Related Experiment Video
Updated: Sep 27, 2025
09:47
Author Spotlight: Advancing Alzheimer's Research – Exploring Early Detection and Multi-Omics Approaches
Published on: December 15, 2023
1.3K
Neurogenerative Disease Diagnosis in Cepstral Domain Using MFCC with Deep Learning
Norah Saleh Alghamdi1, Mohammed Zakariah2, Vinh Truong Hoang3
1Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah Bint Abdulrahman University, P.O.Box 84428, Riyadh 11671, Saudi Arabia.
Computational and Mathematical Methods in Medicine
|April 14, 2022
Summary
Deep learning models can detect voice pathologies linked to neurological disorders. A 2D convolutional neural network (CNN) demonstrated better generalization for early voice pathology diagnosis compared to 1D CNN.
Area of Science:
- Neurology
- Biomedical Engineering
- Speech Science
Background:
- Neurological illnesses like Parkinson's disease, multiple sclerosis, myasthenia gravis, and ALS can manifest motor and non-motor symptoms affecting speech.
- Voice disorders stem from disruptions in neural pathways controlling speech production, impacting systems like the corticospinal tract, cerebellum, basal ganglia, and motoneurons.
- Voice pathology detection technologies offer potential for early assessment and diagnosis of voice irregularities.
Purpose of the Study:
- To develop and evaluate deep-learning computational models for detecting voice pathologies.
- To compare the performance of 1-dimensional convolutional neural network (1D CNN) and 2-dimensional convolutional neural network (2D CNN) for voice pathology detection.
- To assess the models' ability to diagnose voice pathologies caused by neurological conditions or other factors.
Main Methods:
- Utilized voice recordings of sustained vowel /a/ from the German corpus Saarbruecken Voice Database (SVD).
- Applied data preprocessing techniques including padding and segmentation to voice signals.
- Implemented 1D CNN and 2D CNN models, incorporating convolutional layers and Mel-frequency cepstral coefficient (MFCC) feature extraction.
Main Results:
- The 1D CNN achieved a maximum accuracy of 93.11% on test data but exhibited overfitting during training.
- The 2D CNN demonstrated better data generalization with lower training and validation loss, achieving 84.17% accuracy on test data.
- The 2D CNN model outperformed state-of-the-art studies, suggesting that models trained on handcrafted features are superior for speech processing in this context.
Conclusions:
- Deep learning models, particularly the 2D CNN, show promise for the simultaneous detection of voice pathologies from neurological disorders.
- The 2D CNN's superior generalization indicates its potential for reliable early diagnosis of voice pathologies.
- Handcrafted features combined with deep learning, as seen in the 2D CNN approach, may be more effective for speech processing tasks than end-to-end feature extraction.
