Free tool
Tempo & timing grid checker
Estimate a track's tempo, then measure how closely its onsets sit on a sixteenth-note grid, how far that placement spreads, and whether the tempo drifts across the file. Decoding and measurement happen in your browser — the file never leaves your machine.
Short answer
Can timing tell you whether a song is AI-generated?
No. Grid-perfect timing means the rhythm came from a machine clock — a sequencer, a drum programmer, or a quantised take — and that has been normal practice for forty years. What timing genuinely settles is narrower and more useful: whether a file is plausible as a live performance, and whether it was edited or stitched.
Where detection misleads →Drop an audio file here
MP3, WAV, FLAC, M4A, Ogg or Opus. Long files are measured on four minutes from the middle.
What each number means
Tempo comes from autocorrelating an onset envelope — a curve of how much new spectral energy appears from moment to moment — and taking the strongest period between 60 and 200 BPM. It is a solid measurement with one known ambiguity: it cannot tell metrical levels apart, so half or double the tapped tempo is a normal result rather than a failure.
Off grid (median) is the typical distance from each detected onset to the nearest sixteenth-note line, after the tool searches for the grid phase that fits your track best. Values under about 3 ms mean placement is effectively exact. Values between 5 and 15 ms are what tight players and lightly edited takes produce. Larger values mean either genuinely loose timing or a grid model that does not match the feel — heavy swing, triplet subdivisions and drifting tempo all read as looseness here.
On grid is the share of onsets within 6 ms of a line, and the spread figure in the readings is the interquartile range of those distances. Read them together: a high share with a near-zero spread is uniform machine placement, while a high share with a wide spread usually means a programmed backbone with performed parts over it.
Why grid-tight timing is not an origin signal
Quantisation predates generative audio by decades. Every drum machine, every step sequencer, every producer who has ever dragged a hit to the nearest grid line produces onsets that sit exactly on a clock. Whole genres are defined by that precision. A track that measures 0.5 ms from the grid has told you it was sequenced, and nothing more.
The inverse claim fails too. Generated output is not always rigid: models trained on performed music reproduce performed timing, and plenty of generated material shows the same small deviations a session player would leave behind. So neither direction of this measurement supports a provenance conclusion on its own.
What timing does settle
The useful questions are narrower ones about specific claims. If a file is presented as a live take by a band, machine-exact onset placement across four minutes is inconsistent with that claim — humans do not hold a clock that well, and the absence of drift is as telling as the placement. If a file is presented as one continuous recording, a tempo that steps between halves points to an edit. Both of those are checkable statements about a stated provenance, which is a far stronger footing than a general verdict.
For the full argument — why quantisation is not an origin signal, how to read the numbers, and where timing readings mislead — see tempo, timing grids and AI music.
Use it alongside the other measurements here: structure and repetition for copied passages, file metadata for encoder history, a spectrogram for band structure, and stereo field for mix decisions.
Limits of this tool
Measurement is bounded to four minutes from the middle of a file. Onset detection needs transients, so ambient, orchestral and vocal-only material often yields too few to measure. The grid model is a straight sixteenth-note division, which means swung, triplet-based and shuffled feels will report deviations that are entirely intentional. Tempo drift is compared between two halves only, so a slow continuous slide may read as a small step. And the browser's own decoder reconstructs lossy files, so transient positions in a low-bitrate MP3 can sit a millisecond or two from where they were in the master.
Questions
- Does perfectly quantised timing mean a track is AI-generated?
- No. Programmed drums, quantised takes and anything rendered from a sequencer all land exactly on the grid, which covers most electronic music and most modern pop. Grid-tight timing tells you the rhythm was produced by a machine clock, which has been standard practice since the 1980s.
- What is the on-grid share?
- The proportion of detected onsets sitting within 6 ms of the nearest sixteenth-note grid line, after the tool searches for the grid phase that fits your track best. High share with a very small spread means uniform placement; high share with a wide spread usually means a quantised backbone with played parts over it.
- Why is the reported tempo half or double what I tapped?
- Tempo is estimated by autocorrelating an onset envelope, and that measure cannot distinguish metrical levels. A track at 140 BPM with strong eighth notes can peak at 70 or 280. Halve or double the figure to match what you feel; the grid measurements are unaffected because the grid spacing scales with it.
- What does tempo drift indicate?
- That the two halves of the excerpt fit different tempi. Real drift suggests a performance recorded without a click. It can also mean a deliberate tempo change, a stitched edit between two files, or the estimator locking onto a different metrical level in each half.
- Why does my ambient track report no timing at all?
- Because there are too few clear transients to detect. Sustained pads, drones, solo voice and very reverberant recordings produce a smooth onset envelope with no peaks to place on a grid, so the tool says so rather than inventing numbers.
- Is my audio uploaded anywhere?
- No. The file is decoded and measured by your own browser and released when you close the tab. Nothing is transmitted, stored or logged, so unreleased material is safe to check.