Runs in this tab · no daily limit
AI music detector — check a song without uploading it
Drop a track. The frequency fingerprint is computed on your device; the audio never leaves it.
Check a track
No song handy? Play one of ours:
No track yet
Your CPU · the audio stays here
- Model
- Decode
- Fingerprint
- Score
%
chance the track carries AI-generator artifacts
This is a probability from one small statistical model that looks for one kind of artifact. It can be wrong in both directions and is not evidence of how a song was made.
Checking a whole catalogue? Pro runs batches of up to 100 tracks
What the model looked at
How it decides
-
Average the spectrum
The song is mixed to mono, resampled to 16 kHz and cut into 8,192-sample windows. Their power spectra are averaged in decibels, so melody and lyrics wash out and only the fixed colouring of the recording chain is left.
-
Subtract the floor
A running minimum across 10 neighbouring bins traces the spectrum's lower envelope. Whatever rises above it, clipped at 5 dB, becomes the 3,585-number fingerprint for 1–8 kHz.
-
Weigh the peaks
The upsampling layers inside a generator's vocoder print peaks at regular frequency intervals. A logistic regression with one weight per bin scores how strongly the fingerprint matches that comb.
The idea comes from Afchar, Meseguer-Brocal, Akesbi and Hennequin, “A Fourier Explanation of AI-music Artifacts” (ISMIR 2025). The trained weights are lofcz/ai-music-detector. We checked this page against the model's own Python code on the same audio: the fingerprints correlate at 1.0000000 and the scores differ by less than one in a million.
What we got on 11 openly licensed songs
We ran the detector on songs from Wikimedia Commons whose licence lets us test and publish results, then re-encoded each one three ways to see how compression moves the score. Numbers are the chance of AI-generator artifacts, measured on 24 September 2026.
| Track | Made by | Original file | MP3 128k | MP3 64k | AAC 128k |
|---|---|---|---|---|---|
| Toreador Song — Bizet Musopen · CC0 | people | 0.2% | 0.2% | <0.1% | 0.2% |
| Morning Musopen · PD | people | <0.1% | <0.1% | <0.1% | <0.1% |
| Sax, Rock, and Roll Kevin MacLeod · CC BY-SA 3.0 | people | <0.1% | <0.1% | <0.1% | <0.1% |
| Fork and Spoon Kevin MacLeod · CC BY 3.0 | people | <0.1% | <0.1% | <0.1% | <0.1% |
| COW BOP Two Scuffed · PD | Suno | >99.9% | >99.9% | 54.1% | 99.9% |
| Inevitability V. Argonov · CC BY-SA 4.0 | Suno | >99.9% | >99.9% | >99.9% | >99.9% |
| Rise! Seba Tysair · CC BY 3.0 | Suno | >99.9% | >99.9% | >99.9% | >99.9% |
| My God Is Science Syntharia · CC BY 3.0 | Suno | >99.9% | >99.9% | >99.9% | >99.9% |
| The Rise of the ClueBots PD | aimusic.so (engine not stated) | >99.9% | >99.9% | >99.9% | >99.9% |
| WikiWarriors PD | aimusic.so (engine not stated) | >99.9% | >99.9% | >99.9% | >99.9% |
| Wikipedia song PD · 2026-03 | Tencent model | 0.2% | 0.2% | <0.1% | 0.1% |
Two things to take from this. First, a clean miss: the Tencent-made song from March 2026 scores like a human recording, because the model only knows the vocoders of Suno up to v5 and Udio up to 1.5. Second, the short Suno clip slid into ‘uncertain’ at 64 kbps; the longer tracks did not. Eleven songs are a sanity check, not a benchmark, and some human tracks here may overlap the Free Music Archive data the model was trained on.
Questions people ask
Is my song uploaded anywhere?
No. The page fetches one 14.8 KB model file, then decodes and analyses the audio inside this browser tab. Nothing that contains your audio or its fingerprint is sent anywhere, which you can confirm in the browser's network panel. Unreleased demos and licensed stems are safe to check.
How accurate is it, and which generators does it know?
Its author reports 99.88% accuracy and a 0.31% false-positive rate on 17,866 held-out clips of FMA music and Suno/Udio output (SONICS dataset). That is on the same kind of data it was trained on. Coverage is Suno up to v5 and Udio up to 1.5. In our own test of 11 songs, every Suno track scored above 99.9% and every human one below 0.3%, but a Tencent-made song from 2026 was missed completely. Suno v6 came out on 9 September 2026, after this model; we haven't found openly licensed v6 songs to test yet, so read scores on brand-new releases with that in mind.
Why does re-encoding or a low bitrate change the result?
The fingerprint is a set of small peaks a few decibels above the spectral floor between 1 and 8 kHz, and that is exactly the detail lossy encoders throw away first. At 128 kbps MP3 or AAC nothing in our test moved. At 64 kbps the 33-second Suno clip fell from above 99.9% to 54%. Use the best copy you have, and the page will warn you when a file has less than 64 kbps per channel.
It flagged my own song. Can the score be used as evidence?
No. It is one statistical model looking for one artifact, with a published false-positive rate around 1 in 300 even on data it knows well, and it has never seen your recording chain. If authorship matters, in a label dispute or a contest, your session files, stems and project history are the evidence. The CSV export only records what this tool said about a file and when.
Why does the first visit only download about 15 KB?
The detector itself is tiny: 3,585 weights and one bias, a 14,795-byte ONNX file. Almost all the work is signal processing done in JavaScript: resampling, about 700 FFTs of 8,192 points for a three-minute song, and the envelope. On our 2023 test laptop (Apple M2 Max) a three-minute MP3 goes from file to score in about 0.45 s, decoding included. Phones are slower, mostly at decoding, but the arithmetic is small enough that the page is built to run on them too.
Can I check a whole catalogue at once?
The free page checks one track at a time, with no daily cap: run as many as you like, one after another, and save a CSV line for each. Batch checking of up to 100 tracks with a single combined report is planned for the Pro tier.
Why only the first five minutes, and why does it want 30 seconds?
The model's reference code stops at 300 seconds, and we match it exactly so that a score here equals the score from the Python original. Below about 30 seconds the averaged spectrum still has a lot of random ripple, and the only wobbly result in our test was the shortest clip.
Model, licence and examples
Detector weights: lofcz/ai-music-detector by Matěj Štágl, MIT licence (full text in LICENSE-model.txt). Examples, both from Wikimedia Commons: “Toreador Song” from Bizet's Carmen, recorded by Musopen, CC0 (3-minute excerpt, re-encoded to MP3); “COW BOP”, made with Suno and prompted by Two Scuffed, public domain (re-encoded to MP3).