Hi all. Thanks for the interest. I'm going to post an impromptu blog post here just explaining what this is. The rather unfortunate ground stop today made this unexpectedly useful with the pilots here on Orcas just to get a sense of the chatter in our airspace re the situation.
This is an ETTUS B210 attached to a dual 4090 box. It's a polyphase channelizer that takes the entire airband, channelizes it using cuda code on GPU, demodulates it, detects transmissions, throws away data tx and then feeds the tx through a voice model I created for transcription. A separate LLM reads the transcriptions to create a situational awareness picture of the airspace.
I'm interested in open sourcing this and making it available for folks who can run an ETTUS B210 or high quality SDR with enoguh bandwidth and a local GPU machine, so that we can cover more airpsace.
This was born out of an incident we had not too long ago where a pilot around KORS (based at Friday Harbor) lost the engine in his 172, and put it in the water in front of a ferry and was picked up by the ferry. We're a pretty close bunch on Orcas Island and we only heard about it days later via social media. That felt completely unacceptable, and I set out to create a radio that could monitor everything in real-time, all the time, and alert on emergencies. That has extended to transcribing and creating situational awareness for a chunk of airspace.
I'm an IFR pilot. I fly a T206H regularly in IMC around here and I'm a ground instructor. You can email me at mark@defiant.com or find me on x at mmaunder if you'd like to connect. I'm particularly interested in hearing from folks who can run a station on their own hardware and are under bravo or a charlie or even just somewhere interesting re airspace.
That's so cool... I'm a pilot (have 3 planes) and would be interested in hosting one--or just running for my own edification, really.
I'll message you.
Awesome. Thanks.
Edit: OMG I just realized who you are. I used to read your blog obsessively before I became a pilot, and while I was working heavily on mysql. Thanks for that!!
> That felt completely unacceptable, and I set out to create a radio that could monitor everything in real-time, all the time, and alert on emergencies.
I just want to send a heartfelt thank you. My sister and her family live on Orcas, and my mom flies in on Kenmore every couple months, as she is too old to make the long drive. I often worry about the the whole thing, and track the flight on Flightradar24. I have watched too much Juan Brown.
From a tech pov, this is cool as heck. From a you are helping watch my mom on somewhat sketchy flights in case I am not available pov: I would like to really thank you for choosing to do this with your time and money.
To get at that, I would love a mayday/pan-pan type alert subscription system. Do you currently have a system for handling the fastest time to alert? Or, is this only retrospective currently? Or, have I missed something entirely?
Thanks that's very kind of you. Please make sure they know about the mercy flight program on the island.
It's real-time data and yes I'm adding in alerting. Right now I'm focused on improving the voice model and the AI part that processes it's output and once I get to place where I'm confident in its reliability, I'll add in alerting. Should be a few days.
First, What an amazing project! > a voice model I created for transcription. >> The model is outstanding. I literally spent 5 minutes listening and barely understood most of the comms, the model was extremely helpful. This looks frontier/next level. Is it possible to share how you built this model/tips&trixjs?
This is super cool! I've got something similar going for my police scanner, albeit seemingly less sophisticated (running Whisper and a 3050): https://illinois.scanner-frequencies.com/
I was thinking about adding railroad and KORD traffic - you've just convinced me to do so!
Oh woah this is cool.
My office looks out on the final approach into LGA, every once in a while I’ll spot a plane out of place (typically doing a go around) and spend twenty minutes with flight radar 24 and LiveATC trying to figure out why. This would make it so much easier!
I have questions! Does it really require all that hardware? I've spent the last week working on a P25 scanner with 2 RTL-SDRs on a Pi and transcription running on a Ryzen AI laptop. The Pi keeps up with recording fine but it is only 4-5Mhz bandwidth instead of the ~11Mhz you're monitoring. For me transcription only lags by a few seconds but if I ran it on my desktop with a 2080ti it wouldn't even be that much.
What STT model are you using? In my tests I've gotten the best WER with whisper v3 large but it's still not that good.
Is P25 not encrypted?
It can be. My local network has a few encrypted talkgroups for more sensitive stuff but most is not.
I would expect that the hardware involved is massive overkill.
The Ettus SDRs are very nice and are marked up enormously over the BOM cost. You’re paying for an integrated, supported system, with good engineering, not for the parts. These days, for this particular use, you could probably build your own direct (no downconversion) receiver with a decent filter and, say, a 400MSPS DAC. (This was not a thing at any reasonable price point 15 years ago when I last played with this stuff.) Or you can buy or build a little mixer and choose an appropriate LO frequency and use a much cheaper DAC. Or you can buy a cheaper SDR.
~15 years ago I amused myself for about an hour using a USRP (older Ettus gear) and MATLAB to do more or less this but for regular FM radio, one channel at a time, on a decent laptop of the era. The processing power needed to channelize and demodulate a few channels from I/Q data is very, very low. Even coming up with some kind of optimal demodulation scheme wouldn’t use many cycles. (Although I suspect that there is more space to try to jointly optimize demodulation and transcription with FM than with AM.)
