AI composing new old Finnish folk songs

As a first step to music and AI research, I have trained a neural network on a database of Finnish folk songs originally published in 1898 - 1933. The data was digitised in 2004 by a team in University of Jyväskylä, so midi files of the monophonic melodies are available for training of the network. Here's a typical Finnish folk song. It's called 'Kaksipa poikaa Kurikasta', 'Two lads from Kurikka'.

And here's one example of a continuation that a long short-term memory neural network, with two hidden layers of 128 nodes, wrote to two bars of it. The training was performed on a fraction of the database only, so the result isn't probably the most beautiful piece of music you've heard. However, there are some promising features in it. For example, it clearly makes chord changes between the first and the fifth degree, and it almost manages to perform an authentic cadence V7 - I at the end of the phrase.

There is clearly place for improvement in compositional skills of the present network. After this first test, we will see, how skilled a folk song composer the network can be. More complex network architectures need to be considered as well. The database of folk songs also has some interesting metadata, such as geographic origin and lyrics of the first verse, which together with melodies form an interesting training set.

The video below is from a Finnish popular science show 'Prisma Studio' on national TV. It's me performing the LSTM-composed song in Seurasaari, Helsinki. I accidentally play C instead of D as the first note of the last bar, making it sound more like a proper cadence. To err is human.

#Neuralnetwork #folkmusic