Spirit Box

A spirit box is a radio that refuses to settle. It sweeps across the band faster than any station can hold, so instead of a broadcast you get fragments: a syllable from one station, a consonant from the next, a burst of noise between them.

Your ear will not accept that. Speech perception is aggressive — it is built to pull words out of a noisy room full of other people talking, and it does not switch off simply because there is nothing there to pull. Given fragments at roughly the rate of speech, it will assemble words. Given a question, it will assemble an answer.

The effect has a name. It is auditory pareidolia, it is extremely well documented, and it is stronger when you already know what you are hoping to hear. That last part is worth sitting with before you use this instrument.

Instrument 07 sweeps a bank of synthesised sources rather than live radio, because rebroadcasting live radio is not ours to do and, more to the point, because the effect has never depended on the source being real.

Everything you hear is made here

There is no audio file anywhere in this instrument. The static, the stations, the whistles and every word are generated by your own browser, in real time, out of oscillators and filters. That is a plain statement of fact and it is the first thing to know about the instrument.

The words are made with a model of a vocal tract. A buzzing source — a wavetable shaped like the pulse a pair of vocal folds actually produces — is fed through four resonant filters, and the frequencies of those filters are driven to the formants of each speech sound in sequence. That is not an approximation of how a voice works; it is how a voice works. A vowel is its first two or three formants, and moving them is what turns one sound into another.

We could have used the speech engine built into your phone. We did not, for one decisive reason: its output cannot be reached. No browser will route it into the audio graph, so it cannot be filtered, swept or put behind the receiver at all — it would arrive over the top of the instrument as a clean assistant voice from a different piece of software. A synthesised tract lives inside the receiver, comes through the same bandwidth limit as everything else, and sounds the same on every device.

It is also articulate rather than intelligible, and that is the point rather than a limitation. A word you have to reach for is a word you believe.

This is the part that separates Instrument 07 from a button that plays a spooky word, and it is the part we would most like you to check.

The content of every word is synthesised. When it happens is a measurement. The instrument listens to your room continuously and tracks its floor — the level the room sits at when nothing is happening in it. Everything afterwards is measured as a distance above that figure, because an absolute level in decibels means nothing when microphone sensitivity varies by twenty decibels between handsets.

Three causes, and the panel names which

• Answered. Something happened in the room and then the room went quiet. The transient is measured as spectral flux against its own running average, so it works at any level, and the instrument will not answer until the room has settled again. This is the one you will find first: knock once, wait, and it answers within a second or two.
• Unprompted, room settled. Nothing happened, and the room has been sitting on its own floor for a while. Slow — about one word every twenty-five seconds in a genuinely quiet room.
• Unprompted, no acoustic reference. There is no microphone, so there is nothing to respond to. The instrument still works, and every word from that session is marked unprompted everywhere it appears.

The gate is a guarantee, not a rate

Above about eleven decibels over your room’s floor, the rate at which the instrument produces a word is exactly zero. Not small. Zero. Make continuous noise at it and it will not say a single word, however long you keep it up.

That is deliberate, and it is what makes the opposite behaviour mean anything. A rate that merely fell off in a loud room would be indistinguishable from luck. A hard gate is something you can prove to yourself in fifteen seconds, which is the only kind of claim this site is interested in making.

There is no random number generator in it

The word is chosen by a digest of what the microphone was measuring at the instant the utterance fired: level, floor, spectral centroid, flatness, all quantised. The same room in the same state tends to produce the same word twice, and the transcript prints the seed so you can see it happen.

The band is seeded the same way, from your room’s measured floor, which means a different room is a different band — and the same room is the same band when you come back to it a month later.

How clearly it speaks is also a measurement

The quieter your room is against its own floor, the more crisply the vocal tract articulates: the formant bandwidths narrow, the consonants come up, and the phones lengthen. Those are the real acoustic correlates of clear speech rather than a filter labelled “clarity”. Go quiet and the instrument becomes more articulate, and the panel shows you the figure while it happens.

Let it establish the floor. The first second of a session is the instrument measuring what your room does when nothing is happening. Stay still and quiet for it. Everything the instrument reports afterwards is a comparison against that figure, and a floor established while you were moving about is worth very little.

Watch the band scan. Sixty-four channels, and the ones carrying something fill in behind the tuner. Note where they are. Crossing a strong channel changes what you hear, and an operator who has not noticed where the stations are will spend the evening reporting the same three.

Listen to the voice test. Four words, clearly articulated, before anything else happens. That is what this instrument sounds like when it is not guessing, and it is the reference against which every ambiguous word afterwards should be heard. If you skip it you will spend the session comparing words to nothing.

Then ask something, and be quiet. This is the whole method. The instrument answers into the gap after a sound, so a question followed by silence is the correct procedure and a question followed by a discussion of the question is not.

Mark what is worth marking. Nothing is filed automatically. A transcript of every word a machine said is not a record of anything — it is the machine talking about itself. Mark the two or three that made you look up, and those go to your case file with their numbers attached.

Find the sweep rate that works in this room. Operators argue about this more than anything else and they are right to. Standard suits most rooms. Fast is better in a room with a lot of reverberation; slow is better in a very dead one.

What to expect, honestly

Most sessions produce a handful of words and none of them land. Every so often one lands squarely, at a moment that makes it feel deliberate, and that moment is genuinely startling. That is the instrument working correctly — and it is worth knowing in advance that the startling ones and the meaningless ones are produced by exactly the same mechanism.

Ghost Detector Lab is an entertainment site operated as the public terminal of the Thinplace Survey. Nothing here detects the paranormal. See the disclaimer, the privacy notice, and the method for every instrument on the bench.