 Can speech features from natural, unstructured speech (e.g., image description) achieve higher classification accuracy for PD than standardized speech tasks?
• RQ2 How do classical machine learning methods with feature selection compare to deep learning models in classifying PD from Mandarin speech?
• RQ3 Which speech task—free talk, DDK, or poem reading—yields the most discriminative acoustic patterns for PD detection?
• RQ4 To what extent do MFCC-based features and dimensionality reduction (PCA) enhance model generalization and separability in PD classification?
• RQ5 Can an end-to-end deep learning system effectively classify PD from daily-life speech without task-specific design?

Key findings

• The free talk (image description) task achieved the highest classification accuracy of 94.0%, significantly outperforming standard tasks (DDK: 83.5%, poem reading: 91.1%).
• The RBF-SVM model with feature selection achieved 94.0% accuracy, surpassing all previously reported state-of-the-art results in PD classification using speech signals.
• The t-SNE visualization confirmed clear separability between PD and healthy speech patterns in the selected feature space after PCA and t-SNE projection.
• The study demonstrated that natural speech tasks elicit more pronounced speech motor deficits in PD patients, enhancing discriminative power.
• Both classical machine learning and deep learning models achieved state-of-the-art performance, indicating robustness across methodological approaches.
• The results suggest that automatic PD detection from daily conversation is feasible and could be integrated into mobile health applications for scalable early diagnosis.