Hi,
I’m working on similarity song recommendations system called AudioMuse-AI. I first start with Spotify ANNOY and then Moved to Spotify Voyager.
The data that I’m using for this task are 200 size embedidng vector extracted with Musicnn Tensorflow Embedding from Essentia-Tensorflow, then adapted to be extracted with Librosa.
From what I’m looking is that using embbeding vectors to find pattern seems more useful than just stick to the limited at of label of the genre. I don’t have an exhaustive dataset to say if this perform well on all the possibile song, but for what I’m looking (and my small amount of user are looking) the result is not bad.
You can take a look here on the open-source GitHub project: https://github.com/NeptuneHub/AudioMuse-AI
I was thinking if you had test on this kind of data on AcousticBrainz and if yes, what was the result ?
I’m asking because due to the good results obtained (at least till now) working with this embbeding vector, I was thinking if collecting this data in a centralised database like MusicBrainz could be a good idea too.