RADE for Land Mobile Radio: A Neural Codec for Transmission of Speech over Baseband FM Radio Channels

In the 1990s Land Mobile Radio (LMR) systems evolved from analog frequency modulation (FM) to standardised digital systems. Both digital and analog FM systems now co-exist in various services and exhibit similar speech quality. The architecture of many digital radios retains the analog FM modulator and demodulator from legacy analog radios, but driven by a multi-level digital pulse train rather than an analog voice signal. We denote this architecture baseband FM (BBFM). In this paper we describe a modern machine learning approach that uses an autoencoder to send high quality, 8 kHz bandwidth speech over the BBFM channel. The speech quality is shown to be superior to analog FM over simulated LMR channels in the presence of fading, and a demonstration of the system running over commodity UHF radios is presented.