A.D.A.M - ANN Music Generator

A.D.A.M

Procedurally generated ambient music is an area of A.I that, although has been used for games and other media since its introduction in 1986, is still a relatively new way of producing music quickly and effectively.

An artificial neural network was constructed to generate ambient game music and the network was then evaluated, using ambient music from games as a benchmark, to find if features such as repetition and musical themes were prominent or featured in the generated pieces of music.

The network was constructed using the coding language Python and utilized libraries such as Tensorflow and Keras to construct an LSTM centred network capable of generating music based on training data in the form of midi files. The generated pieces show some understanding of the structure of the training examples provided and was able to produce pieces of music that are both appealing and have the potential to be used as music within games. This project evaluates the level of learning achieved by the Artificial Neural Network and the suitability of the pieces of ambient music created for use within games with a discussion covering how to improve the overall network’s learning abilities as well as how to improve the pieces of music generated.

Results:

The network was able to produce music with similar aspects to the training data which composed of music from games series such as Animal Crossing and the Zelda franchise. With further expansion to the network this can prove to be a viable way of producing music or providing a start into music creation for audio designers and game makers alike.

Below is a video showing off some of the pieces produced by A.D.A.M:

(Music is not shown accurately as A.D.A.M was unable to 'pick' a time signature, most music is roughly 4/4 but is not displayed as so)