Quantum Approaches for Dysphonia Assessment in Small Speech Datasets

TL;DR

The paper tackles dysphonia assessment under data scarcity and evaluates whether quantum-inspired learning via QNNs can outperform CNNs on small speech datasets. It introduces Quanvolutional Neural Networks that replace standard convolution with a quantum-processing layer, using angle-encoded 2x2 patches processed by a random quantum circuit and decoded by Pauli-Z measurements. Using Mel spectrograms from the PVQD vowel /a/ subset, the authors compare two scenario pairs (QNN1/CNN1 and QNN2/CNN2) across 10 training-size experiments and report that QNNs achieve higher accuracy and lower variance. The results support the potential of hybrid quantum-classical architectures to improve speech pathology classification on limited data and motivate further exploration of encoding choices and data-augmentation strategies.

Abstract

Dysphonia, a prevalent medical condition, leads to voice loss, hoarseness, or speech interruptions. To assess it, researchers have been investigating various machine learning techniques alongside traditional medical assessments. Convolutional Neural Networks (CNNs) have gained popularity for their success in audio classification and speech recognition. However, the limited availability of speech data, poses a challenge for CNNs. This study evaluates the performance of CNNs against a novel hybrid quantum-classical approach, Quanvolutional Neural Networks (QNNs), which are well-suited for small datasets. The audio data was preprocessed into Mel spectrograms, comprising 243 training samples and 61 testing samples in total, and used in ten experiments. Four models were developed (two QNNs and two CNNs) with the second models incorporating additional layers to boost performance. The results revealed that QNN models consistently outperformed CNN models in accuracy and stability across most experiments.

Figures (6)

• Figure 1: The speech processing.
• Figure 2: The quanvolutional layer.
• Figure 3: The proposed framework: (1a) Quanvolutional Neural Network 1 (QNN1) and (1b) Convolutional Neural Network 1 (CNN1), (2a) Quanvolutional Neural Network 2 (QNN2) and (2b) Convolutional Neural Network 2 (CNN2). The (1) gray, and (2) green boxes represent the models used in the first and second experimental scenarios, respectively. The only difference between CNNs and QNNs is the replacement of the convolutional layer by the quanvolutional layer.
• Figure 4: Comparison of testing accuracy of (a) 1st and (b) 2nd scenario.
• Figure 5: Comparision of testing accuracy (above) and testing loss (below) of QNN1 and CNN1 at 60 and 240 training samples.