Primerjava postopkov globokega učenja za samodejno tvorjenje glasbe

Turk, Tom (author), Štruc, Vitomir (mentor)

V diplomskem delu smo preizkusili možnosti ustvarjanja samodejno tvorjene glasbe z uporabo različnih vrst nevronskih mrež (RNN, CNN, LSTM, GRU). Uporabili smo orodja Google Magenta, Google Wavenet in GRUV. Za uporabo orodja Google Magenta smo morali zvočne podatke pretvoriti v MIDI in MXL predstavitve zvočnih datotek, medtem ko Wavenet in GRUV sprejmeta surove zvočne podatke. Nevronske mreže smo učili s pomočjo množice podatkov elektronske glasbe ter množice podatkov jazz glasbe. Na koncu je sledila evalvacija rezultatov, izvedli smo slušne teste in analizirali rezultate.

• Language: Slovenian
• Keywords: samodejno tvorjena glasba, nevronske mreže, RNN, CNN, LSTM, GRU
• Work type: Bachelor thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2018
• PID: 20.500.12556/RUL-102746
• Publication date in RUL: 07.09.2018

Secondary language

• Language: English
• Title: Comparative analysis of deep learning models for automatic music generation

Abstract

The thesis addresses the possibility of the automatic music generation, using a variety of neural network topologies (RNN, CNN, LSTM, GRU). In this thesis, the capabilities of projects Google Magenta, Google Wavenet and GRUV were explored. Google Magenta operates on the MIDI and MXL files, so we had to convert the files from the source datasets to the MIDI and MXL files first, whereas Google Magenta and Google Wavenet operate on the raw audio files. Listed neural networks were trained on a dataset, containing samples of electronic music, and a jazz music samples dataset. Finally, we evaluated the results using an auditory test and analyzed the results.

• Keywords: generative music, neural networks, RNN, CNN, LSTM, GRU