Media
3 items
Description
Detects Suno-generated music while it plays, from the fingerprint its vocoder leaves between 1 and 8 kHz. Suno Sniffer tells you when the music you are listening to was generated by Suno. Let a track play for about a minute. A small panel in the corner reports what it smells: "mmmh, pure spirit" for a recording with no trace in it, or "ugh, reeks of robot" when it finds one. Everything is measured locally. No account, no upload, no server, and no network request of any kind. HOW IT WORKS Music generators build audio with a neural vocoder. The last stage of that vocoder upsamples the signal, and the operation leaves a mathematically predictable trace: the spectrum is copied at regular intervals, producing a comb of peaks between roughly 1 and 8 kHz. The extension averages the spectrum of what you are hearing over a long window, subtracts the lower envelope so the melody drops away and only the peaks remain, then feeds the result to a small classifier. Real instruments and real voices do not produce that comb: their peaks move with the notes and average away over time. That band is also the reason detection survives streaming. A widely repeated claim says generated music can be spotted because it cuts off above 15 kHz, but that cutoff is the signature of lossy encoding, not of generation, and human recordings at the same bitrate are cut in the same place. The 1 to 8 kHz band is left alone by compression. WHAT IT CAN AND CANNOT TELL YOU It answers one question: does this recording carry the Suno trace? It is not a general detector for machine-made music. Anything built on a different architecture is invisible to it, so a clean reading means "no Suno trace here", never "this is not AI". The panel repeats that limit under every clean result. The measurement covers the whole mix, so a human singer performing over a generated backing track is flagged as well. That is technically correct, but please do not read it as proof that a performer is fake. It cannot separate sections either: a reliable answer needs at least a minute of averaging, so it cannot say which verse is which. A positive is an informed indication worth following up. It is not evidence. HOW IT BEHAVES It needs about a minute of audio and says so instead of guessing before then; the verdict settles around two minutes. Mistakes are pushed to the safe side: a short listening window produces a missed detection rather than a false accusation, and a recording that sits outside the usual range without reaching the threshold is reported as inconclusive rather than clean. Heavy saturation, pitch or time manipulation and resampling will disturb the trace. Ordinary mixing, mastering and streaming compression will not. It reads the audio playing on the pages it is granted access to, and local audio files you open in the browser. Streams protected by DRM cannot be read by the Web Audio API and are out of reach. Click the toolbar icon to switch it off entirely. The panel has its own close button to hide it on a single page. CREDITS The detection model is ai-music-detector by Matej Stagl, used under the MIT licence: https://github.com/lofcz/ai-music-detector