Johann Sebastian Bach is generally considered to be one of the great classical composers, with compositions that exhibit a technical mastery of harmony and counterpoint. One of the things he excelled at was writing short polyphonic hymns known as chorale cantatas (he wrote over 300), which are short 4 voice pieces rich in harmony. As it turns out, computer scientists find these pieces very attractive because of their algorithmic-like structure. The problem is that even though you can teach a computer to compose using a similar algorithm, it’s never been particularly convincing. Until now.
Thanks to the work of Gaetan Hadjeres and Francois Pachet at the Sony Computer Science Laboratories in Paris using the artificial intelligence of a machine they call DeepBach, they’re able to produce very convincing choral cantatas that even some pros think were composed by Bach himself.
Essentially, they trained DeepBach’s neural network by teaching it all 352 of Bach’s cantatas, then transposing them to other keys for a data set of over 2,500 chorales. The machines then does its thing and before you know it, it’s composed a cantata that’s contains so much of the Bach style that even many trained listeners believe it came from the great composer himself.
How much so? A study was launched with 1,600 people (400 were professional musicians or music students) who were asked to compare two different harmonies of the same melody, then determine which of the two harmonies sounded more like Bach. When given the music from DeepBach, about half thought it was the real thing. Keep in mind that when given an authentic Bach piece to listen to, only 75% thought it came from Bach.
This is actually a very interesting step forward not so much from a composition standpoint, but more about music analysis. Bach cantatas follow a very precise structure that most other music doesn’t adhere to, but as a producer, I look forward to the day when I can get a readout as to the inner workings of a hit so I can learn from it. Hopefully DeepBach is a step towards that.
Listen to what DeepBach came up with.