A satellite in orbit played Doom after hearing a human voice

On 27 July 2026, a radio club in Oslo sent a spoken message to OPS-SAT PRETTY, an experimental ESA satellite. The message went out on the 23 cm amateur radio band, a slice of radio space that hobbyists are allowed to use.

On board, with no human in the loop, the satellite did five things. It caught the radio signal, pulled the voice out of the noise, turned it into text with an AI model, found the command in the text and launched Doom.

Then it made a "postcard" image of the moment and sent it down to Earth. The project is called PRETTY DOOMed.

The experiment was led by Georges Labrèche, the principal investigator at Tanagra Space.

There were two sessions on the same day.

• Morning: the team played a synthesized voice from a keyer (a device that plays a stored message) on a loop. The satellite detected the command and launched Doom.
• Evening: an operator spoke the command live into a microphone. The satellite detected that too.

The live message was: "LIMA ALFA FOUR OSCAR, PRETTY PLAY DOOM DOOM DOOM, PRETTY PLAY DOOM DOOM DOOM."

The first part is the club call sign, LA4O, which radio rules require for legal identification. The rest is the command, said twice, with "Doom" repeated each time to give the satellite more chances to catch it.

In real life Think of shouting your order across a noisy stadium. You say the key word three times, because you know some of it will be lost on the way.

Most satellites take orders as telecommands. Those are strict digital packets, sent during scheduled passes and built to rigid standards. This is different: it is plain spoken language, heard from the sky.

Here is the chain, all on the satellite's small onboard computer:

1. A software-defined radio records the signal as the satellite flies over. That means the tuning is done by code, not by fixed circuits, so it can be changed with an update.
2. It finds the voice buried in the noise and tunes in to just that narrow slice.
3. A speech-to-text model, an AI that turns audio into words, writes out what it heard.
4. A matcher checks the words for the command. It allows for mistakes, because the radio link garbles things.
5. On a match, the satellite plays Doom and builds the postcard.

The speech model is a 27 MB quantized English model. "Quantized" means its numbers were squeezed into a smaller size so it fits on limited hardware.

Not great, and that is part of the story.

The live evening command came out as "AND FAR BRIEFLY PLAYING DOOM STEAM DOOM PRECIPATED". It had two exact matches for the word Doom, although one is enough to trigger the game.

The morning detections passed only on the matcher's loose tolerance, with transcripts such as "PLAY DOON".

So the achievement is less about the satellite understanding a sentence. It is more about telling a human voice and one specific command apart from noise, while flying overhead.

The test was shared with a group in Poland, and the command finally got through on the last two runs, on the same day.

By the numbers: in run 6, the morning pass with the looped voice, three of six captures detected the command. In run 7, the evening pass with the live operator, one of five usable captures did.

Early on, the satellite heard the club, but most of what it got was static. About 170 kHz of radio spectrum reached the FM discriminator (the part that turns radio waves into audio) for a signal only about plus or minus 5 kHz wide. In plain words, the receiver was listening far too widely.

Offline analysis found a fix. Find the signal's peak within 100 kHz, shift it to the center, filter it and thin it out before turning it into audio. Tanagra says it is like tuning a car radio to one specific station.

Tanagra reproduced the fix and removed a redundant audio filter. Delivering the update took about a week. Being able to update the satellite's software was what made this possible.

This was a hobby setup, not a space agency dish.

• Radio: an Icom IC-9700, sending narrowband FM on 1296 MHz at about 80 W.
• Antenna: a Wimo helix antenna with 15.1 dBi of gain, pointed by hand.
• Where: a rooftop in Oslo, under the club call sign LA4O.

The Oslo group's campaign took 59 days and about 7,200 NOK, roughly $770, pizza included.

Earlier link tests ran from April to June 2026 with radio amateurs in Legnica, Poland, under the call signs SQ6RDP and SQ6QV. Tanagra says the experiment depended on the amateur radio community, who formed its ground segment.

The satellite's passes were scheduled and run by the OPS-SAT mission operations team.

Doom has become the classic "can it run this?" test for new hardware. It already ran on an earlier ESA satellite, OPS-SAT-1. This time the new part is how it started: a voice.

OPS-SAT itself was ESA's flying laboratory. ESA says it launched on 18 December 2019 and was about the size of carry-on luggage. Outside teams could rent time on it to test ideas that would be too risky for a regular mission.

Over about four and a half years, 134 teams from 26 countries ran more than 284 experiments on it. Strong solar activity in early 2024 increased drag and brought it down early, in May 2024.

ESA says the idea lives on in successor missions, including OPS-SAT VOLT and ORIOLE. The PRETTY experiment flew under the same Space Lab program of outside experiments.

The voice was only the test signal. Labrèche said the same setup could handle other things with a new model.

Tanagra proposes repeat runs that keep the original data and send it down to Earth. That would let researchers test the live-voice result properly, because the live voice was detected on just one of five usable captures.

Larger speech models might detect better, and the team could try them on the ground first. The long-term hope is a satellite that can listen to any type of radio signal, and one day pick up "mayday" calls or radio bursts from space.

The catch: this was one experiment on one satellite, and the model only looks for one trigger phrase. It is a demo, not a product.

• Nothing changes today: you will not be talking to satellites this week. This was a lab demo.
• Small hardware can do a lot: a speech model of 27 MB ran on a tiny orbiting computer.
• Hobbyists still matter: a rooftop, a handheld microphone and a hand-aimed antenna were enough to reach space.
• Updates save missions: the team fixed a bad signal with new software, without touching the hardware.