I heard that the reason why music identifiers don’t work offline is because their databases are huge, with millions of songs (including video game music). Databases containing only video game music may be significantly smaller, even if they contain the soundtracks of over 1000 games.
Google pixel phones have automatic offline music recognition where the database is something like 50MB IIRC and it’s pretty good unless I’m listening to something particularly niche
I assume you could build a similar database if you had the source material to do so.
You’re potentially gonna have an issue with games that have dynamic soundtracks that aren’t exactly the same every time (think in an action game how the music changes based on if fighting or something)
Damn, how do they keep the track metadata in such a small file?
I feel like my iTunes (Apple Music) library data is at least that big.
You had me wondering so I looked it up
I dropped a zero, it’s 500MB (and may be a bit bigger now perhaps)
https://venturebeat.com/media/how-googles-pixel-2-now-playing-song-identification-works
Sounds like quite the endeavor. I’m just curious how you would collect recordings of all video game music for the database. Music is distributed as tracks on an album, and I don’t know if you can easily rip that data from the code, or if someone would have to record it from gameplay.
Music identification services like Shazam get it the same way as Spotify
Right, but they wouldn’t be able to identify Wood Man’s theme from Mega Man 2. As far as I know, nobody has collected and included those in the recognition software.
You’d be surprised what they have in their database
Also, just tested that song and Shazam didn’t have any problem identifying it
This post was a spinoff of another post from a few days ago asking why most video game music isn’t in those databases. So, I started from that assumption. I probably just picked a bad random example, because wood man slaps.
True
Although those databases really have basically everything
Once I was sitting in a train and pixel randomly detected the high pitched wirring of the accelerating train as a random Himalayan pan flute song with a 3 digit playback count on Spotify
You probably could. But what’s the target demographic for this?
How many songs does a typical game have? You could just look up the soundtrack and play from the 12 different songs until you find the right one.
deleted by creator
Possibly. Using meta data to identify sound should be relatively small. Then, sound samples from the first few seconds would allow comparison without the burden of having everything. Getting the data and samples is probably the hardest part.
There are open source tools to do this kind of thing, you should see if they have offline versions.
I’ve used music brainz picard in the past.
Problems:
- Game music isn’t often released on an album
- Some games use vertical mixing (dynamic tracks that separate instruments to allow the game to vary the song’s intensity)
- Some games use horizontal mixing (dynamic tracks that are indexed for the game to dynamically switch beats or add bridges between songs at marked points for seamless soundtrack changes)
- Some games use both
deleted by creator
In terms of raw horsepower as well as storage, probably. People were already doing audio fingerprinting in the early '90s on, like, 386 hardware. It ought to be trivial for anything even vaguely modern. The entire fingerprint data for a 2 minute chunk of audio is 3-4 kB or something. I don’t even think the size of the database will be hugely onerous.
