HI!
Rob from MusicBrainz/ListenBrainz here!
A fully automated process has long been a dream, but sadly its been eluding me for 2 decades now. There are too many nuances (read: stupid marketing BS) to make this an automated process. Let me give you an example as to why this is a tricky process:
Let's say you are trying to identify a track called "1. Sunday bloody sunday - U2.mp3" -- of course it has no ID3 tags. So asking MusicBrainz gives you:
https://musicbrainz.org/search?query=sunday+bloody+sunday+artist%3Au2&type=recording&limit=25&method=advanced
Ugh. OK, no problem, lets ask MB for the "album" that this track appears on. We get:
https://musicbrainz.org/release-group/c6b36664-7e60-3b3e-a24d-d096c67a11e9
There is a whole page of them and most of them CDs! Well, lets pick on and move on, right? Let's say there are 10,000 of these and none of them have tags. For each track you load, acoustid, pick the best guess on an album and move on. Yep, you can make the fully automatic, 100%.
Problem is, your tracks will be IDed, but scattered across many different albums. Or you might have some complete albums, but extra bonus tracks and weird stuff like that? You'll be left with a giant pile of unmatched stuff and a collection that is hard to listen to and nearly impossible to get solid recommendations for.
If you would like to try this for yourself, it is quite easy actually:
- Take a chunk of your music collection (a few artists, maybe 100 albums) and make a copy of it.
- Run MusicBrainz Picard and have it acoustid your collection.
- Accept everything as suggested and hit save on everything. Fix nothing.
- Now point your favorite music tool at the collection and see what the results are.
Yes, you could do a better job than Picard, but you hit diminishing returns very quickly.
Automated tagging is just not there yet!
And I proved this to myself this past week. I wrote this PROOF OF CONCEPT called, get this, auto-tag:
https://github.com/metabrainz/auto-tag
And by the time the POC was done, I basically billed it as:
This tool could be used as a new clustering tool for Picard.
Which needs manual review. Lol.
More seriously, I wrote auto-tag for Funkwhale. I want to work out if we can match whole albums much faster than individual tracks, using the MBID mapping created by ListenBrainz. Given a track name and an artist name it will make the "best educated guess" as to which track you might be interested in. (first digital release albums/CD is a very good guess). If you look up each track in the mapping and then look up ALL possible releases the track could be on, you can start to see a more complete picture of what albums could be contained in a music collection. And if the user interaction could be limited to letting the user pick their albums of the the collection and moving on, I think we'd be at an acceptable level of user supervision.
I'd love for some people to try auto-tag and give me some feedback. Just remember, its freaking proof of concept, not a production ready piece of software!
