Using SincNet for Learning Pathological Voice Disorders

Area of Science:

• Artificial Intelligence
• Speech Processing
• Medical Diagnostics

Background:

• Deep learning, particularly Convolutional Neural Networks (CNNs), shows promise in identifying pathological voices.
• A significant limitation of current CNN models is their lack of interpretability, hindering clinical adoption for voice disorder detection.

Purpose of the Study:

• To develop an explainable deep learning system, SincNet, for pathological voice classification.
• To enhance the interpretability of voice disorder detection models by integrating learnable sinc functions.

Main Methods:

• Replaced the initial layer of a standard CNN with learnable sinc functions to create SincNet.
• Utilized sinc filters as a front-end signal processor to extract acoustic features.
• Evaluated the system on three distinct voice datasets from Far Eastern Memorial Hospital.

Main Results:

• The proposed SincNet achieved a 7% accuracy improvement and a 9% sensitivity improvement over conventional methods.
• Demonstrated superior performance in predicting pathological voice waveforms.
• Provided initial explanations linking system output to extracted speech features.

Conclusions:

• SincNet offers a more interpretable alternative to traditional deep learning models for pathological voice detection.
• The system's enhanced performance and explainability can facilitate the development of robust voice disorder diagnostic tools.