Stvaranje glazbe pomoću generativnih modela

master's thesis

• Title (english): Creating music using generative models
• Author: Sven Kadojić Balaško
• Mentor: GORDAN KREKOVIĆ (mentor)
• Granter: University of Zagreb; Academy of Music; (Wind Instruments Department); Zagreb
• Defense date and country: 2025-02-14, Croatia
• Scientific / art field, discipline and subdiscipline: FIELD OF ART; Art of Music; Composition
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing; Artificial Intelligence

Abstract

Ovaj rad istražuje generiranje glazbe pomoću strojnog učenja, s posebnim naglaskom na transformere i difuzijske modele, koji su se pokazali kao ključni pristupi u modernim generativnim sustavima. Detaljno su objašnjene arhitekture oba modela, uključujući njihove ključne komponente, poput mehanizama pažnje kod transformera i postupaka iterativnog uklanjanja šuma kod difuzijskih modela. Analiziraju se metode kvantizacije zvuka i tekstualne uvjetovanosti. Rad analizira suvremene modele poput MusicLM, MusicGen, Riffusion i Stable Audio Open, ističući njihove mogućnosti, izazove i potencijalne primjene. Generiraju se glazbeni primjeri za svaki model. Također, obrađuju se etička i ekološka pitanja povezana s razvojem generativnih modela, kao i njihov utjecaj na glazbenu industriju. Unatoč ograničenjima, generativni modeli predstavljaju važan alat u transformaciji načina stvaranja i konzumiranja glazbe.

Abstract (english)

This thesis explores music generation using machine learning, with a particular focus on transformers and diffusion models, which have proven to be key approaches in modern generative systems. The architectures of both models are explained in detail, including their core components, such as attention mechanisms in transformers and iterative noise removal processes in diffusion models. Sound quantization and textual conditioning are analysed. The paper examines contemporary models such as MusicLM, MusicGen, Riffusion, and Stable Audio Open, highlighting their capabilities, challenges, and potential applications. Music examples are generated for each model. Additionally, ethical and ecological issues related to the development of generative models, as well as their impact on the music industry, are addressed. Despite their limitations, generative models represent an important tool in transforming the ways music is created and consumed.

• Keywords: strojno učenje; generativni modeli; glazba; simboličko generiranje; audio generiranje
• Keywords (english): machine learning; generative models; music; symbolic generation; audio generation
• Language: croatian
• URN:NBN: urn:nbn:hr:116:479659
• Study programme: Title: Integrated Undegraduate and Graduate University Study Programme for Instrumentalists; specializations in: Flute, Oboe, Clarinet, Saxophone, Bassoon, Horn, Trumpet, Trombone, Tuba, Percussion, Harp, Violin, Viola, Guitar, Piano, Organ, Harpsichord, Euphonium, Tamboura, Violoncello, Double Bass; Course: Clarinet; Study programme type: university; Study level: integrated undergraduate and graduate; Academic / professional title: magistar/magistra muzike (magistar/magistra muzike)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2025-03-05 16:40:45