How a spirit box works

The clearest description of a spirit box is the one hobbyists gave the first mass-market ones: "a scanning radio with the seek mute removed". A normal receiver with auto-seek mutes the audio while it hunts for the next station. Cut that mute wire on a late-1990s Radio Shack scanner (models 12-469, 12-470 and 12-150 were the classics, the so-called Shack Hack) and you hear everything the tuner passes: a quarter of a second of a talk show, a burst of hiss, a syllable of a song, hiss again. That is the whole instrument. The P-SB7 productised exactly that.

Gary Galka's P-SB7 (around 2009, still sold) is a synthesised AM/FM receiver — a microcontroller driving a PLL or DSP tuner — programmed to sweep instead of lock. The figures from its manual:

There is no public schematic; everything above comes from the manuals and from teardowns by hobbyists. The noise between hops is the detail most imitations get wrong: it is not a by-product, it is deliberately generated, and it is what keeps the background constant so that fragments jump out of it.

The P-SB11 is two tuners in one box: two simultaneous, independent sweeps, each AM or FM with its own direction and speed, one per stereo channel. The SB11-ANC variant added the line's first noise control, a per-channel filter that attenuates the hiss between stations so that fragments stand out — an adjustable squelch, in practice. Lower it and the box sounds like a radio; raise it and only the fragments remain.

Frank Sumption built the first box in 2002, and unlike the P-SB7 his design is fully documented: his 2007 notebook with complete schematics is public. The parts tell the story:

• Voltage tuning. AM/FM tuner modules salvaged from digital car radios (or Sanyo LA1135 / LA1816 chips), tuned by a varactor with a 0 to 8.2 V control voltage.
• The random version. The first boxes drove that voltage from a noise source — the reverse-biased emitter-base junction of a 2N3904 transistor, amplified about 240 times, low-pass filtered, rectified and averaged — so the tuning wandered at random across the band.
• The linear version (2007). Sumption concluded the randomness was unnecessary and replaced it with a triangle-wave generator that sweeps the whole band in 3 to 5 seconds in each direction, as a continuous analogue slide with no discrete steps. That is the key difference from the P-SB7, which steps.
• A physical echo chamber. The tuner's audio played through a speaker inside a closed box; a microphone on the opposite wall picked up the reverberant sound, and that was the output.
• "Raw audio". His central idea: a continuous, unsquelched stream of broadcast fragments and inter-station noise. He held that random material containing the frequencies of human speech works better than pure white noise — which is a statement about perception, and it is correct.

A syllable lasts roughly 150 to 250 ms. A dwell of 100 to 350 ms chops strong broadcasts — talk shows, adverts, sung vocals — into grains of phoneme or syllable size, separated by hiss. The speech-perception system joins successive grains, often from different stations, into "words" and "answers". This is auditory pareidolia, and it is not a flaw in the listener: it is the same machinery that lets you follow a conversation in a noisy bar.

The speed is a trade-off. Slow (250–350 ms) lets whole words through: more intelligible, more obviously radio. Fast (50–100 ms) approaches continuous noise: fewer intelligible fragments, but any "word" that is perceived feels more striking. The popular practice settles at 150–250 ms. The Estes method, now the standard protocol, uses 200 and never goes below 100.

The other "legendary" EVP device is not a scanner at all. The Panasonic RR-DR60 is a late-1990s dictation recorder that is technically bad in ways that manufacture voices: it samples at about 6 kHz with an ADPCM-type codec (watery, warbling artefacts), its aggressive automatic gain control turns the gain to maximum in silence and amplifies the circuit's own noise and induced interference through poor shielding, and its voice-activated system concatenates disjoint sub-threshold sounds into continuous audio. It generates its own noise floor with artefacts in the speech band — ideal material for pareidolia. Units resell for thousands.

A phone has no usable radio tuner, so an app cannot be a P-SB7. It can do the same thing with a different source: keep several live internet radio streams open, make one audible at a time on a fixed dwell, and inject noise in the gaps. That is what SpiritBox does, and the engine page gives the numbers. An app that plays recorded phonemes or words from a list is doing something else, whatever its name says.