PD Voice Research
Research Documentation

About the Project

A comprehensive overview of the research background, dataset, methodology, model architecture, and experimental results.

Background

Why Parkinson's Voice Detection Matters

Parkinson's disease (PD) is the second most common neurodegenerative disorder worldwide, affecting over 10 million people and projected to double in prevalence by 2040. The condition is characterized by progressive loss of dopaminergic neurons in the substantia nigra, leading to tremor, rigidity, bradykinesia — and a range of non-motor symptoms.

A critical challenge is that current diagnosis relies primarily on clinical observation of motor symptoms, which typically manifest only after 60–80% of dopaminergic neurons have already been lost. Patients often go undiagnosed for years during which earlier intervention could slow progression.

Voice changes offer a unique early window into neurological health. Approximately 90% of individuals with Parkinson's develop voice or speech disorders — often preceding motor diagnosis by years. These changes include reduced vocal intensity, monopitch, vocal tremor, and increased aperiodicity — all measurable from a simple voice recording.

Non-invasive, low-cost, and scalable, voice-based screening represents a compelling direction for early biomarker research. Deep learning architectures — particularly convolutional and recurrent models — have demonstrated strong performance in audio classification, making them well-suited for detecting subtle spectrotemporal signatures of vocal pathology.

Research Question

Central Research Question

“Can a hybrid CNN-LSTM deep learning model, trained on mel-spectrogram representations of sustained vowel phonation recordings, reliably distinguish between individuals with Parkinson's disease and healthy controls in a subject-independent evaluation framework?”

Input modality
Sustained vowel /a/ recording
Representation
Log-normalized mel-spectrogram
Evaluation
Subject-independent cross-validation
Dataset

Dataset Description

The study uses a composite clinical voice database of sustained phonation recordings from Parkinson's patients and age-matched healthy controls, collected under controlled conditions with a calibrated headset microphone.

252
Total Subjects
128 PD, 124 healthy
756
Total Recordings
~3 per subject
22,050 Hz
Sample Rate
mono channel
~7 sec
Avg. Duration
sustained vowel /a/

Data note: Dataset values are placeholders for this research demonstration. A published study would include full provenance, ethics approval, and informed-consent documentation.

Methods

Processing Pipeline

The pipeline transforms raw audio into normalized mel-spectrogram inputs suitable for deep learning. Each step maximizes signal-to-noise ratio while preserving acoustic features most relevant to vocal pathology.

Step 01Data Collection

Sustained Vowel Phonation

Participants sustain the vowel /a/ for approximately 5–10 seconds at a comfortable pitch and loudness. This task isolates the voice production system, minimizing articulatory confounders.

Step 02Preprocessing

Audio Preprocessing

Raw recordings are resampled to 22,050 Hz mono. Amplitude normalization accounts for recording-level variability. Silence trimming removes leading and trailing noise.

Step 03Segmentation

Middle 5-Second Extraction

The central 5 seconds are extracted to exclude onset and offset transients — capturing the most stable, steady-state vocal production representative of neuromotor control.

Step 04Feature Extraction

Mel-Spectrogram Generation

Each 5-second segment is transformed via STFT with a 25ms window, 10ms hop length, and 128 mel filterbanks spanning 0–8,000 Hz.

Step 05Normalization

Spectrogram Normalization

Log-amplitude compression is applied, followed by per-channel mean and standard deviation normalization using training-set statistics for zero-mean, unit-variance inputs.

Step 06Classification

Deep Learning Classification

Normalized mel-spectrograms are fed into the CNN-LSTM model. CNN learns spectro-temporal features; LSTM captures temporal dynamics. The output head gives a Parkinson's probability score.

Model Architecture

Hybrid CNN-LSTM Architecture

The architecture combines convolutional layers for local feature extraction with recurrent layers for temporal modeling — capturing both spectral texture and temporal evolution of acoustic biomarkers.

Input
128×216×1
Conv2D Block 1
32 filters, 3×3
Conv2D Block 2
64 filters, 3×3
Conv2D Block 3
128 filters, 3×3
CNN Block
Reshape
Seq. of features
LSTM Layer 1
128 units
LSTM Layer 2
64 units
LSTM Block
Dense 128
ReLU + Dropout
Dense 2
Softmax
Classifier
Output
PD / Healthy

CNN Extraction

Three Conv2D blocks with BatchNorm and MaxPooling extract hierarchical spectrotemporal features.

LSTM Temporal

Stacked LSTM layers model temporal dependencies, capturing tremor periodicity and vocal instabilities.

Hybrid Design

CNN outputs are reshaped into a sequence and passed to LSTM, combining spatial and temporal biases.

Independent Eval

Training and testing use entirely different subjects, preventing data leakage and measuring true generalization.

Results

Experimental Results

Results on the held-out test set under subject-independent 5-fold cross-validation. All metrics are macro-averaged. Shown values are placeholders — replace with your experimental results.

Performance Metrics

Accuracy87.3%
Sensitivity (Recall)91.2%
Specificity83.5%
F1 Score88.6%
AUC-ROC92.3%

Model Comparison

Random Forest
68.9%
SVM + Handcrafted
71.8%
LSTM Only
74.1%
CNN Only
78.4%
CNN-LSTM (Ours)
87.3%

Accuracy on subject-independent test set (placeholder values)

Key finding: The hybrid CNN-LSTM outperforms both CNN-only and LSTM-only baselines, validating the complementary nature of spectral and temporal feature extraction for voice-based PD detection.

Limitations & Future Work

Limitations & Future Work

Current Limitations

  • 1Dataset size is limited — larger, more diverse datasets could improve generalization across demographics and equipment.
  • 2Recording conditions (microphone quality, background noise) vary and may affect real-world performance.
  • 3Binary classification (PD vs. healthy) does not account for disease severity or staging.
  • 4The model was evaluated on English-speaking populations; cross-linguistic validity is unknown.
  • 5Vowel phonation alone may not capture the full scope of vocal biomarkers.

Future Directions

  • Extend to connected speech and reading tasks for richer acoustic feature coverage.
  • Incorporate longitudinal data to track disease progression over time.
  • Explore multi-modal fusion with tremor data, gait analysis, and clinical scores.
  • Conduct prospective clinical validation with a certified IRB-approved study.
  • Develop a lightweight mobile-compatible inference pipeline for real-world screening.

Ready to See It in Action?

Try the interactive demo to upload a voice recording and walk through the processing pipeline.

Launch Interactive Demo