What is audio fingerprinting?

How it works

A fingerprinting system first turns audio into a spectrogram, a picture of which frequencies are loud at each moment. It then keeps only the strongest points, the peaks that stand out above their surroundings. Those peaks survive background noise, low bitrates and cheap speakers far better than the raw waveform does.

Pairs of nearby peaks are combined into small hashes (roughly: this frequency, then that frequency, this far apart in time). A song produces thousands of them. To identify a clip, the system hashes it the same way and looks the hashes up in a database of known recordings. The right match is the one where many hashes agree and line up at a consistent time offset, which is also how the system knows how far into the track the clip starts.

An example

A video editor has 300 vlog clips and needs to know which ones have a particular licensed track playing in the background before publishing. Fingerprinting the track and scanning the clips flags the ones where it appears and at what time, even under talking and street noise. A transcript would not help here, and neither would comparing file bytes.

Broadcasters and platforms use the same technique at a much larger scale, to check uploads against reference recordings from rights holders and to log which tracks aired and when.

What it is not, and where it fails

Fingerprinting is often confused with two other things. It is not transcription: it does not know what is being sung, only what the recording sounds like. It is also not a file hash: two files with different formats or bitrates have different bytes but the same fingerprint, which is exactly the point.

• Covers and live versions usually don't match, because they are different recordings. Recognizing the underlying song is a separate problem.
• Heavy pitch or tempo changes, such as sped-up edits, can break matching.
• Very short clips may not contain enough peaks for a confident match.
• No database, no name. A fingerprint can only be matched against recordings someone has already fingerprinted.

Why it matters for searching media

In a personal media library, the question is usually less “what is this song?” and more “which of my videos have music in them, and where?” MediaFind has a Music & songs search mode and a Sound events mode among its ways of searching, both running on your computer, so you can look for footage by what it sounds like as well as by what was said. For duplicate files, byte-identical copies are caught by exact-duplicate cleanup for free, and re-encoded copies call for near-duplicate detection. More in our note on song recognition.

Frequently asked questions

Can audio fingerprinting identify a song I hum?

Usually not. Humming produces a different recording. Recognizing a hummed melody needs melody matching, which is a different technique.

Is an audio fingerprint the same as a checksum?

No. A checksum changes if a single byte changes. A fingerprint stays the same across formats, bitrates and moderate noise.

Does fingerprinting work on speech?

It can match the same speech recording, but it does not tell you what was said. For that you need transcription.