How Computers Recognise a Song From Its Sound Alone

· 7 min read

You have a file called "track07.mp3" with nothing else to go on. No title, no artist. Yet within seconds an application can tell you it is a specific song, from a specific album, from a specific year. How?

Not from the filename

First, a correction: this kind of recognition does not read the filename or the metadata at all. A file called "aaa.flac" with empty tags is identified just as reliably.

What gets read is the sound itself.

Acoustic fingerprints

The underlying idea resembles a human fingerprint. You do not need to store someone's entire body to recognise them again — a small, highly distinctive, unchanging pattern is enough.

For music, the process runs roughly like this: software analyses the track and records how energy shifts across frequency ranges over time. At which second the high frequencies swell, when the bass enters, how that pattern repeats.

The result is not a fragment of the song but a sequence of numbers. The original audio cannot be reconstructed from it — much as a fingerprint cannot be used to rebuild a face.

This is why the process raises neither privacy nor copyright problems: what travels to the server is a mathematical pattern, not the music.

Why fingerprints still match across different files

The impressive part: fingerprints of the same song still match even when the files differ in format, bitrate, or slightly in volume.

That is because what gets recorded is the relative pattern, not absolute values. Raise the whole track's volume by 10% and the relationships between its parts stay identical — and those relationships are what is stored.

So a 128 kbps MP3 and a FLAC of the same song produce matching fingerprints, despite being wildly different as data.

AcoustID: the fingerprint library

AcoustID is an open database holding millions of fingerprints along with the identity of each recording. When an application submits a fingerprint, AcoustID finds the closest match and returns the recording's identifier.

The database grows through user contributions — every time someone identifies a track and confirms the result was correct, that data strengthens the match for the next person.

MusicBrainz: the encyclopaedia

AcoustID answers only "which recording is this". Knowing the title, artist, album, year, track number, and cover art requires a separate database — and that is MusicBrainz's role.

MusicBrainz is an open, volunteer-edited music encyclopaedia, similar in spirit to Wikipedia but dedicated to release data. It holds remarkably fine detail: the difference between Japanese and European editions, who played which instrument, release dates per country.

The two work as a pair: AcoustID recognises the recording, MusicBrainz explains it.

Why results are sometimes wrong

The limitations deserve an honest account.

  • One recording can belong to many releases. The same album is reissued as a remaster, a deluxe edition, or a country-specific version with a different tracklist. The fingerprint is identical; the metadata is not.
  • Rare recordings may not be registered. Local music, small independent releases, and older recordings are often absent from the database.
  • Live and studio versions can be swapped when the arrangements are very similar.
  • Fingerprints need enough duration. Very short excerpts are hard to identify with confidence.

This is why a well-built application always shows the confidence of a match and asks for your approval before writing anything to a file.

How EarTea uses it

EarTea computes fingerprints directly in your browser, sends them to AcoustID, then fetches release details from MusicBrainz.

The music files themselves never leave your computer — what crosses the network is the numeric fingerprint and the search results.

Suggested metadata is shown first so you can inspect it. If the match is wrong — say you own the remaster but the original was matched — you can reject it and choose for yourself.

Try it on the messiest files you own: untitled files are exactly where this kind of recognition proves itself.

Open EarTea

Why this deserves support

Both AcoustID and MusicBrainz are run as open, non-profit projects, and their data is free for anyone to use.

Most applications that can "magically recognise a song" stand on two decades of unpaid volunteer work. If you correct a single wrong entry in MusicBrainz, you have improved it for everyone.