March 21st 2024 - Recently, I discovered RM Noise, an AI-powered program that is designed to completely remove and or reduce noise on SSB and CW transmissions.
RM Noise has been trained using a vast dataset of noise recordings, which are unique to each setup and location depending on where your at. You might have a high noise floor that isn't even caused by the atmosphere. I'm always on the lookout for software that not only goes with my skills with computers and radios, but this project caught my attention due to my experiences with AI as well. My friend near me was the first to show this off to me.
RM Noise's website if you'd like to check it out is here:
https://ournetplace.com/rm-noise/
At its core, RM Noise is an AI model trained on noise recordings as mentioned above. The process begins when the client program captures the noisy output from the radio. This audio is then sent away to the servers, where it undergoes real-time processing to strip away the noise. The cleaned audio is promptly sent back to the client, allowing for usually very immediate, hopefully better audio. Based on my testing, the latency seems to be around 0.200 to 0.300 ms which isn't too bad... It is for sure getting into the range where we can tell of a delay but for me I didn't seem to notice it when I was using actively for QSO's. It is rather high due to the processing time and then being sent but is still quite impressive that it's managed that time. However this does bring up one of the issues with the program and that is the need for a constant internet connection.
In the grand scheme, it may not be a very big deal.. considering how powerful of a rig you would need to run these models in general but it's worth mentioning. This dependency means that your usage into noise-reduced transmissions are tethered to the availability and reliablity of your internet connection. If your internet isn't very stable, you might encounter a good bit of packet loss and or very increased latency then the 0.200 and 0.300 ms times I saw. Which in that case, you can and will usually notice the delay. And if you have no internet, you just won't be able to use the software. This can also be a problem even if you don't live in a region that has a nearby processing server. Currently only three servers exist.
Again this is partly not exactly the program's fault with having to be ran completely online, the AI models are very VERY hard to run on consumer grade machines. You will usually need a very beefy computer, typically with lots of VRAM. This situation isn't unique to RM Noise. The landscape of AI models in general have been rough, let's take the latest advancements in AI-driven natural language processing or image generation for instance (I understand this isn't related to this model, but it's still something to point out). Models like OpenAI's GPT series or DALL�E are really neat in their capabilities but are anchored by the immense processing power behind them, making local deployment very challenging without substantial hardware. If they even release the weights and code behind them to make that even possible. RM Noise's documentation does seem to suggest that they may release the model to a degree. This is probably because of the next issue long term which will likely get worse as more users use the program.
The necessity for powerful hardware for running these AI models translates into financial ones. Maintaining the models to support such models *in general is usually costly, with expenses that could balloon alongside the computational demands. This brings into focus the economic long term issues of RM Noise's current free-to-use model. I don't want to come off as "that guy" but it is something to point out. Offering a service like RM Noise for free is really neat and makes this accessible to a wide audience. However, it does raise questions about how this can maintain it's operations over time without starting to impose some sort of fee to use. The processing power needed to clean up radio signals through AI, while providing clear and usually pretty good results, involves not just powerful servers but also likely a significant amount of electricity and maintenance.
In my tests, I found the AI model to work pretty well for my noise floor, it includes two different models currently for SSB and a few other models for CW and an experimental one for FM as well at the time of writing this. Again, it seemed to work very well, and the program has a customizable slider as well that lets you adjust how much AI filtering/processing is done. Of course this isn't going to magically pull someone that's terribly low in the weeds, but it helps a ton if you are able to hear them but that pesky noise is bothering you and your radio's built in noise reduction is doing not much to help. I can see this being also pretty helpful for people in very noisy areas, like a city. As long as you can somewhat hear them, this program should be able to help out.
I've also found it VERY helpful for the static crashes and random pops that happen. Which is very nice if your trying to have a nice long ragchew and you are just getting bothered by the constant static crashes.
Another neat feature that they mention is the "time shift", which offers the ability to rewind and replay portions of a transmission. To use it, you simply click onto the graph that is on the program on a timeframe in the past, the program seems to let you go back about 4 minutes in time. This feature can be useful in cases where you might've missed a piece of a convo and just want to go back, like maybe getting a callsign or something from a quick QSO. I've used it a few times and it seems pretty neat.
RM Noise is a neat example of how AI can be harnessed and this is also a good example of ham radio expanding into the software world. While it's still got a good bit to grow, the potential for clearer, to a degree or almost noise-free communications is pretty exciting.
I'd for sure recommend you check it out, again it isn't magical, but if you set your exceptations right, this is a fantastic program to use.