It takes years of practice for a pianist to become skilled enough to improvise, hitting the ivories in a way that is both instantaneous and something a listener would want to bop their head to. Now, even the most tone-deaf among us can take a whack at it with the help of an AI system called Piano Genie.
The team behind Piano Genie trained a neural network on a dataset made up of about 1,400 pieces of piano performances from an international youth competition. The system learned what notes typically follow one another, much like your phone’s predictive texting tries to guess what you want to write next.
Chris Donahue, a PhD student at UC San Diego and an intern at Google’s Magenta project, where he was one of the researchers that worked on Piano Genie, says some inspiration came from Guitar Hero, the video game that lets players bang on a plastic guitar’s buttons, following the notes displayed on the screen.
In the spirit of Guitar Hero‘s simplified guitar “playing,” they reduced the standard number of 88 keys on a piano into just eight buttons to keep it manageable for non-musicians. (A web demo can be played with here, and a video of what it looks like when an actual piano player tries it can be watched on YouTube.)
Most AI-created art is nonsensical or derivative, like this screenplay or this Harry Potter chapter. Humans still need to be in the loop for something that is traditionally entertaining to come out of AI models. Fortunately, music researchers have been working on AI-assisted products since the 1980s. That work has spawned tools like Logic, which uses AI to create drum patterns and another Google Magenta project called NSynth Super that relies on neural networks to combine instruments and create new sounds.
Other AI music projects like Amper and Flow need some configuring before popping out a completed (but still configurable) piece of music. Piano Genie is ready to go after loading and improvises note by note, learning from the player as they go. And every time a user refreshes the demo site, the neural net behind the system is reset and starts creating anew.
“A nice property of these generative systems is they can surprise you,” Donahue says.
Donahue says the team, which also includes DeepMind’s Sander Dieleman and Google’s Ian Simon, are looking into similar technology that could be useful for other musicians. There is no compilation of guitar-playing samples that match the quality of the piano competition dataset, so the more immediate goal is to grow Piano Genie’s musical knowledge. Training the system to consider aspects like chord progression and scales would create a sophisticated tool that could let even novice musicians jam with more experienced players.
Even though Piano Genie is geared toward novices, it hasn’t reduced the enjoyment of playing for Donahue, who has played piano for 20 years.
“It’s fun to wail on it,” he says.