// WHAT COMES OUT OF THE NOISE
The voices are real. They're the training data.
Whisper learned to transcribe from subtitled video. Asked to transcribe audio with no speech in it, it falls back on what that video says when nothing is happening: thanks, goodbyes, a narrator moving to the next step.
of 301,317 non-speech clips got text from Whisper, plain white and pink noise included.
of those hallucinations were "thank you". Another 10.32% were "thanks for watching".
distinct lines came out of our own spike run over generated noise before the app was built.
| LINE, AS WRITTEN | TIMES | ON THE STUDY'S LIST |
|---|---|---|
| Thanks for watching! | 55 | ◆ RESIDUE |
| I'm going to go to the next station. | 26 | NOT LISTED |
| Where are you? | 26 | NOT LISTED |
| I'll be right back. | 15 | ◆ RESIDUE |
| Oh, my God. | 9 | ◆ RESIDUE |
| "Around the corner of the mountain" | 7 | NOT LISTED |
| I'll see you next time. | 6 | ◆ RESIDUE |
| I can hear you. | 4 | NOT LISTED |
| I am here. | 3 | NOT LISTED |
// A SESSION
Choose a source, then listen.
A session is five steps. The second is optional, and it's the one people come back for.
-
1SOURCE
Generated noise, a radio-style sweep of noise bursts, or the room itself.
WHITEPINKBROWNSWEEPROOM -
2ASK
Hold ASK and speak a question. It becomes the model's prompt for what follows.
-
3LISTEN
Captions surface over the live signal. A phrase the model repeats stutters on screen.
-
4MARK
Tap a line to flag it, with its timestamp.
-
5SEAL
Keep the audio and transcript, and export a vertical clip of the captions over the static.
// THE VEIL
The veil is the model's decoding settings, on one dial.
Turn it toward THIN and the model gets hotter sampling, loses its silence checks, stops describing sounds and picks from more candidate words. Every setting the dial touches is listed, here and in the app.
Curve veil-v1, adopted from the yield spike. Try it.
- Sampling temperature
- 0.72
- Silence checks
- OFF
- Retries on a weak guess
- 0
- Sound descriptions
- SUPPRESSED
- Candidate words (top-k)
- 5
// ASK
Your question is the only steering.
Ask is transcribed normally, then handed to the recognizer as prompt text, the way Whisper is told what a recording is about. Whatever comes out of the static next is conditioned on it. Nothing else in the app nudges what it writes.
Is anyone here?
80 windows of noise, base model, after one question
The base model tends to answer. The small model tends to repeat the question back. Either way, it's the prompt at work, and the Codex says so.
// THE CODEX, PAGE ONE
Nothing here contacts the dead. A speech model raised on human video is asked to transcribe silence, and it answers with what people say when the screen goes dark.
ONE TAP FROM EVERY SCREEN IN THE APP
// HONEST BY DESIGN
Eerie is fine. Lying about the words isn't.
- NO STAGED LINESEvery caption came from the model during your session. Nothing is scripted or seeded.
- MIC ALWAYS SHOWNThe mic opens only in Room mode or while you hold ASK. The whole screen gets a green border while it's open.
- FILTERED BY DEFAULTWhisper's noise output can include ugly text. Slurs and sexual text are always withheld, shown as ▒▒▒▒.
- STAYS ON YOUR PHONEThe model runs on the device. Audio and transcripts leave it only when you export them.
- NO ACCOUNTNo sign-up, no analytics, no ads.
The green border is Apple's rule and ours.
// FREE, THEN ONE UNLOCK
Try it free. Pay once to keep it.
FREE
- Pink noise
- Three sessions a day
- Exports carry a DEAD//AIR band
- The full Codex
UNLOCK
- Every source, Room included
- Unlimited sessions
- Your past questions, across sessions
- Exports with a small credit you can turn off
One-time purchase. No subscription.
Something is transcribing.
DEAD//AIR · iPhone first, Android after
DEAD//AIR is an instrument for a documented glitch in speech recognition. It names no one and promises no contact with anyone. Whisper is a model by OpenAI, run on the device through WhisperKit.