Quantum-assisted speech enhancement via a two-stage hybrid neural network

Traditional deep neural networks (DNNs) demand substantial computational resources for single-channel speech enhancement, while quantum computers are constrained by limited qubit resources. To address these challenges, this study introduces HQCSE (Hybrid quantum-classical speech enhancement), a hybrid quantum-classical framework optimized for near-term quantum computers with low qubit usage. HQCSE incorporates innovative optimization strategies and a novel loss function that mitigates the sensitivity of traditional mean squared error (MSE) to outliers, thereby enhancing model robustness. A novel learnable quantum state embedding method, complemented by advanced feature preprocessing, significantly improves the efficiency of mapping classical data to quantum states compared to previous approaches. The framework's two-stage network design employs quantum neural networks in the preprocessing stage, reducing reliance on classical architectures like LSTM, and collaborates with LSTM in the post-processing stage to enhance information complementarity. Experiments on the TIMIT+ and VoiceBank+Demand datasets demonstrate that HQCSE achieves competitive performance with only half the parameters of baseline models and exhibits superior robustness on unseen datasets. Additionally, quantum noise resistance experiments confirm the framework's robustness against quantum circuit noise, highlighting its potential to advance quantum machine learning in speech enhancement applications.