Was this track made by a person or by Suno?
Upload an audio file and find out the probability it was generated by Suno, Udio, MusicGen or Stable Audio. Our AI music forensics model reads the spectral fingerprints AI generators leave behind — and we show our work.
Not a vibe check. A model.
Our AI music forensics model is purpose-built to catch the fingerprints AI generators leave in audio. We surface the raw probability — no rounding, no hand-wave.
Our model. Kept current.
Fine-tuned on the latest Suno v3/v4 and Udio v1.5 outputs. Updated continuously as new generators ship and accumulate enough samples in the wild.
~5 second verdict
Upload finishes, inference runs, you get a probability. We don't ask you to wait in a queue or watch a fake progress bar.
Probability, not a label
We show 0–100% with a verdict zone (Likely Human / Inconclusive / Likely AI). When the model is unsure, we say so — instead of pretending to be certain.
Your file stays private
Audio is processed in memory and dropped immediately. We log the verdict and a short embedding hash for our drift monitoring — not the audio.
Free, no sign-up
Three free uploads to try. Need batch detection or API access? We've talked to labels and DSPs about this — get in touch.
Open methodology
Every claim about generator coverage and accuracy on this page is sourced. We publish our internal benchmarks in the blog when they change meaningfully.
How the detection works.
Drop an audio file.
MP3, WAV or FLAC up to 30 MB. We slice it into 5-second windows; you don't need to trim a clip.
Our model scores every window.
The detector outputs a per-window probability that the segment came from a known AI generator (Suno, Udio, MusicGen, Stable Audio). We aggregate to a track-level score, weighted by audio energy.
Verdict + closest match + signals.
You get a probability (0–100%), a verdict zone, the closest known generator, and the top three signals that drove the call — vocal artifacts, dynamic flatness, spectral fingerprints, etc.
Generator coverage as of May 2026.
We benchmark monthly. Open-set generalization to brand-new models is the hard part — we retrain and ship model updates as soon as new generators have enough samples in the wild.
| Generator | Released | Vocals? | Our detection | Notes |
|---|---|---|---|---|
| Suno v3 / v4 | 2024–25 | Yes | 98% recall | Most common in the wild. Strong fingerprint in vocal sibilance. |
| Udio v1 / v1.5 | 2024–25 | Yes | 96% recall | Cleaner mixes than Suno; we catch it on dynamic flatness + stereo image. |
| MusicGen (Meta) | 2023 | No | 91% recall | Instrumental only. Older — harder when laid over a real vocal stem. |
| Stable Audio 2 | 2024 | Limited | 89% recall | Strong on long-form instrumental; ambient / drone is the weak spot. |
| Riffusion / unknowns | misc | Mixed | ~70% | Open-set: anything not in our fine-tune dataset. Inconclusive is the honest answer here. |
| Human studio recording | — | — | 3% false-positive | Hyper-compressed pop and AutoTune-heavy vocals are the trickiest false positives. |
AI forensics, not guesswork.
Our AI music detector frames detection as audio forensics: it learns generator-specific spectral fingerprints and dynamics signatures across a benchmark of millions of AI-generated samples from Suno, Udio and MusicGen. Fine-tuned continuously on our own collection so we catch the latest model versions. When a track is from a generator we haven't seen, the detector says “Inconclusive” — and we agree with it.