Quantifying Quanvolutional Neural Networks Robustness for Speech in Healthcare Applications

TL;DR

The paper investigates the robustness of hybrid quanvolutional neural networks (QNNs) for speech healthcare tasks, comparing against classical CNNs under four non-adversarial acoustic corruptions on AVFAD and TESS. It evaluates three quantum circuit templates (BEQC, SEQC, RQC) across multiple depths, using corruption metrics $CE$, $mCE$, $RCE$, and $RmCE$, and contrasts against CNN-Base, ResNet-18, and VGG-16 baselines. Key findings show QNNs outperform CNN-Base under pitch shift, temporal shift, and speed variation (with up to 22% gains in $CE$/$RCE$ under severe temporal shift), while CNN-Base remains more robust to Gaussian noise; QNN-Basic tends to excel on AVFAD and QNN-Random on TESS, with fear being the most robust emotion and neutral the most vulnerable under Gaussian noise. Additionally, QNNs converge up to six times faster than CNN-Base, illustrating potential practical advantages for deployment in healthcare contexts, though additive noise remains a challenge and deeper entanglement does not universally improve robustness.

Abstract

Speech-based machine learning systems are sensitive to noise, complicating reliable deployment in emotion recognition and voice pathology detection. We evaluate the robustness of a hybrid quantum machine learning model, quanvolutional neural networks (QNNs) against classical convolutional neural networks (CNNs) under four acoustic corruptions (Gaussian noise, pitch shift, temporal shift, and speed variation) in a clean-train/corrupted-test regime. Using AVFAD (voice pathology) and TESS (speech emotion), we compare three QNN models (Random, Basic, Strongly) to a simple CNN baseline (CNN-Base), ResNet-18 and VGG-16 using accuracy and corruption metrics (CE, mCE, RCE, RmCE), and analyze architectural factors (circuit complexity or depth, convergence) alongside per-emotion robustness. QNNs generally outperform the CNN-Base under pitch shift, temporal shift, and speed variation (up to 22% lower CE/RCE at severe temporal shift), while the CNN-Base remains more resilient to Gaussian noise. Among quantum circuits, QNN-Basic achieves the best overall robustness on AVFAD, and QNN-Random performs strongest on TESS. Emotion-wise, fear is most robust (80-90% accuracy under severe corruptions), neutral can collapse under strong Gaussian noise (5.5% accuracy), and happy is most vulnerable to pitch, temporal, and speed distortions. QNNs also converge up to six times faster than the CNN-Base. To our knowledge, this is a systematic study of QNN robustness for speech under common non-adversarial acoustic corruptions, indicating that shallow entangling quantum front-ends can improve noise resilience while sensitivity to additive noise remains a challenge.

Figures (8)

• Figure 1: The proposed Quantum-Classical robustness evaluation framework for speech classification. Clean audio is converted into log-Mel spectrograms and used to train QNN models (QNN-Basic, QNN-Strongly, QNN-Random) and classical CNNs (CNN-Base, ResNet-18, VGG-16) classifiers. Model robustness is evaluated on corrupted audio (Gaussian noise, pitch shift, temporal shift, speed variation) using metrics such as CE, mCE, RCE, and RmCE.
• Figure 2: Examples of three quantum circuit architectures for $n=4$ qubits: (a) BEQC, (b) SEQC, and (c) RQC, each illustrated for circuit depths $d=1$ and $d=2$.
• Figure 3: Architectures of (a) QNN and (b) CNN-Base classifiers. In (a), the quanvolutional layer includes encoding, quantum circuit, and decoding stages, where inputs are encoded by rotation gates ($R_y$), processed, and decoded into four classical channels (ch.1--ch.4). These channels are then passed through convolutional, pooling, flatten, and fully connected layers. The main difference lies in the first layer: QNN uses a quanvolutional layer, while CNN-Base uses a classical convolutional layer.
• Figure 4: Accuracy of QNNs (QNN-Basic, QNN-Strongly, and QNN-Random) and CNN-Base under four corruption types: Gaussian noise, pitch shift, temporal shift, and speed variation on AVFAD, TESS datasets. The corresponding circuit depths ($d$) for QNN-Basic, QNN-Strongly, and QNN-Random are summarized in Table \ref{['tab:depth_circuit']}.
• Figure 5: Confusion matrix for CNN-Base (red) and QNN-Random (green) on the clean and under the most severe level of four corruption types. The corresponding circuit depth ($d$) of QNN-Random is $d=1$.