This month I’ve been working on an alternative fine timing estimation algorithm for RADE V2 kindly proposed by Jean-Marc. He suggested a pure DSP algorithm as an alternative to the hybrid DSP/ML algorithm I was previously testing. This is intriguing as it’s simple DSP and so avoids some of the complexities of ML, and also outputs a frequency offset estimate. Turns out it works pretty well. However, like my ML/DSP estimator it exhibits some jitter when tested with multipath channels. Unfortunately the jitter upsets the RADE V2 decoder which was trained with a constant (jitter free) timing offset. So I added a smoothing stage with hysteresis to avoid short term timing jitter but still allow it to acquire new signals rapidly.
Here is a plot of the ML network “loss” against SNR that illustrates the performance of the new timing estimator.
The Genie curves use the ideal timing estimate, and the FT curves the new fine timing estimator. The curves are just about on top of each other which shows that when we use the timing estimator there is very little additional distortion – just the result we want.
The next topic for this month was “signal detection”. How can we determine a RADE V2 signal is present so we can start decoding? Using the output from the timing estimator, I applied a statistical model to make a decision about signal/no-signal. This is working reasonably well, even on low SNR and multipath channels.
To help track RADE V2 development I’ve developed the table below which summarises the new algorithms we’ve had to develop. This is real R&D with many failure paths so it’s hard to estimate schedule, but we can track progress to get a feel for how much algorithm development work is left to do. Once all of the algorithms have been developed, we can integrate them into a set of Python tools and start testing over the air, for example using stored file tests. If our stored files tests are successful, the development becomes a bit more straight forward, e.g. a C port, packaging and documentation is time consuming but low risk.
| Function | Requirement | Status |
| Phase Equalisation | Handle phase distortion of HF channel, and fine timing offsets of +/-1 ms | OK |
| Fine frequency offsets | +/- 2 Hz | OK |
| Coarse Freq offsets | +/- 50 Hz | under test |
| Coarse and fine timing | timing to within +/-1 ms | OK |
| Signal detection | One false detect every 10 minutes | under test |
| Frame sync | Align Rx symbols with RADE decoder | OK |
| Amplitude equalisation | Keep Rx signal within dynamic range of ML | TBD |
| End of Over Detection | Reliably detect the end of transmission and “squelch” within 100ms | TBD |
I’m a new user; QRV for about a week now. The last few days I’ve had one QSO with Dieter VK3FFB and a few attempts, including today. His signal on 28.330 is right at the threshold for me; we made it work the first time but the last two days it was garbled and no communication was possible. It occurs to me that although 8 kHz of audio bandwidth sounds wonderful, a narrower audio bandwidth (e.g. 2.5 kHz) for DX work, if accompanied by a narrower RF bandwidth and improved sensitivity down to, e.g. -10 dB SNR or so, would do wonders for enabling DX work on FreeDV. I know this is never going to equal the sensitivity of WSJT-X but what are the possibilities of improving it? Thanks!
RADEV2 is supposed to reduce the RF bandwidth significantly and slightly improve minimum SNR (though I can’t promise any specific numbers). There’s a screenshot on https://freedv.org/david-august-2025-bbfm-write-up-rade-v2-ota/ that shows the difference vs. RADEV1 and regular SSB. Hope that helps!
Hi Edgar,
A similar thought has occurred to us, however the relationship between audio bandwidth and SNR is not straight forward due to the non-linear spacing of the freq bands in the vocoder features and the correlation between them. For example in classical vocoders only 20% more bits are required for 8 kHz audio compared to 4 kHz. So the gains in SNR are unlikely to be 10log10(8000/2500). It’s an R&D topic that would take some time to explore.
– David