EVP Labs
How EVP Labs measures a sound
What the app measures
EVP Labs measures whether a short stretch of your recording has the acoustic shape of a human voice. It measures the shape of a sound. It does not measure what made the sound.
Those are two separate questions. Many ordinary things make voice-shaped sound: a person nearby, your own breath, a radio or television in another room, air moving through a vent, an animal, electrical interference picked up by the recorder. A high score means "this sounds voice-like and is worth a careful listen". It says nothing about the source.
What the app does not claim
- It does not say what made any sound.
- It does not identify a sound as paranormal, and it has no setting that could.
- It does not turn a faint or unclear sound into a clear one. Slow-down, looping and the level boost change playback only. Your original file is kept as recorded, with a fingerprint you can check.
The app can help rule a sound out. When it finds no voice-like structure in a moment where someone heard a voice, the likeliest explanation is that the listener's brain filled in a pattern, which is normal human hearing (see "Why people hear voices in noise" below).
How the measurement works
- The recording is read in one-second windows that overlap. Each window is turned into a picture of how loud each pitch range is over time, the same kind of picture a spectrogram shows.
- A small model that runs on your phone compares each window with six acoustic patterns: human speech (spoken or whispered), and five kinds of ordinary sound (knocks and impacts, rustling and friction, squeaks and ringing, electrical or engine noise, animal sounds).
- Neighbouring windows that fit the speech pattern are joined into events. An event counts as voice-like only when its best few windows agree on average: a real utterance is long enough to be heard by several overlapping windows, while a knock or a glitch tends to stand out in one. Each event gets a voice-like structure score from 0 to 100, an artifact risk score with the best-fitting ordinary explanation, a confidence band that depends on how far the sound stands above the background and how long it lasts, and the plain reasons.
- A score of 60 is the line for the review queue. Events a little below the line are kept in a separate faint list, labeled Faint, for a thorough second listen. Short or faint events always carry a caution.
- The same file gives the same result every time. Nothing is uploaded. The analysis runs on your phone.
How it is tested
A test needs a known right answer. Nobody can supply a recording that is known to hold a paranormal voice, so recordings claimed to be EVP are not used to train the app or to score it. The right answer we can know is human speech: words that real people spoke or whispered.
- Hidden voices. Short spoken and whispered clips are mixed into recordings of ordinary places at set loudness levels, from clearly quieter than the background to somewhat louder, at known times. The test then checks whether the app lists each one.
- People it has not heard. Every voice, every background and every ordinary sound in the test comes from people and recordings that were held back from training. The real recorded whispers used for testing were kept out of training completely.
- Recordings with no voice. The app also analyzes recordings that contain no voice, with knocks, rustles, bells, engines and animals in them. Every voice-like candidate it lists there is counted as a false candidate.
- Phone conditions. Most test recordings pass through the same compression a phone uses, and half the voices have room echo added.
- Fixed rules. The pass marks are fixed in writing and are not adjusted to fit the results. The sensitivity setting is chosen on one half of the voice-free recordings and the false candidate rate is reported on the other half.
This test used 864 voices hidden in 144 two-minute recordings, plus 90 two-minute recordings with no voice in them.
Measured results for this version
Ranges in grey are 95% confidence intervals: the span the true value is likely to fall in given the size of the test.
False candidates
In recordings with no voice, the app listed 4.0 false voice-like candidates per hour (1.5 to 8.7, 6 in 1.5 hours). Expect some in any long recording. This is why every candidate needs a person to listen.
Hidden human voices found
| Hidden voice and its loudness | Share found |
|---|---|
| Spoken words | |
| Quieter than the background | 77% (70% to 83%, 112 of 145) |
| As loud as the background | 96% (86% to 99%, 45 of 47) |
| Louder than the background | 94% (87% to 97%, 90 of 96) |
| Real recorded whispers | |
| Quieter than the background | 43% (36% to 52%, 63 of 145) |
| As loud as the background | 46% (32% to 60%, 21 of 46) |
| Louder than the background | 78% (69% to 85%, 76 of 97) |
| Computer-made whispers | |
| Quieter than the background | 75% (68% to 82%, 107 of 142) |
| As loud as the background | 96% (87% to 99%, 49 of 51) |
| Louder than the background | all 95 found (true rate likely above 96%) |
Computer-made whispers are spoken recordings converted to whispers by removing the pitch of the voice. They are easier for the app than real whispers, so read the real-whisper row as the honest one.
The real whispers come from 35 separate recordings. Clips cut from one recording share a person, a microphone and a room, so the fair question is how the app would do on a new recording. Counted that way the app found 56% of real whispers (range 47% to 64%). It found at least half of the whispers in 27 of the 35 recordings and none in 2.
A second set of real whispers
The whispers in the table above were shared online by many different people and include odd material: single words, stage whispers and sound effects. As a second opinion the app is also tested on plain read-aloud sentences whispered by 4 people it was not trained on, recorded in a studio and on laptop microphones. There it found 76% (71% to 81%, 220 of 288). Counted per person the range is 71% to 81%. With only 4 people, treat this as a second opinion, not a precise rate. On this version, clear whispered sentences were found more often than the short or unusual whispers in the table.
