A pure-C port of the RADE V1 and RADE V2 (Radio AutoEncoder) digital-voice modem, ported from the PyTorch reference implementation (https://github.com/drowe67/radae).
RADE V1 & V2 is a neural joint source/channel speech codec for HF radio: a neural encoder compresses speech features into a latent vector that is mapped directly onto an OFDM waveform; a neural decoder reconstructs the features at the far end. This repository ports the RADE Python logic (the encoder/decoder graph wiring and the full OFDM modem DSP) to C, and reuses the Opus DNN / FARGAN / LPCNet engine unmodified for the neural-network kernels and the vocoder.
The deterministic DSP core builds and self-tests with only a C compiler + libm:
cmake -S . -B build -DRADC_WITH_OPUS=OFF
cmake --build build
ctest --test-dir build
The full codec (encoder/decoder/vocoder) additionally needs the Opus DNN engine and the generated weight arrays.
BSD-2-Clause (see LICENSE). Reuses Opus (BSD-3), KISS FFT (BSD-3, Mark
Borgerding), and the Opus DNN kernels / weight-export tooling (BSD-2,
Amazon / Jan Buethe). Third-party copyright notices and license texts are
bundled in NOTICE and must accompany redistributions of the full codec.