Quantum Approaches for Dysphonia Assessment in Small Speech Datasets.

Area of Science:

• Medical signal processing
• Quantum machine learning

Background:

• Dysphonia, characterized by voice impairments, necessitates accurate diagnostic tools.
• Machine learning, particularly Convolutional Neural Networks (CNNs), shows promise for audio classification but struggles with limited datasets.
• Quantum-based approaches offer potential solutions for data-scarce scenarios.

Purpose of the Study:

• To compare the efficacy of standard CNNs, CNNs with random nonlinear layers (RANDOM models), and hybrid quantum-classical Quanvolutional Neural Networks (QNNs) for dysphonia assessment.
• To evaluate model performance using limited audio speech data.

Main Methods:

• Audio data preprocessed into Mel spectrograms (243 training, 61 testing samples).
• Developed six models: two CNNs, two RANDOM models, and two QNNs, with enhanced versions.
• Utilized angle encoding in the QNN quanvolutional layer for complex feature representation.

Main Results:

• QNN models consistently outperformed CNN models in classification accuracy (82.22%–89.26%) and convergence speed.
• The hybrid quantum-classical approach demonstrated robustness with limited data.

Conclusions:

• Quanvolutional Neural Networks (QNNs) present a promising, high-performing alternative for medical data classification, particularly for conditions like dysphonia.
• Quantum-based machine learning can enhance diagnostic capabilities in data-limited clinical settings.