Your OpenAI Jukebox scientists were so preoccupied with whether they could, they never stopped to think if they should.
Details on the Neural Network, from the YouTube page:
What is OpenAI Jukebox?
Simply put, it's a neural net that generates music – read more about it on the official blog: https://openai.com/blog/jukebox/
Can I use Jukebox myself?
Yes! If you have a very powerful graphics unit, you can download the program from the official Github. If you don't, you can give Google CoLab a try – https://colab.research.google.com/git...
Is there a tutorial? Yes, I made a tutorial explaining the basics about generating samples through Google Colab. https://www.youtube.com/watch?v=mNtmg… I'd also recommend checking out this site, it has some useful info towards the bottom: https://github.com/openai/jukebox/iss…
How does it know the lyrics?
You can feed it lyrics when setting it up.
What has Jukebox been trained on?
According to OpenAI, it has been trained on 1.2 million songs. Unfortunately, OpenAI has not publicly disclosed which songs are in this dataset.
Are there longer versions of these samples?
No, they would have to be generated first. It would take a long time and the sound quality would likely deteriorate even more.
What sampling temperature are you using?
It varies. But always between .975 and 1.
Does the original artist get compensated if you use their song to train the AI and upload the result to YouTube?
When doing a song continuation like this, in most cases, yes. YouTube's algorithms can detect most songs, even if they're only played for a few seconds. The record label will then automatically claim the video and make money off of it.