AI music detector

A new kind of uploaded music

Text-to-music models such as Suno, Udio and Google’s Lyria generate songs complete with lyrics, singing and production from a short description. The results range from novelty jingles to tracks that pass for professional releases. Streaming services now see a steady stream of AI-generated uploads, some made to collect royalties at scale, some attributed to fictional artists, and some imitating the style or voice of real ones.

Listeners, playlist curators, labels and music supervisors increasingly want to know what they are hearing. Disclosure rules are still being written, so detection fills the gap.

What GPTTrace measures in a song

Tags

Songs downloaded directly from generator websites sometimes carry comments or URLs naming the service. Distribution services usually rewrite tags, so their absence is normal.

Bandwidth and the upper band

Music generators work with compressed internal representations of audio. Some produce a softer, smeared top octave or a ceiling well below what the file format would allow, and some add a characteristic high-frequency shimmer. GPTTrace measures where the spectrum ends, how abruptly, and whether narrow, constant tones sit in the upper band. The spectrogram view shows the same thing visually — look at the top of the picture.

What it skips

The pitch-jitter, pause and noise-floor checks are designed for a single speaking voice and would mislead on music, where steady pitches and continuous sound are normal. GPTTrace detects that the audio isn’t speech and leaves them out.

Music is a harder target than speech for signal-based detection, because mastering, compression and effects shape real songs in similar ways. Expect “Inconclusive” more often, and combine the result with the artist-level checks above.

Checking a track you found on a streaming service

Streaming services don’t let you download the original file, so use the best copy you can legally obtain — a purchased download, the artist’s own upload, or the audio from an official video. Run it here and look at the spectrogram: generated tracks often show a uniform, slightly blurred texture at the top of the frequency range and a hard ceiling where the energy stops. Then check the artist page. Fully AI catalogues tend to share tell-tale traits: many releases in a short time across unrelated genres, generic AI cover art, no songwriting credits beyond one name, no tour dates, no interviews and no social media history before the first release.

Listening still helps. Sustained vocal notes with a faint metallic warble, consonants that blur into the music, lyrics that rhyme neatly but never say anything specific, and crash cymbals that hiss rather than ring are typical of current generators. None is conclusive on its own; three or four together, plus a clean result for the artist checks above, is a reasonable basis for doubt.