thesis
[2021]
This thesis implemented a Neural Network model for the style transfer of music between three specific genres (genre-to-genre). Instead of audio, MIDI files are used, through which symbolic data of music representation can be extracted and provide information similar to sheet music. This approach brings out the compositional differences between the genres.
A simplified diagram describing the transfer process. Each genre has a separate decoder in a shared Autoencoder architecture.
Presented below are samples of a specific transfer, reproduced as audio tracks after being given timbre through a Digital Audio Workstation.
Original Track Input
While the piano sound can be obstructive, the characteristics of a standard rock track can still be found in this snippet. The melody is strictly in a minor/major scale, notes have moderate lengths and are restricted in one position of the scale.
Produced Genre Transfer 1
In this genre transfer from rock to jazz, the track loosens in form, both in terms of harmony and rythm. The melody isn’t confined in a single scale and the notes vary greatly in length.
Produced Genre Transfer 2
Taking on compositional characteristics of electronic music, the notes become shorter in length and “jump” quickly between different positions of the scale, giving an artificial undertone to the track.
Since MIDI tracks can be given the characteristics of essentially any instrument, a piano sound was chosen to reproduce the total melody of each track as it was considered to have fairly neutral timbre, enabling an easier comparison at a compositional level.