Year: 2021
Type: dissertation
Abstract: Automatic polyphonic music generation is a challenging temporal task. Asidefrom the temporal dimension, a model dealing with polyphonic music shouldunderstand which combination of notes are consonant and can be groupedtogether into chords. The ability of deep generative models to learn from bigcollection of data makes it a suitable approach to model complex data distri-butions such as the one underlying music. Among them, the original MusAEmodel leveraged an adversarial autoencoder to generate smooth interpola-tions between two polyphonic musical sequences made by monophonic se-quences. A problem presented by the original MusAE is that the generationof musical sequences by sampling from the latent space lead to inconsistentsamples. This thesis extends the original MusAE integrating a state-of-the-artmodel for sequence to sequence tasks: the Transformer. Moreover, the eventdata representation is replaced with a new type of representation that, to-gether with the relative positional encoding, allows to deal with polyphonicinstruments. The experiments show that the new MusAE is able to deal withpolyphonic instruments and to reconstruct songs with a similar style to thosein input. Furthermore it allows for smooth interpolations between two mu-sical sequences and it is able to generate new musical sequences with appre-ciable and coherent dynamics and structure.
Author: BERTI, STEFANO
Institution: University of Pisa
Language: English
Subfield: Music
Field: Arts and Humanities
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7002463227