There is no upload and no queue. Opening this page downloads no analyser at all. Once you choose a track it loads into the tab, and the audio is decoded by your own browser.
01Hand it a track
Drag in an MP3, WAV, M4A, OGG, WebM or FLAC, or click to browse. Ten seconds is the minimum and 50 MB the ceiling, and the file stays on your device.
02The spectrum is measured
The audio is downmixed to mono at 16 kHz and run through the same window, FFT and lower-envelope calculation the model was trained against.
03You get a number and the caveats
A 0 to 100 figure on five published bands with the decision line at 50, plus what was measured about the file that should temper it.
Tracks do not sort themselves into human and generated. Modern production already involves a great deal of software, so the question this model answers is narrower and more specific than it first sounds.
AI-generated
88typical score
Carries a vocoder fakeprint
Regularly spaced residue in the frequency spectrum, of the kind neural music generators leave behind. That residue is what the model was trained to find.
Inconclusive
48typical score
Between the two populations
A low-bitrate transcode, a heavy remaster or a pitch shift can move a track towards the middle. Near the line, the score is not a classification.
No AI signal
12typical score
Behaves like a recording
The spectrum shows no trace of the fakeprint. That says nothing about who wrote the song, only that this particular signature is absent.
In the wild
What people bring to a music check
Generated tracks arrive faster than any catalogue can review them. These are the cases people describe, and most of them have money or credit attached.
Catalogue floods
Hundreds of tracks uploaded in a day by an artist with no other trace.
Sync and library licensing
A track licensed as original work for an advert, a film or a game.
Demo submissions
A demo that sounds finished from an artist who cannot play it back live.
Artist impersonation
A new single in a known voice, released by somebody who does not own it.
Royalty claims
A claim on a recording nobody involved remembers making.
Competition entries
An entry to a songwriting prize with rules that assume a person wrote it.
Catalogue floods
Hundreds of tracks uploaded in a day by an artist with no other trace.
Sync and library licensing
A track licensed as original work for an advert, a film or a game.
Demo submissions
A demo that sounds finished from an artist who cannot play it back live.
Artist impersonation
A new single in a known voice, released by somebody who does not own it.
Royalty claims
A claim on a recording nobody involved remembers making.
Competition entries
An entry to a songwriting prize with rules that assume a person wrote it.
Playlist placements
Filler tracks placed to collect streams on playlists nobody curates closely.
Sample and stem packs
A pack sold as recorded performances, generated in an afternoon.
Broadcast beds
Background music cleared for broadcast on a licence that does not cover it.
Posthumous releases
A new recording by an artist who died before the model that made it existed.
Cover art and packaging
The artwork often fails a picture check before the audio fails an audio one.
Coursework submissions
A composition handed in for assessment as somebody own original work.
Playlist placements
Filler tracks placed to collect streams on playlists nobody curates closely.
Sample and stem packs
A pack sold as recorded performances, generated in an afternoon.
Broadcast beds
Background music cleared for broadcast on a licence that does not cover it.
Posthumous releases
A new recording by an artist who died before the model that made it existed.
Cover art and packaging
The artwork often fails a picture check before the audio fails an audio one.
Coursework submissions
A composition handed in for assessment as somebody own original work.
Reading the number
What the score says, and how firmly
The scale runs 0 to 100 and the decision line sits at 50, not at 65. This is a different model from the one behind the picture, video and voice detectors, trained on a different problem, and moving its threshold to match the others would misrepresent what it measures.
The five bands below are the model's own, published here so a result means the same thing every time you run one.
Decision line · 50
The five bands, what each one means for a track and what to do about it
Score
Band
What it means
What to do next
0 – 20
No strong AI-music signal
The measured spectrum does not show the vocoder fakeprint this model was trained to find.
Behaves like a recorded or produced track. Provenance still tells you more.
20 – 40
Probably human-made
A weak reading, of the kind ordinary mastering and encoding produce.
The track falls on the human-made side, but provenance still matters more than one score.
40 – 60
Inconclusive
The spectral evidence sits too close to the model's decision line.
Do not classify the track from this result. Ask for stems or a project file instead.
60 – 85
Likely AI-generated music
Over the line, in the region where generated tracks usually land.
The track carries a learned Suno or Udio style fakeprint. Confirm with provenance or source project files.
85 – 100
Strong AI-music signal
As firm as this measurement gets on a two-minute window.
The detected fakeprint is strong, but the score still does not prove authorship or identify a generator.
The bands and the line at 50 are fixed and published, so two tracks can be compared honestly.
None of it is proof of authorship. It is a measurement of a spectrum, and a spectrum does not know who wrote the song.
Under the hood
What the model is actually measuring
Neural music generators reconstruct audio through a vocoder, and vocoders leave regularly spaced peaks in the frequency spectrum. The model measures that residue directly. It does not guess from genre, from lyrics, from how polished a production sounds, or from whether a voice sounds processed, all of which would be prejudice rather than measurement.
