Hear some samples!
MusicVAE is a machine learning model that has previously been trained on a lot of MIDI files, and has learned to generate melodies or trios. It represents each music sample as a vector of 256 features, and every different vector of 256 numbers sounds like a different piece of music. Here are some sliders for 90 of all these possible features -- we picked the most important ones to keep things simple. Slide some of them around and see how it affects the sample!
🆕🆕🆕 This new model we trained learned to summarize those 256 sliders as 4 super sliders. The super sliders control the features that make the random samples sound like your input. When you use them, the sample won't randomly change like it does when you use the regular sliders.
You can also whenever you want.
use these sliders to change some features of the sample
And this is what that sample sounds like:
The thing about MusicVAE is that it only knows generic music; when you ask it for a sample, it just picks something from the space of all possible melodies. It can't give you something that sounds more like Twinkle Twinkle Little Star, or more like your favourite Beatles song, without having to be retrained on a new corpus of data, and that takes several days on a really big GPU (not to mention millions of data points!) Enter....Midi Me!
Choose your sound
We are now going to train a small model (right here in your browser!!!), based on MusicVAE, that will try to make sure all the sampled variations will sound like some music you give it. You can or if you don't have a MIDI available.
This is what your file sounds like:
Personalize the model!
The MIDI file you uploaded had several instruments, so this is what we think the melody is. This is what we will be training on.
We're ready to start training on the MIDI that you uploaded! If you think about MusicVAE as representing the
space of all possible melodies as a giant sphere, we are now building a smaller sphere
inside of it, which represents all the melodies that are musically in the same neighbourhood as yours. The more steps you
train for, the better the results are.
When you're ready, for steps.
We've now added 4 new MidiMe sliders to Step 1: these super slliders control the features that make the random samples sound like your input. Go slide them around and see how the samples change!
Training step: 0 / 100
To check how well we're training the model, we can track its error reconstructing your uploaded file. We're not looking for it to become exactly 0 -- just to go down and flatten out.