Comparative Analysis of CNN and RNN for Voice Pathology Detection.

Area of Science:

• Speech pathology
• Computational linguistics
• Biomedical engineering

Background:

• Acoustic analysis of voice is crucial for pathological speech diagnostics.
• Accurate detection of speech noise is essential for precise voice health metrics.
• Disease pathology can potentially be identified from voice characteristics.

Purpose of the Study:

• To investigate the role of computerized acoustic examination in early diagnosis and monitoring of pathological speech.
• To develop and evaluate deep learning models for disease detection using voice features.
• To assess the accuracy of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) in pathological speech classification.

Main Methods:

• Feature extraction was performed on the Singular Value Decomposition (SVD) dataset.
• A 27-layer neural network, combining convolutional and recurrent architectures (CNN and RNN), was employed.
• The dataset was split for training and testing, with 10-fold cross-validation used for performance evaluation.

Main Results:

• The CNN model achieved an accuracy of 87.11%.
• The RNN model achieved an accuracy of 86.52%.
• 10-fold cross-validation confirmed the classifier's performance.

Conclusions:

• Computerized acoustic examination shows significant potential for early pathological speech diagnosis and monitoring.
• Deep learning models, specifically CNN and RNN, demonstrate high accuracy in classifying pathological speech from voice data.
• The developed system offers a promising tool for objective assessment of voice health and disease detection.