Two Musical AI Business Ideas

Original article was published on Artificial Intelligence on Medium

Two Musical AI Business Ideas

Source: John Petrucci, Dream Theater- Illumination Theory

There are two business opportunities in music that would likely take less than 6–12 months of full time work by a team of 3–5 people. In short, you could train an AI to automatically tabulate songs (i.e. turn sound into sheet music) and automatically splice songs into their individual components (i.e. turn sound into individual tracks).

Source: Guitar Pro

The first business opportunity would be to create a way of automatically turning songs into sheet music. There have been a few attempts to do this, but there are not any affordable commercial options currently available (that I am aware of). Here’s an example of a CNN that transcribes music to guitar tabs.

A company such as Ultimate Guitar or Guitar Pro could use their vast library of sheet music to train an AI. Input music files, output sheet music. I’m not claiming this would be an easy task.

Due to the complexity of the task, only a company with vast amounts of already tabbed songs would be able to take advantage of this. An alternative method would be to have musicians play songs while a network observes their behavior and have someone manually create sheet music. However, this would be cumbersome.

Nevertheless, this would be a highly valuable algorithm that could save millions of dollars in sheet music costs, as well as allow musicians to quickly learn a new song.

Source: Virtual DJ

When you are writing a remix for a song (not that I have ever done that), you often need to get access to the individual tracks for vocals, guitar, bass, drums, etc. in order to isolate parts of the song that will be mixed in a new song.

Most songs are recorded with individual components and are later merged into one single song or album. However, it is not usually easy to find those individual tracks for the purposes of mixing, since the final song has everything playing at once.

Skilled DJs can manually isolate certain tracks such as vocals by changing settings like pitch, frequencies, volume, etc.

Source: Towards Data Science

To automate this process, one would have to use a form of unsupervised learning such as principle components analysis. In PCA, individual structures are discovered from the underlying data without necessarily labeling those structures (dimensionality reduction).

Creating such an algorithm that took a raw finished song and broke it up to its individual tracks could undoubtedly be worth millions. This could be incorporated into existing software and would make it much easier for the next kid in his mom’s basement to make it big with a Lady Gaga/Fleetwood Mac remix.