arXiv HF RADE paper
The paper “RADE: A Neural Codec for Transmitting Speech over HF Radio Channels” that Jean-Marc and I have been writing has been submitted to a conference, and posted on arXiv. This paper is a professional engineering level description of RADE – covering the problem we set out to solve, design, details on training, and evaluation. Part of our ARDC grant commitments to document our technology to a professional level. Thanks Jean-Marc for working with me on this paper, and Bill and Jan for reviewing.
During review of the paper, I found a potential issue in RADE V1 that may affect HF channel performance. The RADE encoder outputs QAM symbols, that are mapped to OFDM carriers in a certain sequence. I found a difference between the mapping used for training and at run time. Initially I though this was a bug, and we might be able to gain some improved performance by fixing it. Unfortunately experiments designed to address the issue made no difference in performance. However the exercise gave us some clues about how RADE handles multipath fading – what matters (and what doesn’t) with the RADE system and HF channels. It also showed the value of writing up the work in spotting issues with complex systems.
ML Frame Sync and Fine Timing
This month I worked on two more building blocks required for RADE V2, frame sync and fine timing. For RADE V1 we used the location of known (pilot) symbols to implement these algorithms. While this approach works, the injection of pilot symbols requires an overhead of about 3 dB of SNR and 500 Hz in RF bandwidth. RADE V2 is designed to work without pilot symbols, so the challenge is to develop ML versions of frame sync and fine timing that can work off the information that is present in the payload data symbols alone.
This month I developed a ML network to handle frame sync, informal tests show it works at a range of SNRs and channel conditions. This could be used to drive a sync state machine, and indicate if the receiver has valid output (squelch). It senses sync state every 20-40 ms, much faster than RADE V1 (120ms), which will lead to faster acquisition and a faster PTT turn around.
To address fine timing I’ve been exploring just how the prototype V2 models handle timing errors over OFDM. To train a ML system you need to tell it where the “goal posts” are, so a good understanding of OFDM timing is required to inform the training process. Now that I have a framework for fine timing, I’m designing and building a prototype ML fine timing estimator.
Baseband FM (BBFM)
This month I worked with Tibor Bece to design and prototype an analog FM simulator that we can use to compare with the RADE BBFM system. FM is pretty complex, so we have developed a linearised channel model that we use for training and simulation. It can simulate AWGN and land mobile radio (LMR) fading channels. I’ve also updated the draft BBFM paper that documents this project. Once we’re happy with the analog FM simulation, we will compare RADE with analog using the automatic speech recognition (ASR) framework, and generate some samples of the system running over real UHF radios.