How the checker identifies music
CopyrightChecker doesn't run its own song database. It sends a 12-second clip of your audio to a music recognition service, which fingerprints the audio and compares it with a catalogue of known recordings. Our main provider is ACRCloud, which says its fingerprint database covers more than 150 million tracks and is constantly being updated. When ACRCloud can't answer, we use AudD, which says its database contains 160 million songs. The copyright detector page explains fingerprinting in plain words.
So the accuracy of a check comes down to two things: whether the recording is in the provider's catalogue, and whether enough of it survives in your clip to match.
How we tested
We wanted to know how recognition holds up in the situation creators actually have: a song playing under someone talking. So we built test clips that way.
- 16 songs. Popular tracks in several languages, taken from the 30-second previews iTunes provides.
- A voiceover on top. Each song was mixed under a synthetic voiceover, with the music at −14 dB, the way background music sits under someone talking.
- 12-second clips. Each mix was cut to 12 seconds, the same length as a free check.
- 4 control clips with no released music: voice only, synthesized chords, a sine-tone melody, and street noise. A good checker should return nothing for all four. Any match here would be a false match.
We sent every clip to ACRCloud and AudD separately and recorded what came back.
| Provider | Songs identified | False matches on controls |
|---|---|---|
| ACRCloud (main) | 15 of 16 | 0 of 4 |
| AudD (fallback) | 10 of 13 tested | 0 |
ACRCloud's one miss was "bad guy" by Billie Eilish, which it matched to a different recording instead of the original. That's the kind of error to watch for: a confident answer that names the wrong version. It's why every report links to the recording it matched, so you can listen and compare.
AudD was tested on 13 songs, not 16, because our account hit AudD's free-tier limit before the last three. We report the 13 rather than guess at the rest.
Full scans
We also ran two synthetic test files end to end through the full scan, which splits the whole soundtrack into 12-second segments and reports where each song plays. The first file had 3 songs and the second had 6. Every song was found and placed in the correct 12-second segment: 3 of 3, and 6 of 6.
What these numbers do and don't show
They show that recognition works well on clean commercial recordings under a voice at background level, and that it didn't invent matches for audio with no released music in it.
They don't show how it does on real creator videos. Our clips were built in a controlled way: one voice, one music level, high-quality previews of well-known songs. Real videos have room echo, music played through a phone speaker, crowd noise, songs that fade in and out, and tracks that are far less famous. Sixteen songs is also a small sample, so one miss moves the rate a lot. Don't read 15 of 16 as "94% accurate on your video".
A benchmark on real creator videos is planned. We'll publish its method and results on this page when it's done, including the misses.
What lowers accuracy
These are the situations where recognition is more likely to miss a song, or to match the wrong version of it:
- Heavy voiceover or noise. The louder everything else is compared with the music, the less of the music there is to fingerprint.
- Very short music. A stinger or a few seconds of a song gives the provider less to compare than a 12-second stretch.
- Covers and remixes. A cover is a different recording, so it won't match the original. Remixes may or may not match, depending on how much of the original recording is left.
- Live versions. A live performance only matches if that live recording is itself in the catalogue.
- Pitch and speed changes. Sped-up, slowed or pitch-shifted edits change the timing and frequencies a fingerprint is built from. Small changes often still match; large ones can stop a match.
- Unreleased or very new music. If a recording isn't in the catalogue yet, there's nothing to match.
If your clip has one of these problems, check a different moment of the file where the music is clearer, or run a full scan.
What the match score means
When the provider returns one, the report shows a match score from 0 to 100. It's the provider's measure of how closely your clip's fingerprint lines up with the recording it matched. A high score means a close, consistent match. A lower score means fewer points lined up, which happens with heavy voiceover or edited audio, and is a good reason to listen to the linked recording yourself.
The match score isn't a probability that YouTube or TikTok will act on your video. Platforms use their own systems and their own reference files. Content ID, for example, matches against files that rights holders give YouTube. Our report tells you what the recording is. The risk level describes what platforms commonly do with that kind of recording.
Why "No match" isn't permission
No match means the clip didn't match any recording in the catalogue. It doesn't mean the music is free to use. Everything in the list above can produce a no match on music that is fully copyrighted, and a platform's own system may know recordings our providers don't. Treat it as "we don't recognize this recording", then check where the music came from. How to tell if a song is copyrighted walks through that. For platform-specific rules, see the YouTube copyright checker and TikTok copyright checker pages.