Likely AI-associated
Several signals lean synthetic
Multiple measurements are consistent with generated audio. Strong enough to look closer, never strong enough to accuse.
Free tool
AI Music Detector
Analyze a song in your browser to estimate whether its acoustic characteristics are consistent with AI-generated audio. A probabilistic signal analysis, not proof of authorship — and this page is honest about the difference.
Analyze a track
Drag and drop audio here, or choose a file. MP3, WAV, M4A, AAC, FLAC, OGG or WebM up to 25.0 MB. At least 10 seconds; 30 seconds or more is more reliable. No account needed.
Audio is decoded and analyzed in this browser. Nothing is uploaded, nothing is stored, and detector files never enter ANYANO's Sound training system.
Method
How it works
Decode in your browserSample sections across the trackMeasure spectrum and dynamicsCombine evidence conservatively
- 01
Your file is decoded locally
Decoding and every measurement happen in this browser tab, in a Web Worker so the page stays responsive. The file is never uploaded and nothing is stored.
Sample rate, channel count and duration are read from the decoded audio, not from the file name.
- 02
Several regions across the track are measured
Up to five six-second regions spread across the recording are analyzed with a Hann-windowed FFT over many frames each, so a conclusion never rests on the intro alone.
- 03
Nine signals, none of them decisive
Spectral ceiling and its sharpness, cross-section spectral agreement, crest factor and dynamic movement, spectral-centroid variability, high-band energy share, stereo correlation and width, flatness, level movement and transient density.
- 04
You get an estimate with its confidence
Likely human-made, mixed, likely AI-associated or inconclusive — with a probability, a separate confidence level, and every signal shown with its measurement and interpretation.
Honest answer
Can AI music really be detected?
Partly. And the way it works is not what most people assume.
Detectors do not identify AI music by judging whether it sounds creative, generic, emotional or repetitive. Those judgements are unreliable in humans and worse in machines. What carries some signal is the production side: bandwidth limits, how uniform the tonal balance stays across a whole track, how little brightness and level move over time, and how static the stereo image is. ANYANO measures those properties directly, on your device, and tells you which of them leaned which way.
That has two consequences worth understanding before you trust any tool of this kind. First, detection is a statement about a production pipeline, not about authorship — a human who ran their own performance through a codec, a limiter and a low-pass filter can look generated. Second, none of these measurements is unique to generated audio, which is why this tool is deliberately conservative: it is a multi-feature acoustic detector, not a neural classifier, and until a properly labelled benchmark exists here no accuracy figure is published at all.
Anyone quoting a single accuracy figure without saying which generators it was measured on, and how it did on generators it had never seen, is quoting a marketing number.
Reading it
What the result means
Likely human-made
No synthetic signature found
The measurements look like conventionally produced music. This does not certify human authorship — a clean generated track can measure the same way.
Mixed or inconclusive
Signals disagree
Common for short clips, heavy compression, lo-fi recordings and AI-assisted work. Reported honestly instead of rounded into a verdict.
Limits
Why it is not 100% accurate
Every limitation below is a known, measured weakness of this class of detector, not a disclaimer added out of caution.
- No labelled benchmark yet. The weights and thresholds are reasoned from DSP principles, not calibrated on a labelled dataset, which is why the probability range is deliberately narrow.
- Re-encoding. Streaming, messaging apps and low-bitrate MP3s all erase part of the evidence.
- Simple edits. Pitch shifting, time stretching and resampling measurably reduce detection performance.
- Human electronic music. Clean digital synthesis and heavy processing are the most common cause of false positives.
- Hybrid production. Real workflows mix human performance with AI mastering, separation and voice tools.
- Short clips. Less audio means fewer sections and less evidence, so files under ten seconds are refused rather than guessed at.
Definitions
Human, AI-assisted, or AI-generated
Human-made
Performed and produced by people
Conventional production: instruments, voices, samples, synths and a person making the decisions, with no generative model writing the audio.
AI-assisted
Human work, AI tools in the chain
Stem separation, AI mastering, voice processing, generated one-shots. Authorship is human; some of the audio has passed through a model.
AI-generated
Audio written by a model
A generative model produced the music itself, usually from a text prompt, with the human role limited to prompting and selection.
Trained on your own music
A different thing entirely
A private model trained only on material you own or have permission to use. Generative, but grounded in your own catalog rather than someone else's.
Privacy
Does ANYANO store my music?
No. Your audio stays on your device. The file is decoded and analyzed entirely in this browser tab, in a Web Worker, and released when you reset. Nothing is uploaded, nothing is stored, no account is required and your filename is never sent anywhere.
That wording is deliberately precise, and it is checkable: open your browser's network panel and analyze a track. No request carries the audio, because the whole analysis runs in JavaScript on your machine. Usage counts are recorded without filenames, audio, or extracted measurements.
FAQ
Frequently asked questions
Train your sound. Generate your song.
If you would rather build with AI than argue about it: prepare your own music, train a Sound that stays private to your account, and generate from your own material.
Related