What our video analyses show

Every recognized track carries a confidence score returned by the engine (ACRCloud). Out of the 988 tracks found, 57% scored 70% or more (or came from the official TikTok sound), 9% scored between 50 and 69%, and 34% scored below 50%.

Below 50%, the site shows “double-check”: the engine found a resemblance, not a certainty. That happens a lot when the music plays under a voice, is sped up or remixed, or is a cover. Public pages, song pages and the chart only keep tracks at 70% or more.

28% of the videos listened to returned no track. The usual reasons: an original sound that was never released (voice, sound effects, music made by the creator), music buried under speech, a live version or a cover missing from the catalog, or a clip that was transformed too much.

The description sometimes helps: in 1 videos it named songs found in the catalogs; in 0 of them, none of those songs was actually heard. The site shows them separately, never as heard.

On top of that, 5 links could not be opened (private, deleted or restricted videos) and 230 videos were longer than the analyzed length. These failures are not counted in the percentages above.

When a TikTok uses a sound from the official catalog, its title and artist are read directly, without listening: that was the case for 8% of the 410 TikToks analyzed. Most of the others show “original sound”: the song was re-uploaded under the creator’s name, and only listening can find it.

One sound travels through thousands of videos. When someone pastes a video whose sound was already listened to here, the result is reused instantly, with no new analysis: 1 videos were solved that way.

50 videos longer than 5 minutes were analyzed, with a median length of 9:09. They returned 217 tracks in total, 4.3 on average, and up to 25 in a single video.

Past 5 minutes, the video is listened to in 10-second samples every 30 seconds: a song that lasts less than about fifteen seconds can slip between two samples, and the result page says so.

The numbers are computed on every page refresh from the analyses stored on the site, the 3,000 most recent at most. A video counts once, even if several people analyzed it.

Excluded: files uploaded by hand (nothing guarantees they match a public video), analyses stopped by the daily budget or the free quota (they say nothing about the video), and analyses still running.

These are the videos visitors wanted to identify: mostly sounds whose title isn’t shown. They are not a representative sample of all published videos. No personal data is used: no link, no video title, only counts.