magistrsko delo
Žiga Kleine (Author), Luka Čehovin (Mentor)

Abstract

Računalniško generirana glasba lahko skladateljem služi kot pomembno orodje za krepitev glasbene ustvarjalnosti in širjenje znanja o glasbeni kompoziciji. Zaradi tega je računalniško generiranje glasbe dobro raziskano področje, ki se je z razvojem globokih nevronskih mrež močno razcvetelo. Trenutno predstavlja velik izziv na tem področju generiranje daljših stilsko konsistentnih glasbenih sekvenc. Drug izziv, s katerim se pogosto srečamo, je nadzor nad lastnostmi generirane glasbe. Obeh omenjenih izzivov smo se lotili z implementacijo difuzijskega verjetnostnega modela za odpravljanje šuma, sposobnega ustvarjanja izvirnega glasbenega zapisa, ki sledi obliki glasbenih spremljav videoiger igralne konzole NES, hkrati pa je proces generiranja mogoče voditi s pomočjo avtomatsko pridobljenih čustvenih oznak. V okviru magistrskega dela nam je uspelo implementirati difuzijski model, ki smo ga ovrednotili s pomočjo glasbenega Turingovega testa, v katerem so računalniško generirane melodije anketirance v 43% primerih preslepile za melodije, ki jih je napisal človek.

Keywords

pogojeni generativni modeli;difuzijski modeli;generiranje glasbenega zapisa;računalništvo in informatika;magisteriji;

Data

Language: Slovenian
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [Ž. Kleine]
UDC: 004.8:78(043.2)
COBISS: 178611715 Link will open in a new window
Views: 32
Downloads: 3
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: English
Secondary title: Conditioned symbolic music generation with diffusion models
Secondary abstract: Computer generated music can serve as an important tool for composers to enhance their musical creativity and expand their knowledge about musical composition. This makes computer music generation a well researched field that has expanded greatly with the development of deep neural networks. Currently, generating longer stylistically consistent music sequences is a big challenge. Another challenge we often encounter is controlling the properties of the generated music. We tackled both of the mentioned challenges by implementing a denoising diffusion probabilistic model capable of generating original music notation that follows the form of the soundtracks for the NES game console, while the generation process can be guided by automatic emotion annotations. We managed to implement a diffusion model, which we evaluated with the help of a musical Turing test, where in 43% cases computer generated melodies confused the test subjects for human written melodies.
Secondary keywords: deep learning;conditioned generational models;diffusion models;symbolic music generation;computer science;computer and information science;master's degree;Globoko učenje (strojno učenje);Glasba;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS): Master's thesis/paper
Study programme: 1000471
Embargo end date (OpenAIRE): 1970-01-01
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: 81 str.
ID: 21492991