The browser reproduces the reference pipeline exactly: downmix to mono, resample to 16 kHz, apply a periodic Hann window, take an 8192-point FFT, convert to decibels, compute the lower envelope, and hand all 3,585 features to the model. Preprocessing that differs from training is the usual way a correct model produces wrong numbers, so it is copied rather than approximated.
At least ten seconds of audio is required and up to the opening two minutes is read. Short, quiet, clipped and truncated inputs are reported beside the score instead of being treated as equally reliable.
Coverage is stated above. A generator newer than that may leave no residue that is recognised, and a low score on such a track is a gap in the coverage rather than evidence about the music.
What is measured, and what is not
Measured: the frequency spectrum of the opening two minutes, at 16 kHz mono
Measured: duration, level, clipping and truncation of the file you supplied
Not measured: lyrics, melody, genre, or how the track was produced
Not measured: authorship, ownership, or who performed on the recording
Not measured: which stem in a mix a signal came from
Accuracy
Where it holds up, and where it slips
Coverage is stated rather than implied: Suno up to v5 and Udio up to v1.5.
Outside that coverage the honest answer is that nobody knows. A generator released since may leave a different residue, or none that is recognised, and there is no independent evaluation that would let us publish a figure for it.
What pushes a result towards the middle
Low-bitrate transcodes, which reshape the high end where the residue lives
Remasters, heavy limiting and loudness normalisation
Pitch shifting and time stretching applied after generation
Tracks under ten seconds, or with a long quiet or ambient opening
Live recordings with heavy room noise
Generators released since the stated Suno v5 and Udio v1.5 coverage
A false positive is most often a heavily processed master rather than a generated track. Aggressive limiting and resampling can leave periodic structure that resembles the pattern being measured, which is why the readout reports what it measured about the file.
A false negative is likely for anything outside the training coverage. The absence of a recognised fakeprint is not evidence that a track was played by people, and this page would rather say that plainly than publish a reassuring number.
Read the result as evidence, never as a verdict. For anything with money or credit attached, a creation trail beats a spectrum every time.
What it cannot prove
Reliably catch generators newer than the stated Suno v5 and Udio v1.5 coverage
Prove authorship, ownership, performance or royalty entitlement
Separate a generated stem from a human stem inside one finished mix
Survive every remaster, pitch shift, EQ change or low-bitrate transcode
Say anything about lyrics, melody or whether a song is derivative
Settle a rights dispute, a distribution appeal or a contest entry on its own
Use provenance as the second check
Ask for stems, a project file, a rough take, a phone recording from the room, or anything at all from before the master. A spectral score is one signal; a consistent creation trail is far stronger, and it is the thing a generated track cannot produce afterwards.
What you get
Everything here, at no cost, with nothing held back
There is no paid tier of this AI music detector holding back the useful part. The score, the band, the measured caveats and the model's stated coverage are all in the free version, because there is no other version.
A measurement, not a guess
Vocoder residue in the spectrum, measured directly. Genre, lyrics and production polish are never consulted.
Training-matched preprocessing
The same 16 kHz downmix, Hann window, 8192-point FFT and envelope the reference pipeline uses, reproduced in the browser.
Its own scale, stated
Five bands with the decision line at 50, published rather than quietly reused from another detector.
Coverage named
The generators the model was trained against are stated, so you know when a low score means nothing.
Nothing uploaded
Opening the page downloads no model. The track is decoded and scored in this tab and never sent anywhere.
Saved on your device
Results are kept in your own browser storage so you can come back to them, and clearing your site data removes them.
Formats, limits and requirements
Formats read
MP3, WAV, M4A, AAC, OGG, WebM, FLAC
Track length
10 seconds minimum; the opening 2 minutes are analysed
File size
Up to 50 MB
Per check
The opening two minutes, as one spectrum
Analysis
Mono 16 kHz, Hann window, 8192-point FFT, 3,585 features
Scale
0–100 over five published bands, decision line at 50
Runs on
Your device, via WebAssembly. Nothing is transmitted
Privacy
Where your track goes: nowhere
This is the part where most AI music detection tools ask you to trust a policy. There is no policy to trust here, because there is no transfer. For unreleased music that is not a nicety: handing a demo to a website is publication of a sort, and it is the one thing an artist checking their own work cannot undo.
ChosenThe page reads the file straight off your device.
ScoredAudio is decoded and the model runs in this tab.
DroppedClose the tab and the track and the decoded audio are gone.
A measurement is the reliable path, but generated music still has habits, and they are easier to hear on a second listen than on a first.
Lyrics that scan perfectly and say nothing specific
A vocal that never strains, breathes oddly, or sits slightly outside the mix
Cymbals and hi-hats that smear rather than ring out
An arrangement that repeats without variation, verse to verse
Instruments that never quite articulate: a guitar with no pick noise, a piano with no pedal
A stereo image that stays fixed while the arrangement changes
Endings that fade rather than resolve, because the model had nowhere to go
Room sound that is identical on every element, or absent from all of them
A catalogue of tracks uploaded in one day by an artist with no other trace
Artwork that fails a picture check, which is often the faster tell
Every one of these appears in ordinary music, especially in demos and in some electronic genres. Use them to decide what to check, not to reach a conclusion.
