In order to train the AI system, Sony’s researchers compiled data from 665 different songs from a wide range of genres including pop, rock and electronica. The songs all feature ryhthm instruments, bass, kick and snares that were available as separate 44.1kHZ audio tracks. With the contexual signals of those tracks, the researchers created drum kicks by setting a drum samples at all amplitude peaks. The AI system conditionally generates the kickdrum patterns based on the characteristics of the other material that it is placed around, regardless of the song’s tempo and changes in speed or duration.
Sony isn’t the only company that has toyed around with AI-assisted music. Facebook has used artificial intelligence to convert music of one genre to any number of other styles. Google likewise has experimented with AI’s ability to create art, building a project called Magenta that is tasked with generating musical and visual creations. Others have managed to use machine learning to create endless guitar riffs and complete musicals, though the results are pretty questionable in terms of quality.