EVP: Recording, and Reviewing It Honestly

Electronic voice phenomena is a description of a filing location rather than of a cause. It means: a speech-like sound present in a recording that nobody noticed in the room. That is all the term commits you to, and it is worth keeping it that narrow, because the interesting work is in what you do next.

The Survey reviews submitted recordings and has done since the 1980s. In our experience between ninety and one hundred per cent of marked regions in any given submission are attributable to identified ordinary sources, and the identification is usually possible from the same recording. One representative file resolved forty-one of forty-two.

Recording

Uncompressed, or do not bother. A phone voice memo hands you a lossy file. A lossy codec is a perceptual model — it works by deciding which parts of the signal you cannot hear and discarding them, and quiet broadband detail thirty decibels down is the first thing to go. That detail is the entire subject of the exercise. Record WAV at 44.1 or 48 kHz.

Automatic gain control off. AGC pumps the noise floor up in the quiet and down after a loud event. Every one of those excursions is a slow swelling and fading of broadband noise around a transient, which is precisely the shape people mark as a whispered word. If the device cannot disable AGC, the device is a source of the phenomenon you are studying.

Off your hands. On a tripod, or on a folded cloth on a solid surface, never held. Body contact puts clothing rustle, pulse and handling noise directly into the diaphragm at levels far above the room.

Announce everything. Out loud, on the recording, at the time: that is my chair. That is me swallowing. Car outside. Someone's stomach. This feels ridiculous for about four minutes and then it becomes the most valuable habit you have, because it converts an unexplained noise into an explained one at the only moment when explaining it is easy.

Identify the building first. Before you record anything, find and listen to the fridge, the boiler, the pump, the extractor, the loft tank and the flue. Write down their cycles. A fridge compressor with a four-minute duty cycle will appear in your recording fifteen times an hour and you will not hear it in the room after the first ten minutes, because attention adapts and microphones do not.

Reviewing

Reviewing is where recordings become evidence or become nonsense, and it is almost entirely a discipline of what you refuse to do.

Raise the gain — and understand what that does. To review you need the noise floor audible, which means heavy compression with a low threshold, a great deal of make-up gain, and usually a band-pass across the speech range. What rises when you do that is everything that was under the threshold, which is the room. A room brought up thirty decibels is full of structure. It will sound like a building with people in it. That is not a finding; it is what noise sounds like when you make it loud.

Do not listen with the question in your head. If you are looking for a name, you will find the name. Review blind where you can: get somebody who does not know what was asked to listen to the marked regions cold, and take their answer seriously when it is "nothing".

Never play it to somebody and tell them what it says. This is the single most damaging habit in the hobby. Once a listener has been told the words, they cannot un-hear them and neither can you, and every subsequent listener you tell inherits the same certainty. If you want a second opinion, hand over the audio and nothing else.

Look at it as well as listening. A spectrogram will settle a great many arguments in five seconds. Human speech has visible harmonic structure and formant bands that move. A door, a knuckle, a pipe and a stomach do not.

What actually counts as a mark

The Survey uses three conditions and requires all three. They are worth stating because a threshold on level alone marks everything in a kitchen and nothing in a stone church.

1. Above the local floor — the floor tracked across the recording, not a fixed number. A quiet building and a noisy one need different thresholds, and the same building needs different thresholds at 9pm and 3am.
2. For the length of a syllable — roughly 90 to 680 milliseconds. Shorter is a knuckle or a click; longer is a vehicle, a plane, or something with a motor in it.
3. And the level moved while it did. This is the condition that earns its keep. Without it, the most heavily marked object in every recording ever made is the fridge, which satisfies the first two perfectly by being a steady hum of the right duration.

A mark satisfying all three is not a detection of speech. It is a region worth listening to. Our instrument prints that sentence next to the count everywhere the count appears, for the straightforward reason that a number with no rule beside it will be read as a score.

The instrument here

Instrument 09 records uncompressed, from the microphone to the file, with no codec anywhere in the chain, and writes a WAV you can drag into an editor. It runs to a fixed length and has no stop button, deliberately: a take you can stop is a take you stop at the moment you think you heard something, which is the moment you most need to have kept recording, and it makes every take a different length so no two can be compared.

If the capture is interrupted — a call, another app taking the audio session, the screen locking — it abandons the take and says so, rather than filing thirty seconds of file containing eleven seconds of room with the join invisible. A recording with a hole in it is not a recording, and losing a take beats reviewing a gap as an event.

Nothing you record leaves your device, and nothing is written to your device's storage either: close the tab and the audio is gone, which is why there is a download button. What goes to your service record is the measurements and a small envelope, never the audio.