A Multi-Scale Low-Bitrate Speech Codec for LEO Communication With Deep Feature Preservation Losses

This paper introduces a low-bitrate neural speech codec specifically designed to address the bandwidth limitations of Low Earth Orbit satellite communication. To tackle the challenge of ensuring speech quality under such strict bitrate and bandwidth constraints, we propose a novel framework that effectively balances compression and perceptual quality. The first innovation is a multi-scale residual network that significantly enhances coding efficiency while maintaining low complexity. The second innovation involves a set of deep feature preservation losses that improve the perceptual naturalness of reconstructed speech. Experimental results demonstrate that our approach achieves high-quality speech reconstruction at a bitrate of just 1.5 kbps with a 24 kHz sampling rate, making it highly suitable for low-bitrate speech transmission in LEO communication scenarios.