The faint list
Events a little below the review line go into a separate list labeled Faint. It finds more of the hidden voices and, by design, many more ordinary sounds: in recordings with no voice the faint list added 13.3 events per hour (8.1 to 20.6) on top of the main list. Listen to it when you want to be thorough, and expect most of it to be ordinary sound.
| Hidden voice | Main list | Main and faint lists |
|---|---|---|
| Real whispers as loud as the background | 46% (32% to 60%, 21 of 46) | 65% (51% to 77%, 30 of 46) |
| Real whispers quieter than the background | 50% (36% to 64%, 23 of 46) | 65% (51% to 77%, 30 of 46) |
| All real whispers | 56% (50% to 61%, 160 of 288) | 66% (61% to 72%, 191 of 288) |
| Whispered sentences (second set) | 76% (71% to 81%, 220 of 288) | 85% (80% to 89%, 245 of 288) |
| All hidden voices | 76% (73% to 79%, 658 of 864) | 82% (80% to 85%, 711 of 864) |
Compared with a general-purpose voice detector
For context, a widely used open voice detector (Silero VAD) was run on the same test recordings under the same rules, held to the same limit of six false events per hour. It is built to find clear speech in calls and meetings, which is a different job, so read this as context for how hard faint voices are.
| Hidden voice and detector | Share found |
|---|---|
| Spoken words | |
| EVP Labs | 86% (81% to 89%, 247 of 288) |
| General-purpose detector | 53% (47% to 58%, 152 of 288) |
| Real recorded whispers | |
| EVP Labs | 56% (50% to 61%, 160 of 288) |
| General-purpose detector | 4% (2% to 7%, 12 of 288) |
| Computer-made whispers | |
| EVP Labs | 87% (83% to 91%, 251 of 288) |
| General-purpose detector | 18% (14% to 23%, 53 of 288) |
Ordinary sounds listed as voice-like
| Ordinary sound | Share that drew a voice-like candidate |
|---|---|
| Birds, dogs, cats, insects | 1% (0% to 3%, 2 of 251) |
| Fans, engines, hum and buzz | 1% (0% to 4%, 3 of 228) |
| Rustling, scraping, zippers, paper | 1% (0% to 3%, 3 of 252) |
| Knocks, taps, footsteps, doors | 1% (0% to 4%, 3 of 244) |
| Squeaks, bells, chimes, ringing | 3% (1% to 6%, 6 of 211) |
A check on your own recording
The numbers above come from our test recordings. Your room, your phone and your background noise are different, so the app also tests itself on each recording you analyze.
After the analysis it takes quiet moments of your recording, mixes a known human voice into a copy of each one (a short spoken or whispered clip from people the model was not trained on) at three loudness levels, and runs the same analysis on those copies. It then shows how many of the added voices it listed, for example "Whisper, as loud as the background: 9 of 12 listed".
That tells you how faint a voice the analysis would have listed in that recording, and that a fainter one could have been missed. The copies exist in memory only. Your recording is not changed and the added voices do not appear in your results.
Known limits
- Faint and short sounds are unreliable. When a sound is quieter than the background or very short, both people and software make more mistakes. The app flags these.
- Real whispers are harder than spoken words. A whisper has no pitch and is close to shaped noise, so the app misses more of them and some wind, breath and rustle can resemble one.
- Voice-like does not mean a voice. Animals, squeaks, bells and some machines have voice-like features. The table above shows how often each is listed.
- People nearby are the usual source. A real voice-like sound in a recording most often comes from someone in or near the room, or from a device playing speech. Note who was present and what was switched on.
- The test is not the field. Results come from controlled mixes of real recordings. Your room, phone and distance will differ.
Why people hear voices in noise
Human hearing is tuned to find speech. Given unclear sound, people often hear words that are not in the signal, and they hear them more often when told beforehand that a voice may be present. Listeners also tend to disagree about what the words are. Researchers call this auditory pareidolia. It is ordinary perception and happens to careful listeners too.
Software does it too. We ran a large, widely used speech-to-text model (OpenAI Whisper, large-v3-turbo) on the voice-free test recordings, cut into 2160 pieces of 5 seconds. It wrote words in 25% of them, typically short phrases such as "Thank you". This is why EVP Labs does not write out words for a sound: a transcript would put words in your head before you listen.
A good habit: before you tell anyone what you think a sound says, ask them to listen and write down what they hear. If they hear something different, or nothing, treat the sound as unclear.
Sources
- Fant, G. (1960). Acoustic Theory of Speech Production. Mouton. The source and filter model of speech, including whispered speech.
- Nees, M. A., and Phillips, C. (2015). Auditory pareidolia: effects of contextual priming on perceptions of purportedly paranormal and ambiguous auditory stimuli. Applied Cognitive Psychology, 29, 129 to 134.
- Paranormal experiences, sensory-processing sensitivity, and the priming of pareidolia (2022). PLOS One. doi:10.1371/journal.pone.0274595.
- SWGDE (2018). Best Practices for Digital Audio Authentication, and SWGDE (2022). Best Practices for Forensic Audio. Keeping the original recording and avoiding absolute conclusions.
- IEC 60268-16 (2020). Speech Transmission Index. Measuring the rhythm of speech needs about 1.6 seconds of sound, one reason very short events are flagged.
- Radford, A., et al. (2023). Robust speech recognition via large-scale weak supervision. ICML. The Whisper speech-to-text model used for the reference above.
- Test audio: LibriSpeech, FSD50K, MUSAN and the OpenSLR room recordings. Authors are listed on the credits page.