Deep Neural Network for Automatic Classification of Pathological Voice Signals

Area of Science:

• Medical acoustics
• Machine learning
• Signal processing

Background:

• Pathological voice detection is crucial for early diagnosis and treatment.
• Computer-aided methods offer efficient screening solutions.
• Deep neural networks (DNNs) show promise in complex pattern recognition.

Purpose of the Study:

• To propose and evaluate an automatic pathological voice detection method using DNNs.
• To compare the DNN model's performance against support vector machines (SVM) and random forests (RF).

Main Methods:

• Extracted 12 Mel frequency cepstral coefficients (MFCCs) from voice samples.
• Developed a DNN comprising a two-layer stacked sparse autoencoder and a softmax layer.
• Utilized identical training and testing datasets for DNN, SVM, and RF models.

Main Results:

• The DNN achieved high performance metrics: 97.8% sensitivity, 99.4% specificity, 99.4% precision, 98.6% accuracy, and 98.4% F1 score.
• DNN classification results surpassed SVM and RF by at least 5% across all evaluated metrics.
• The DNN effectively learned advanced features from raw acoustic data.

Conclusions:

• The proposed DNN model demonstrates superior capability in distinguishing pathological from healthy voices.
• DNNs offer a powerful tool for pathological voice analysis, with potential for broader clinical application.
• Further research is warranted to explore DNNs in diverse experimental and clinical settings.