AI music detector: is this song AI-generated?

Why demo inboxes need an AI check

Generating a finished-sounding song now takes one prompt and a few seconds. For an A&R, a publisher or a playlist curator, that changes the job: a demo can sound polished and still have no artist, no session files and no rights you can sign behind it.

Listening alone is no longer a reliable filter, and checking every submission by hand does not scale. An AI check at intake lets you spend your listening time on the tracks that have a real project behind them.

How the detector works

Every analysis runs two independent checks:

1. File metadata. Generators often leave a signature in the file tags. Tag-per-Track reads them and recognises the marks left by Suno, Udio, Soundraw, Mureka, Boomy, Riffusion, Stable Audio, MusicGen and AIVA. A signature found here is treated as a confirmed result.
2. The audio itself. A 12-second excerpt, taken from the body of the track (around 0:45, where vocals and energy are stable), is examined by a specialised detection service for the artefacts generative models leave in the signal. If that excerpt is inconclusive, a second one is taken from the start of the track.

The result is a verdict, a confidence level from 0 to 100 and, when the signature is recognisable, the name of the most likely generator.

How to read the verdict

A detector that only answers "AI" or "human" hides how sure it is. Tag-per-Track shows the confidence and sorts it into levels:

The same care applies to human verdicts: "Verified human" is only shown from 80% confidence, "Likely human" from 50%, and below that the report says "Human" with a note that the detector is not very assertive: likely, but not certified.

What a detector cannot tell you

No AI music detector is proof, and this one is no exception. A few things worth knowing before you rely on a result:

• It is a probability, not a ruling. Treat "Possible AI" as a reason to ask questions, not as an accusation.
• Electronic instrumentals are harder. Heavily synthesised or sample-based beats often land in "Uncertain origin", which is why that range never lowers a score.
• It checks an excerpt. A track that mixes generated and recorded parts can be judged on the wrong 12 seconds.
• Generators evolve. A new model version can go undetected for a while.

When the stakes are real (a signing, a sync, a release), back the verdict with what only a human author can provide: the session or project file, separate stems, earlier versions of the song.

More than a yes or no: the full A&R report

The AI check is one part of every Tag-per-Track analysis. The same upload also gives you:

• An A&R score from 0 to 100, combining the acoustic profile of the track with the artist's audience.
• BPM, key, mood, genre and instrumentation, extracted from the audio.
• The artist's live Spotify traction: monthly listeners, followers, popularity.
• Optional lyrics transcription.

Screening a whole folder of demos

One track at a time is fine for a doubt. For a submission inbox, use batch mode in the studio, or connect Tag-per-Track to Claude Desktop or Cursor with the MCP server: one sentence, and your assistant analyses the whole folder and returns the demos ranked by priority, with AI suspicions listed apart.

Developers can call the same analysis through the REST API or the tag-per-track-mcp package.

Frequently asked questions

Which AI music generators can it detect?

The acoustic check looks for the artefacts of generative music models in general, with Suno and Udio as the most common cases. The metadata check recognises signatures left in file tags by Suno, Udio, Soundraw, Mureka, Boomy, Riffusion, Stable Audio, MusicGen and AIVA. When the signature is recognisable, the report names the most likely generator.

Is the AI music detector free?

Your first analysis is free, with no card and no account, and it includes the AI check. After that, an analysis costs one credit: Studio Packs are €5 for 33 credits or €15 for 100 credits, and credits never expire. Developers and AI agents can also pay per use in USDC.

How accurate is it?

It returns a confidence level rather than a blanket claim, and no detector is right every time. Results are most reliable on full songs with vocals and least reliable on electronic instrumentals. A verdict below 70% should be treated as a lead to verify, not as a conclusion.

What should I do if my own track is flagged as AI?

False positives happen, especially on synthesised or heavily processed productions. Keep your project file, your stems and your earlier bounces: they are the strongest evidence of authorship, and they are what a label will ask for.

What happens to the audio I upload?

The file is stored temporarily for the analysis and deleted as soon as the report is generated. The acoustic AI check is run on a 12-second excerpt by a specialised external detection service; if you turn on lyrics transcription, that step sends the full track to an external transcription service. Tag-per-Track does not keep your audio and never uses it to train AI models.

Can it detect AI vocals or voice clones?

The detector assesses the track as a whole, not the identity of a voice. It can report voice-conversion tools when their signature is recognisable, but it cannot tell you whose voice was cloned.