Real-Time Morse Code Decoding: Exploring Domestic SoC Applications

Although Morse telegraphy via shortwave commu-nication has fallen out of widespread use, it maintains its unique value in specific fields such as military communications, emergency services, and navigation. We have designed a PyTorch-based neural network architecture of Convolutional Recurrent Neural Network(CRNN) and Bidirectional Gated Recurrent Unit (BiGRU), and the trained model enables fast decoding of Morse signals. In this paper, the pt-format model of PyTorch is converted to an RKNN model, which is then deployed to Rockchip's industrial-grade general-purpose System on Chip(SoC) RK3568. The real-time decoding efficiency of Morse signals is significantly improved through the implementation of a sliding window mechanism and NPU hardware acceleration. Empirical data from our experimentation reveal that on the RK3568 platform, the decoding time for a one-minute-long audio signal is significantly reduced from 12.63
$s$
using CPU decoding to 1.67 s with NPU decoding, which indicates an approximate seven times to eight times improvement in decoding efficiency.