Get in the Halloween Spirit with This AI-Generated Spooky Music
The results are pretty eerie, but you can help them get even creepier.
Horror movies wouldn’t be horror movies without the creepy soundtrack that accompanies them. The mood, tension, and even jump scares are accentuated by carefully crafted, and eerily familiar scores. It’s a skill that’s not as easy as it looks, as evidenced by the latest Halloween-themed artificial intelligence experiment from MIT: The Uncanny Musicbox.
Here’s the “Psycho” track of the Uncanny Musicbox:
You can definitely hear hints of the original Psycho soundtrack, as well as random spooky sounds such as a crow or an evil laugh.:
MIT researchers useda huge number of midi files and a handful of horror movies soundtracks as “primer melodies” to give the AI a starting point to make up the rest of the soundtrack.
“Our initial plan was to train the AI on a corpus of horror movie soundtracks. However, that quickly proved to be unfeasible since the model needs a much larger training set to work well,” Pinar Yanardag, who helped build the AI while she was a postdoctoral associate at MIT media lab, told me via email. Yanardag is now the CEO of AI Fiction, a creative agency in artificial intelligence.
“In the spirit of Halloween, since we wanted to generate scary songs, we primed the AI with five to ten seconds of famous horror movie soundtracks ( Exorcist, Halloween, Psycho, etc.) and had it generate new melodies as a response to those.”
The same team has created horror-themed AIs around Halloween every year since 2016. The first year they trained an AI to produce scary images, then in 2017 they trained an AI to write scary stories. This past April Fool’s, they also made a “psychopathic” AI that only sees death.
In the interactive music box, users can use sliders on a virtual mixer to add or reduce different spooky elements until the music is just creepy enough.
Over time, the researchers will collect user-generated songs, and will use them to teach the AI to auto-generate better, more authentically scary songs.
Yanardag told me they’re also hoping this data will help them get a better understanding of what makes a song scary or not.
“Is there a right combination of scary sound effects that appeals to a large audience, or do people usually have distinct tastes when it comes to scary music? If so, can we generate personalized horror movie soundtracks with such a dataset?”