It would be a fun project to try to kludge up an ESP32 of some sort to do this. The tricky bit would be that the WiFi/BT hardware doesn’t go to DC.
i'm amazed at the transcription quality here. i spent a couple minutes listening, just in awe at the often nearly unintelligble crackling over the speakers, and then a half second later it prints a transcription that seems totally plausible to what i just heard, but without the transcription i don't think i would have been able to understand it.
I've honestly never understood how pilots or the tower are able to understand anything that's happening over the radio. Not only is there always constant static and interference, but they all speak in short rapid fire bursts that even beyond the lingo and shorthand is hard to process. I'd never be able to get a pilot's license.
"Delta blarg-mumble-twenty-blurgle-fargle-*screeeeeeeech*eading-foxtrot-squidgle-skee thanks."
It's absolutely amazing that the AI model is able to transcribe all that noise.
90% of radio communications is standard and expected.
If I get the call sign and a few key words, I can normally figure it out. For the most part you kinda know what people are gonna say before they say it, or at least have it narrowed down to just a few possibilities. You also just get an ear for the narrowband AM sound after a few hours.
“(Garbled) n123a (garbled) joining (garbled) one three” heard at an untowered field is going to be “local traffic, Cessna N123a is joining the downwind leg for runway 13.”
Or if you’re on final approach and the tower calls you up, there are really only a few things they are going to tell you, most likely is “Cessna 123a cleared to land”.
Plus, I’m guessing you have o my listened to very busy feeds at commercial airports. Students at smaller airports will often announce that they are a student, and controllers and other pilots will try to slow it down and give them a little extra room and time.
No idea how AI does it, but I assume it’s not too hard since there are essentially unlimited hours of recording to train on.
i'm guessing the model is doing something similar, and has a fairly confident set of the possible phrases to map sounds to.
That feels like a recipe for a mixup due to going by assumptions rather than facts. Given all the regulations, rules, checklists, etc. otherwise present in aviation, I'm surprised. But I guess there are not that many known cases of serious mistakes in comms?
Sounds that way, but it isn’t an issue in practice. The first example I gave where you piece together garbled reception is an example of an advisory, non-required communication at a non-towered airfield. The communication isn’t required and just serves as a backup to standard VFR flying. At some fields you aren’t even required to have a radio in the aircraft. Reading between the garbled lines isn’t an issue for these purposes.
There’s redundancy built in for critical comms. For critical messages (e.g. “cleared to land”, ) the receiving party reads back what they heard. If a controller doesn’t get the right readback with all of the details, they will correct it. You do occasionally get errors that aren’t caught, or someone that does a readback and then has a brain fart and does something that they aren’t supposed to, but that’s inherent to humans in the system.
Also, keep in mind that what you hear on liveATC is not what the pilots and ATC hear. Typically, you get much better reception because you are either on the ground transmitting to an antenna on a tower right next to you, or you are transmitting from thousands of feet up. It’s not unusual at all to hear crystal clear radio calls from 100 miles away.
The comms system is surprisingly robust, especially since it can degrade gracefully. AM is much easier to receive at low signal strengths, if there’s some sort of emergency in the tower, they can and have just sent a guy out with binoculars and a handheld to get people on the ground. If you have an electrical issue in the plane, you can grab a backup handheld out of a bag, etc. AM radios have been standard since WWII so all planes have the same equipment. Pilots and aviation regulators think a lot about safety, but replacing the aviation radio system is just not something that comes up.
You mean "Typically, they get much better reception" right?
I suppose that’s ambiguous.
To be clear: I, the pilot, get better reception in the plane than what anyone might hear on liveATC recordings.
In the air it’s usually way more clear than the stuff you hear on liveatc etc. if I’m 5k above terrain, have a nice radio and antenna, I have a solid 50+ mile range with anything local being quite clear. Ground recorders have way worse reception unless someone rented space on a tower.
I spent a good 10 seconds chuckling at your very apt portrayal there.
Now I wonder if it can transcribe Cradle of Filth - Funeral in Carpathia?..
Well timed! SSL cert date 2026-11-12?
!! That's awesome, I'm across the water in Port Townsend & have been playing around with a pair of RTL-SDR's doing something similar with police & SAR bands. this is awesome work man.
ps: grew up in obstruction pass, I suspect we have some friends in common. Nice work dude.
Very nice! We just started building something similar for our SAR team so we could scroll back and look at any pertinent Sheriff/Fire/SAR transmissions we have missed or mis-heard.
The voice recognition model on Adrian's thing is amazing.
Thanks! I visit Spruce Goose a lot. Yum! My wife and I are friends with some folks in your neighborhood. You Port Townsenders are cool!
Same here in Seattle area; but I believe PSERN is flipping on encryption like yesterday so I completely abandoned working on something similar.
I wonder if you can partner with ham radio enthusiasts running their own Web SDRs… maybe have a way to point a pipeline at a websdr. Some of em do have air band.
cool stuff!