VoNTN: Real-Time and High-Quality Voice Services Over Narrowband Non-Terrestrial Networks

Non-Terrestrial Networks (NTNs) are crucial for guaranteeing emergency communication services. However, providing voice services over narrowband NTNs remains challenging due to their extremely limited data rate. Therefore, this article presents a novel Voice over NTN (VoNTN) mechanism that combines a parametric voice codec and an artificial intelligence (AI) reconstruction model, achieving remarkably low data rate requirements of sub-1 Kbps. The parametric voice codec operates by first compressing and encoding vocal acoustic features at an ultra-low bitrate of 450 bps. These compressed features then enable real-time streaming inference through the AI reconstruction model, which synthesizes high-quality audio waveforms. The reconstruction model has a three-layer neural network specifically engineered for resource-constrained satellite equipment, achieving low inference latency with lightweight computational load. We implement and evaluate VoNTN in our testbed using an actual narrowband satellite. Experimental results show that VoNTN can provide a low-latency and high-fidelity voice service without high-performance computing hardware.