Most free AI music detector sites follow one pattern: take the upload, hold the audio, return a percentage with no working, and name a generator they cannot actually identify.
Capability
Original or AI
Typical free checker
Runs on your device, nothing transmitted
Yes
NoYour track goes to a server
Says how much of the track was read
YesOpen and published
No
States which generators it covers
YesSuno v5, Udio v1.5
No
Published band thresholds
YesFive bands, line at 50
No
Reports duration, level and clipping
Yes
No
Reports duration, level and clipping
Yes
No
Unreleased music stays private
Yes
NoUploaded and stored
Works without an account
Yes
PartlyOften after a sign-up wall
Result kept only in your browser
Yes
No
Names the generator that made a track
NoNobody can do this reliably
PartlyFrequently claimed
Separates AI stems inside one mix
NoNot possible from a master
No
Proves authorship or ownership
NoNo detector can
PartlySometimes implied
“Typical free checker” describes the pattern shared by the free web tools we have used, not one named product. Where a competitor does better on a row, that row is wrong and we would like to be told.
Yes, and without an asterisk. There is no account, no card and no trial. It runs on your own processor, so there is no server cost to recover. The limits are technical: ten seconds minimum, 50 MB maximum, and the opening two minutes analysed.
Does the track get uploaded?
No. Opening this page downloads no model at all. Once you choose a file, the analyser loads into your tab, the audio is decoded by your browser, and the analysis happens there. For an unreleased demo that is the difference between a private check and a publication.
What audio formats can it read?
MP3, WAV, M4A, AAC, OGG, WebM and FLAC — in practice, anything your browser can decode. A bounce straight out of a session works, and so does a file saved from a streaming download. When decoding fails the tool says so rather than reporting a number.
Why does it need ten seconds?
Because the measurement is spectral, and a very short clip does not contain enough spectrum to measure. Ten seconds is the floor at which the feature set the model expects can be computed at all. Below that the tool declines rather than returning a figure it cannot stand behind.
Can I check a link to a streaming track?
No. Fetching audio from a streaming service would mean routing it through a server, which is exactly what this tool avoids, and would also mean handling files we have no right to. Download or export the track yourself and hand the file to the tool.
What the result means
What does the score actually measure?
It measures how strongly the opening two minutes of the track show the regularly spaced spectral residue that neural music generators leave, on a 0 to 100 scale with the decision line at 50. It is a statement about a spectrum, not about a songwriter.
Why is the decision line 50 rather than 65?
Because this is a different model solving a different problem, and its published threshold is 50. Moving it to match the picture and voice detectors would make the four look tidier and would misdescribe what this one measures, so the number stays where the model put it.
Can it tell me whether Suno or Udio made a track?
No, and the distinction matters. What is measured is a characteristic shared by neural vocoders in general, not a signature belonging to any one service, so the same residue appears whichever of them produced a track. Any tool naming a specific generator from a finished master is guessing, however confidently it presents the guess.
Does a low score prove a human made the music?
No. A low score means this particular fakeprint was not found, which is a narrower statement than it sounds. A generator outside the model's stated coverage, or a track transcoded at low bitrate, can read low while being entirely generated.
Why did my own recording get flagged?
Usually mastering. Heavy limiting, loudness normalisation and resampling can leave periodic structure in the high end that resembles what is being measured. Try an earlier bounce with less processing, and check the caveats shown beside the score for clipping or truncation.
Limits and next steps
Can it detect AI vocals over a real backing track?
No, not as a separate finding. The measurement is taken across the whole mix, so a generated element sitting inside a produced arrangement is diluted by everything around it. Separating stems from a finished master is a different problem this model does not attempt.
Will it work on a track made by a generator released this month?
No, not reliably, and that is worth knowing before you rely on it. The stated coverage is Suno up to v5 and Udio up to v1.5. Anything newer may leave residue the model has never seen, and a low score on such a track says more about the training set than about the music.
Is a result enough for a distribution appeal or a contest entry?
No. A spectral score is one input, it carries stated coverage limits, and no platform should decide a rights or eligibility question on it alone. What actually resolves these is provenance: stems, a project file, a rough take, anything from before the master.
Should I check the cover art as well?
Yes, and it is often the faster answer. Generated releases usually carry generated artwork, and the picture detector reads a still image far more confidently than any audio model reads a master. A flagged cover is a reason to look harder at the track.
Are results saved anywhere?
Yes, in your own browser and nowhere else. Each check is written to local storage on this device so you can reopen it from the results page. We never receive them, they do not sync between devices, and clearing your site data deletes them permanently.
Settle it before you sign anything
Drop in the track you were unsure about and read what the spectrum says. Free, nothing uploaded, and the coverage limits are stated beside the number.