Imagine if you could create music like Coldplay or background scores like Hans Zimmer! Would be really cool, right? Well, that's precisely what the Open AI Jukebox let’s you do with its deep learning algorithm!

You provide the Jukebox a genre, artist and the lyrics and it creates a medley of pieces imitating the different artists you chose. Now, what about a scenario where you have an initial melody but just not able to complete it? Again, Jukebox is your friend. You provide it with the first few seconds of your piece, and it auto-completes your music.

How did they come up with this?

Music generation via computers has been around for more than 50 years. The most common approach is to record each instrument separately and then combine them together using computer software. But most of the generators are not fully capable of capturing the subtle notes. This is where generative models were being used. 

By using an autoencoder, you can compress the raw audio and discard the noise (irrelevant bits of music) and then train your Neural Network with the compressed audio and then resample it back to raw music. 

The team at OpenAI, created a dataset of 1.2 million songs for neural network learning. The metadata includes artist, album genre, and year of the songs, along with common moods or playlist keywords associated with each song. The neural network then starts clustering - it groups the similar artists and similar genres, such as clubbing Maroon 5 with Sia and Miley Cyrus.

With a few song samples already generated, they are still a work in progress to achieve their ultimate goal - creating a perfect Hans Zimmer score!