RADE V1 AI-assisted C Port

Last month I mentioned Peter’s innovative work using Claude Code to perform a C port of RADE V1, and my concerns around Quality Assurance for AI generated code. The FreeDV project leadership team discussed this and we decided to apply the same quality standards to the C port as the reference Python implementation. We adapted a suite of automated tests developed for the Python version to the C port, and wrote a few new ones. Some of these were developed with the assistance of Claude Code. Through a team effort between Peter, Mooneer and myself these tests are now passing on the C port of RADE V1, which is now ready for initial deployment.

Claude Code for RADE V2 development

I have started using Claude Code for my development and it is indeed a game changer. It’s like having another DSP engineer to work with. It is more knowledgeable than myself in many areas, and is indeed teaching me some DSP tricks like SNR estimation for OFDM. Claude is much faster and more accurate than me at coding, especially around scripting and automated tests. This month Claude helped me prototype a End of Over (EOO) system and SNR estimator for RADE V2, various automated tests, and with refactoring the V2 Python code into a class.

The EOO system is a good example of AI assisted DSP development. The transmitter sends a EOO symbol at the end of it’s transmission. The receiver detects this symbol and cleanly stops decoding, preventing a trail of R2D2 noises at the end of the over. A first pass worked, but was unreliable, only detecting about 40% of the EOO symbols on multipath channels. I then “worked” with Claude on various tests and algorithms he suggested, until we had adequate performance. The constructive criticism went both ways – some of the ideas I had were flops, but I also had objections to some of Claude’s ideas. I learnt quite a bit, as Claude’s knowledge is encyclopedic and he had algorithm ideas I had never heard of. I quite like learning so this was an added bonus. Claude did all the coding and is way faster than me. Eventually we reached an algorithm that passes all the tests we have thrown at it so far, and works down to all but the lowest SNRs.

Turning to SNR estimation, here is a plot of the SNR estimator Claude helped me come up with. Turns out SNR is related to the autocorrelation of the cyclic prefix which we already compute. This works quite well for AWGN channel but (as usual) has significant errors for multipath channels. So I asked Claude to fit a correction that minimises the error over three channel – AWGN, multipath good (slow fading), and multipath poor (fast fading).

Tx Drive Levels and ALC

The next phase V2 development is stored file over the air (OTA) tests. For each test I generate a file consisting of a chirp – SSB RADE V1 – RADE V2 in samples, which are scaled so they have the same peak power.

As a first step I connected my HF transmitter to a spectrum analyser via a RF sampler and attenuator so I could measure the actual Tx power. I encountered a significant difference in Tx power for each sample, despite the peak drive level being the same. This improved (the difference was reduced) when I increased the drive level from 60% to 80% on my IC7200 test transmitter:

The 60% drive level followed the conventional FreeDV wisdom of setting Tx drive to “just tickle the ALC”. To explore this a little more I conducted some “over the cable” tests where I connected a FT-817 receiver to the transmitter via attenuators and coax (so no HF channel or noise to deal with). In this case I measured no difference in the measured loss metric of the received RADE V1 and V2 signals between two drive and power levels. This suggests that for modern FreeDV waveforms like RADE V1 and V2, the Tx drive level and amount of ALC is not critical. Indeed driving the radio a little harder so the ALC is engaged may increase the Tx power. It also shows that the path through the two radios introduces just a small amount of distortion to the RADE signals, a pleasing result. Of course this is just one result, with one radio, so I would encourage other experiments in this area.

Test	Condition	USB Drive	Power setting	ALC (bars)	Peak Power	RADE V1 Loss	RADE V2 Loss
Control	Software					0.108	0.082
Low power	OTC	50%	L	1	20W	0.108	0.079
High Power	OTC	80%	100%	4	70W	0.107	0.087

RADE V2 Stored file OTA testing

So with all the algorithms for RADE V2 in place, I started testing V2 over the air (OTA) using stored file tests, similar to tests performed during RADE V1 development. As mentioned above I generated a file consisting of a chirp – SSB RADE V1 – RADE V2 in sequence, then transmitted that from my home station to various KiwiSDRs around Australia. The chirp is used to estimate SNR of the channel, and SSB and RADE V1 are included for comparison. By sending them all together, we have them passing over roughly the same HF channel.

The purpose of these tests was (a) see if RADE V2 actually works over real world HF channels (b) compare it to RADE V1 to see if we get any improvement at low and high SNRs. The simulations predict a 3 dB increase in sensitivity at low SNRs. At high SNRs the ML network “loss” is significantly lower for V2, but we are not sure what that means in terms of speech quality.

Here are some samples from the OTC/OTA testing. The first set of samples is over the cable (OTC) so very high SNR. Can you hear any difference between RADE V1 and V2? Which one sounds closer to the input speech?

This set of samples is from a low SNR (approx -5 dB) path 2000 km across Australia on 14.230 MHz:

Objective Loss Measure for OTA tests
It can be hard to tell fine differences in speech samples using ad hoc listening tests such as playing the samples above. To help with this problem, the OTA test software also measures the “loss” between the vocoder features at the input of the RADE encoder, to the output of the RADE decoder. This can be considered a number that represents the distortion of speech passed through RADE. We use the same metric when training the ML networks, where the weights are adjusted to minimise the loss. We observe during training that speech quality increases as the loss is reduced. So we can argue that in general a lower “loss” represents increased speech quality. Here is a table of the “loss” for a selection of OTA and OTC samples:

Sample	Speaker	V1 software	V2 software	Channel	V1 OTA	V2 OTA
v208	David	0.102	0.077	NVIS	0.132	0.132
v209	David	0.102	0.077	NVIS	0.141	0.125
v211	David	0.102	0.077	DX	0.520	0.154
v216	Brian	0.115	0.086	OTC	0.123	0.096
v217	Brian	0.115	0.086	DX	0.311	0.198

It can be seen that when sent over the same channel at (almost) the same time, the RADE V2 samples achieve an equal of lower loss than RADE V1. It’s interesting to compare the loss values for v216 and v217 to the samples above – do you agree with the ordering suggested by the loss values?

Conclusions

My conclusions so far:

RADE V2 works over the air for the channels (40m ground wave and NVIS, 20m skip) and two speech samples tested so far. No obvious functional problems over a total of 17 tests.

At low SNRs there is some evidence of V2 performance gain (measured loss values), but it is subtle and may be hard to hear using ad-hoc listening tests. Likewise for high SNR voice quality. It is fair to say V2 is no worse than RADE V1. Both RADE V1 and V2 are doing better than SSB.

The RADE V2 waveform has a fraction of the bandwidth of RADE V1 and SSB (99% power contained in 860 Hz).

After a few laps with Claude the End of Over (EOO) system is working on most OTA samples. The SNR estimator is functional but not accurate, with performance varying across channel types. End users will need to be briefed about it’s accuracy (up to +/- 3 dB error).

Some features I still need to test/demonstrate are the lower latency and streaming text feature.