AcousticBrainz: Making a hard decision to end the project (2022)
blog.metabrainz.org
blog.metabrainz.org
Reading between the lines here -- AcousticBrainz can't store copies of any actual audio for copyright reasons. Instead they store various fingerprints and derived properties of the audio signal (example[1]). When the project started they didn't anticipate the right kind of input data required for these newest AI techniques, so to make progress they would have to throw away 7 years of data and start over their data acquisition from scratch. Instead they decided to shut it down.
Just another way the RIAA is making the world a worse place to live.
[1] https://acousticbrainz.org/api/v1/6df2d3b5-25e1-4d4c-a196-04...
"…we don’t have the resources and developer availability to perform this kind of research ourselves, and so we rely on the assistance of other researchers and volunteers to help us integrate new tools into AcousticBrainz, which is a relationship that we haven’t managed to build over the last few years."
If the RIAA was to blame, Pandora's Music Genome Project and other music analysis/recommendation systems wouldn't exist.
I wonder if that license extends to deriving a ML model from it. I know there's some ML models out there already that can produce music based on a prompt, but that'll be limited to what music the authors have available to them. Spotify (and Apple Music, and the others) have millions upon millions of tracks available to them.
Do they?
I have ‘melancholic’ and ‘I want to hype myself up’ playlists. Both from the tone of the music, voice, and lyrics (even the titles, I’d say) it is extremely obvious what the overarching vibe of the playlist is. Yet the radio-mode of those playlist or the recommendations below the playlist are fairly obviously just “other people who played the songs in this playlist also liked these songs”.
As far as I can tell there are no deep analysis smarts at work, or even something like EveryNoiseAtOnce.
I'm sorry it didn't work out.
Is there a possible pivot to something else non-music related, or adapting your technology to another industry?
See [2] for the list of models currently available. It includes updated ML models for the classifiers used in AcousticBrainz and many new models (e.g., a 400-style music classifier, models for arousal-valence emotion recognition, and feature embeddings that can be used for audio similarity or training new classifiers).
[1] https://essentia.upf.edu/ [2] https://essentia.upf.edu/models.html
The Weather apps in the latest versions of iOS, iPadOS, and macOS are very obviously re-styled versions of DarkSky. They even rebranded and released the API.
I suspect that they just didn't have the right people on the team, or those people were busy and haven't gotten around to it.
This is sad, but it's an important cautionary tale for businesses that plan to collect data first, and later hire a data person to turn that data into gold. Many, many businesses in the 2010s probably failed because they thought they could do this.
The truth is that you need data people involved from the beginning, and continuously throughout the project, in order to monitor and evaluate the data being collected, building proof-of-concept models along the way, and to adjust the data collection process as problems are discovered and new techniques are developed.
And yes, another vote here for early proofs of concept and rigorous testing. Not everybody gets what really matters in data quality, but the surest way to find out is to try to build something and see if it really works.
You are correct, those issues were apparent right from the beginning, and they never really got better. The acoustic fingerprinting worked sorta OK for very popular albums -- but even for that, it was never 100% accurate. It never worked well for live performances, imports, classical music, jazz, or jam bands.
Using this software always required a lot of manual intervention, which at least for me negates the whole point of using it in the first place.
Is.
Hard.
Understanding it and building useful models of it are really really hard.
Even building a functional content-driven also-bot is hard, although you can always solve that problem by cheating and aggregating playlist preferences.
Supposedly simple concepts like 'track BPM' just don't work reliably in the real world. (What's the tempo of a recording of a symphony, or even just a folk album that wasn't recorded to a click track? Or an EDM track with subtle tempo shifts - which quite a few tracks have?)
If they'd known more about music when they started they'd have understood this.
Sure it does, you just have to give up on the idea of there only being a single value. A song could have multiple different BPM values, so in your database you'd just record a set (or a tuple in Python parlance). For subtle tempo shifts, you'd run the music through some kind of filter algorithm which would isolate sections with significantly different tempos, then for each section would find the average, then would make a set with those values and store it. The only place this would fall down is some weird music with constantly-changing tempo, where you'd end up with an average over the whole song.
Echonest used to provide source code, but after being bought by spotify, they stopped updating their code, disabled fingerprint server, etc: https://github.com/spotify/echoprint-codegen
So if I am sad, I am actually also really glad and impressed the people behind a project recognized when to stop, and rationalized how to do it.
Looks like they stored computed outputs from the “Essentia” tool, and the values are not accurate, so training a model on top of that will render equally innacurate results.
To improve it you’re [pun not intended] essentially starting from scratch.