It's not very significant. Basically it stores some distances between peaks in each song's spectrogram which can then be further compressed.

Even supporting a database of millions of songs would be possible.

If each signature occupies an effective space of 1k (not sure how feasible this is), then for 1000 songs this would take 1M, 10M for 10k songs, etc

Every year gives us around 100 popular songs (add a % of location-specific popular ones), so it seems the plan is feasible.

Spotify has 30 million songs.

Even if it takes up a small amount of space, it's basically a non-feature.

>Spotify has 30 million songs.

Many of those songs have never been played[1]. There's a really (really, really, really) long tail.

[1]: http://forgotify.com/

The index can be highly compressed, see for example

https://blog.afterthedeadline.com/2010/01/29/how-i-trie-to-m...