Problem to Be Solved
The Skiptune database is a database of 83,000 melodies from around the world and across several centuries. These melodies are encoded in a special way (a pitch difference followed by a duration ratio), and we want to explore whether an AI model can learn melodic structure from this representation.
The Goal
Anyone can ask a computer to generate a melody by asking it to randomly generate one note after another, but doing so sounds like noise, not music. Our goal is to get the machine to generate melodies that sound as if they could have been written by humans. The 83,000 melodies in the Skiptune database represent about 5 million notes. Whether that is enough to train AI is something we’ll find out.
The Method
Nobody really understands how AI models really learn things, only that they do, and that different models are better for some types of learning than others. We can take a clue from large language models, and try to treat music notes (or, rather, pairs of them) as a “word” in a language. Music has its own grammar and syntax, and if we feed the “words” of a melody to AI, we will find out if it can learn the vocabulary, grammar, and syntax of music well enough to produce melodies of its own that sound like a human could have written them.
Mistakes Will Be Made
Because we are new to AI, we will be muddling our way through the AI training. Errors will be included at no extra charge and every bad idea documented. We will basically being experimenting with AI using the Skiptune database in controlled confusion, from which we will learn how far we can take AI in learning about our melodies. We will be intentionally candid about the development process to see how one stumbles through failure after failure before finally reaching success.
Structure of the Website
The AI Lab menu will contain the blog posts, roughly one new one each week. The rest of the site contains information about the Skiptune database itself, as constructing it was a 20-year project of entering melodies manually. You don’t need to explore the other menu items unless you are curious about the database itself. We will refer you via links to pages under those menus as needed. The website is blissfully ad free and does not earn money. It is a project of love and curiosity.