MP3 tags. Where are they?

[jz78817]

@Testy

@jz78817

“Vista” doesn’t tag anything unless you do so yourself from within Explorer. If the tag info is wrong then it was because of whoever originally ripped them or the software used to do so.

jz78817

Thank you and that is probably what happened, either someone mis-tagged the file or maybe the software they used. I understand that Vista is not trying to tag anything on its own but simply reports what is in the file tags.

Thanks again

Testy

oh believe me, I ripped all of my CDs using CDEx. CDEx uses freeDB for tagging ripped songs, and I had to spend a not-insignificant amount of time fixing the retarded crap that various idiots had populated the freeDB database with.

[Testy]

I don’t know about neural networks (except that they might be able to learn how to spell my user name right ;)) but you might be able to train a Bayesian statistical filter to sort music in the same way you can train it to sort email into ‘spam’ and ‘ham’ categories. (SpamAssassin is probably the go-to implementation of this idea, but it’s very common in the real world.)

Ultimately, though, the concept is doomed because ‘genres’ are as much social constructs as anything else, and aren’t always bound very strongly to what the music actually sounds like. For example, is the 1812 Overture heavy metal? You need a lot of domain knowledge (what humans like to call ‘common sense’) to sort that stuff out, and computers don’t do that very well.

You’d likely do a lot better to use computers to recognize which song it is, and then rely on a website-based tagging system to allow humans to sort songs into genres. Spammers, trolls, assholes, and crazies will get in eventually and sort every single Justin Beiber song into ‘DICKBUTT’ and ‘CH34P V14GR4’, but that’s just how it is.

Derleth
:o Woops! Sorry for the misspelling. I’ve been trying to join the ten-fingered typists of the world. It’s easier to do than I assumed but my spelling has really taken a nosedive.

I take your point about genre being a social construct and this project is starting to sound doomed. I was spending more time than I cared for building playlists and like any good IT guy, I started wondering if I could automate the process. Building playlists using tags is one way but they are often wrong and I wanted playlists that would be unique with each run of the application.

The reason I was thinking about NNs was that the Wikipedia has quite a list of the various characteristics of Blues music (4/4 time, 12 bar format, flattened 3rd, 5th and 7th of the associated major scale, and other bits I don’t understand.) With the defining characteristics, I was thinking I might be able to snip appropriate bits out of the tunes in my library and use them to train a NN to recognize those bits in a particular song and then either add it to a playlist or drop it and go on to the next.

I was thinking about using some kind of assisted training where you presented a variety of the correct inputs to get a True state in the output and kept doing that until the NN adjusted itself to select the correct answer from other inputs. I have no idea if such a thing is even possible.

The more I think of the potential complexities of this, the more I think it may be one of those good ideas that just gets a harsh reality check from the real world. Alas!

Thanks for everything and I’ll take a look at Spamassasin.

Testy

[Testy]

@jz78817

@Testy

jz78817

Thank you and that is probably what happened, either someone mis-tagged the file or maybe the software they used. I understand that Vista is not trying to tag anything on its own but simply reports what is in the file tags.

Thanks again

Testy

oh believe me, I ripped all of my CDs using CDEx. CDEx uses freeDB for tagging ripped songs, and I had to spend a not-insignificant amount of time fixing the retarded crap that various idiots had populated the freeDB database with.

jz78817
Interesting. I had no idea that something queried a DB to get genre tags. I’ll have to give this a try and see what it does. (I’ll try to avoid CDex! )

Thanks

Testy

[si_blakely]

In the beginning, there was CDDB. This was a free internet accessible database that grabbed a unique identifier from a CD and returned the artist/track information. All this information was entered by dedicated individuals with the CDs from all over the internet. However, because not everyone can spell/proofread/capitalize you can get multiple matches and inconsistent case. Eventually the people who owned the CDDB servers decided to monetize the work of the millions of individual contributers by making access to the CDDB servers pay-only (for the developers of software that used CDDB). This annoyed both contributers and developers and so the data was cloned into freedb, which was set up running in parallel. Most free software uses freedb, and the data is just as inconsistent as it was in the CDDB days. You can usually find a good entry that matches your preferences, though, and with good software you can adjust the genre settings to your taste.

I prefer to use MediaMonkey to manipulate Music files and ID3 tags. It can deal with large collections, will auto-organize files, and can collect album art/song names from Amazon. It is really good and the free version is not limiting.

As for your idea of autoclassifying genre, you need to identify beat (which is still not trivial after years of trying), chordal structure (which I haven’t seen done in software), instrument weighting (again, does not seem to be happening), melodic/harmonic weighting and the wider structural elements of the song (verses, chorus, bridge). There are some fingerprinting approaches for melody, but trying to combine all the elements would be complex. With all these elements you may be somewhat closer to a working bayesian classification, but I would not count on it - crowdsourcing seems to me to be an easier solution.

Si

[Derleth]

Testy, I can’t imagine that automatically recognizing things like chord structure, beat, and (especially) instrumentation would be any easier than speech recognition, which seems to be stalled at an 80% accuracy rate no matter how much more data and processing we throw at the problem. I honestly have no way to translate that into a prediction of how well automatic music processing (as opposed to audio processing) would work, but I certainly wouldn’t invest in any companies founded on the assumption that it bodes well.

Maybe in a decade there’ll be a company that makes software that can transcribe a mumbled aside in a noisy café, spellcheck a drunken scrawl on a cocktail napkin, and explain Finnegans Wake through computational philology. That company will create the software that automatically sorts music into genres. But I doubt it.