AI Music Detector — Was This Song Made by AI?
An online tool that checks whether a song was made by an AI such as Suno or Udio. It reads generator tags in the file and measures the sound, and the verdict comes from the trained SONICS classifier, which listens to the whole track. Runs in your browser, no installation needed.
How to check if a song is AI
Upload a song — MP3, WAV, FLAC, M4A, OGG or a video
The tool reads generator tags, measures the sound's bandwidth and starts the classifier on its own — the model (73 MB) downloads once
Wait for the verdict: usually 10–20 seconds on a computer, a couple of minutes on a phone
Read the verdict and the explanation under it, and download the PDF report if you need it
Check a track with a trained classifier and see what the result means
What the detector does
Why you can trust this detector
The verdict comes from a trained classifier, not a set of heuristics; the AI threshold is set so that in our tests no human recording was accused; the tool says “inconclusive” when a file can't be checked and explains every number.
Frequently asked questions
What do the verdict and the percentage mean?
The percentage is the classifier's score: how much the sound resembles the AI songs it was trained on. “AI-generated” starts at 99 %, “Possibly AI” at 70 %, below 30 % it says “Likely human-made”, and in between “Inconclusive”. Under the verdict the page always says what it means for your file.
Why “Inconclusive” when it is an ordinary song?
There are two reasons. The score sits in the middle of the scale, so the classifier leans neither way. Or the recording has no sound above 7.25 kHz — that is how calls, voice messages and phone-speaker recordings sound; the classifier errs on such audio, so we give no verdict. Check the song's original file.
Which AI generators does it recognize?
Suno and Udio songs best: 91 % of the Suno 5.5 and 6 songs the model had never seen scored 70 % or more. Songs from MiniMax, Mureka, HeartMuLa, ACE-Step and other 2025–2026 generators: 71 %. It recognizes instrumental music from generators like MusicGen, Stable Audio and Mubert least reliably (34 %), and the same goes for generators missing from training. It does not name the generator from the sound — only from a tag in the file.
Can I check a voice message, a call or a phone video?
A video, yes, if its sound is complete: the tool extracts the audio track. Calls and voice messages are usually cut off above 3–4 kHz, and on them the classifier can't tell AI from human — the tool will say so.
Does it detect a cloned voice or AI vocals over live music?
Not reliably. The classifier judges the whole song and was trained on songs made entirely by AI. If you suspect a cloned voice in a call from a “relative”, a family code word helps — the Caller check tool is built for that.
Why are there no separate percentages for the vocal and the music?
The page used to show separate scores for the vocal and the instrumental, and we removed them: they did not tell AI from a human recording. We also tested the honest version — splitting a song into vocal and music and scoring each part with the same classifier — on 548 songs. A part's score is wrong noticeably more often than the whole song's: for almost one in four human recordings the vocal scored 70 % or more. And on test mixes where only the vocal or only the music was AI, the breakdown pointed at the right part in about one case out of four. So the verdict is given for the song as a whole, and the detail it offers is the score by time section for tracks longer than two minutes.
There is no generator tag — does that mean a person made the song?
No. Almost any re-encode or upload to a platform strips the tag, so its absence proves nothing. And if there is a tag, it describes the file, not the music: it can be written by hand.
We can! Just send us a quick message with your idea. If you'd like to discuss it in detail, leave your email and we'll get back to you. You can stay anonymous.
How do you rate this tool?
Comments
I also now have fixes for typo.
