Song recognition not working? How Shazam-style apps work and why they miss
You hold your phone up, it listens for a few seconds and names the song, even over chatter and clinking cups. Then it fails on something that seems easy. Both make sense once you know what it's actually matching.
What the app actually listens for
A recording is turned into a spectrogram: time along the bottom, pitch up the side, loudness as brightness. The app keeps only the brightest spots, like stars on a map, because the loud parts of a song stay loud even when someone is talking over it. It then pairs each star with a few that come just after it (“this pitch, then that pitch, this far apart”) and turns each pair into a short number called an audio fingerprint. Those numbers are looked up in an index built from millions of recordings. Lots of pairs match songs by chance, but for the right song they all line up on the same moment, so the votes pile up there.
You can play with a small version of this, with made-up tunes and no microphone, in the free Wirebook song recognition lesson.
Why it fails, and what to try
| Situation | Why it misses | Try |
|---|---|---|
| Live gig, cover, acoustic version, remix | A different performance has different timing and pitches, so its fingerprint doesn't match the studio one. | Note a few lyrics and search them; some apps do know popular live albums. |
| Humming, whistling or singing it | One voice with no instruments, slightly off pitch and tempo, leaves almost none of the original fingerprint. | Use a hum-to-search feature (in some search and music apps), which matches the shape of the melody instead. |
| Music very quiet or far away | Background noise is louder than the song's peaks, so the stars it picks are the noise. | Move closer to the speaker, point the phone's bottom microphone at it, wait for a louder part. |
| Only a second or two, or a talky intro | Not enough matching pairs to beat chance matches. | Let it listen for 10–15 seconds, ideally over the chorus. |
| Sped up, pitch-shifted or slowed clip from a short video | Changing speed or pitch moves every star, so the pairs no longer match. | Search the video's sound page or caption, or find the original speed version. |
| Very new, obscure or unreleased track | It isn't in the index yet. | Try again in a few weeks, or search a lyric. |
| No result at all, instantly | Usually no microphone permission, or no connection for the lookup. | Check the app's microphone permission; offline apps save the clip and look it up later. |
| Music playing through your own earbuds | The microphone can't hear sound that only goes into your ears. | Use the app's “listen to other apps” or on-screen recognition if it has one, or play it on the speaker. |
Is it always listening?
A song recognition app listens when you press the button (or for a set time if you turn on auto mode). Some phones have a separate on-device “now playing” feature that matches against a small song list stored on the phone without sending audio anywhere; it can be turned off in the sound or lock screen settings. Either way, what's compared is the fingerprint, not a recording of the room.
Does it use much data?
Very little. A fingerprint is a few kilobytes, far smaller than the audio, which is why recognition works even on one bar of signal. Playing the song afterwards is the part that uses data; see how much data streaming uses per hour.
Related
How a phone knows a song (interactive lesson) · Data per hour of streaming · Mobile data running out fast · What airplane mode turns off
Questions
Why can't Shazam find my song?
It matches the exact recording. Live versions, covers, remixes, sped-up clips, humming and very quiet or very short snippets don't match the fingerprint in its index.
How does song recognition work?
The app turns a few seconds of sound into a pitch-over-time picture, keeps the loudest points, and looks up pairs of those points in an index of recordings; the song where the most pairs line up at the same moment wins.
Can song recognition work in a noisy place?
Usually yes, because the loudest parts of the music survive background chatter. It fails when the noise is louder than the music.
Why doesn't humming work?
Humming is a different performance with no instruments and slightly different pitch and timing, so it leaves almost none of the original fingerprint. Hum-to-search features match the melody shape instead.
Published .