This is our second update on Signal2Speech, our Hack the Nest 2026 project that lets someone answer questions with an EEG headset. If you missed the first post, the short version is that the app suggests replies, the wearer picks one with a P300 response (a brain signal that shows up about 300 milliseconds after the tile you’re watching flashes), and they confirm by closing their eyes. Nothing is spoken until they confirm. It has been about two hours since that post, and the biggest change is not one we planned.
The headset didn’t make it
At the end of the last post, we said the next step was connecting the g.tec Unicorn headset and testing on an actual wearer. That isn’t going to happen at this event because our headset was lost in transit. Judging for the Mind and Machine track, led by g.tec with PiEEG, runs on simulation, and there are similar chips available, so the app can still be judged. However, it meant we needed a different way to show that the brain side of the project actually works, not just the code around it.
We did this two ways. First, the evidence view can now play g.tec’s own hackathon recording of five real people doing a P300 speller task through our filters and averaging. Within about a minute, the response to target flashes clearly pulls away from the non-target response around 300 to 500 milliseconds. Our chooser’s method classified single flashes at 0.59 to 0.85 balanced accuracy across the five people, where 0.50 is chance. The banner always says the data was recorded by g.tec, not our headset or wearer, so it is never passed off as ours. Second, PiEEG’s own EEG simulator now streams into our g.Pype pipeline over Lab Streaming Layer (LSL, a standard way to send live sensor data between programs), and its simulated eyes-closed alpha burst confirms a preview. The stream uses a name unique to our laptop, so another team’s simulator on the venue Wi-Fi can’t drive our app.
The full live path is also built in case a headset turns up. --input live calibrates on the app’s own tiles, then the P300 picks and an eye closure confirms. We rehearsed that exact path on simulated data end to end, but nothing has been tried on a real wearer.
Claude got the demo question wrong
Once we added an API key, we found out that AI Answering had a real problem. When asked “Which county’s AQI rose more?”, every model we tried answered “Wayne.” That is wrong. Washtenaw rose 186 points, and Wayne rose 168. Wayne ended with the higher AQI, which is likely why the models mixed them up, but it was still hallucinated information.
Claude now gets a “Verified facts” block built from our EPA and NASA data, with the four AQI values, the changes, and the sources, and it is told to state only those facts and to say “I don’t know” otherwise. A reply can also be a map action now, like “Compare counties.” For those, the spoken line comes from the dataset and never from Claude’s words, so the numbers spoken always match the map. We checked it with real calls: the AQI question, “Do you want water or juice?”, and a question the data can’t answer, which got an honest “I don’t know.” We also switched to Claude Sonnet 5.5 because it answered in about 1.4 to 1.9 seconds, compared to 2 to 3 seconds for Opus, which matters when someone is waiting on a reply (it's also much cheaper, which helps).
Proving “speaks exactly once”
In the last post, we said only one place in the code can speak, and old decisions get thrown away. We wanted actual evidence for that instead of just saying it. A new fuzz test (a test that throws huge numbers of random inputs at the code) sends random event sequences at the selection logic and checks that it never speaks twice, never speaks without a fresh preview, and never lets an old decision act in a new round. A long run of 3,000,000 operations found zero violations. To make sure the test actually catches problems, it also breaks the code three ways on purpose, and it catches each one.
A bug in our own tooling
One of our git hook tests was making commits and changing settings in the developer’s real repository instead of its own scratch copy. When the pre-push hook ran the tests, git passed along variables pointing at the real repo, so the test ended up switching it to “bare” (which broke git status), replacing the developer’s name and email with “Hook Test,” and leaving a fake .env file staged. The fix gives that test a clean environment, and the hook now clears git’s variables before running anything. A new test checks on every run that it can’t reach an outer repo.
Smaller things
- Bundled audio: The two OpenOrbit answers now play from bundled recordings, so they sound the same on every laptop with or without Wi-Fi.
- Response times: Every confirm shows how fast the app responded. On a Mac, the map changes in 2 to 30 milliseconds, and speech starts in about 15 milliseconds.
- Crash-safe launcher: One command starts the demo fullscreen and brings it back if it crashes.
- Editable phrases: Saved phrases now live in
data/phrases.json, so a user or caregiver can change them without touching code.
What’s next
Next, we are rehearsing the full demo on the judging laptop with Wi-Fi off and recording the demo video. The script, shot list, and video graphics are already done. We also have to test our app using an Octopus chip, which doesn't have quite the same functionality as the chip that got lost in transit, but it is still good enough for proof-of-concept testing.
Top comments (0)