Back to search

Ionian University

Conditional machine learning architectures for harmonic and rythm music generation

Abstract

dc:description

Algorithmic Music Composition (AMC) deals with the creation of a rigid, well-defined system for the process of composing music. It is considered as one of the most popular applications of Music Information Retrieval (MIR) research field with countless approaches over the past years. This dissertation deals with the developing of innovating AMC architectures, specifically for harmonic and rhythm generation. Besides, it contributes indirectly to related MIR applications such as with the creation of a dataset and developing a voice separation algorithm in the symbolic domain. Specifically, the most significant contributions of this thesis can be summarised as follows:- Creation of the first symbolic dataset containing Greek traditional and popular music called the Greek Music Dataset (GMD); a collection of 1400 Greek tracks which offers: pre-computed audio, lyrics & symbolic features for immediate use in MIR tasks, manually annotated labels pertaining to mood & genre styles of music, generic objective metadata, a manually selected MIDI file (available for 500 of the tracks) and a manually selected link to a performance/audio content in YouTube for each track. According to Related Research, GMD can be considered one of the most extensive datasets in the symbolic domain. - The refinement of a Rule-Based Voice Separation algorithm in the symbolic domain called VISA3. This revision builds upon its previous editions by the introduction of new characteristics that adhere to general perceptual principles, address assignment errors that accumulate affecting the precision and tackle more generic musical content. It also differs with related attempts by defining the term “voice” as a perceptually independent, and not necessarily monophonic, sequence of notes or multi-note simultaneities. In addition, a manually annotated subset of the GMD is introduced. Experimental results indicated the significant performance amelioration the proposed revision achieves in relation to its predecessors. - The introduction of the first modular probabilistic system for examining Bass Voice Leading in Melodic Harmonisation. Since most of the relevant approaches do not take into account the horizontal relation of notes between successive chords, a combination of probabilistic and statistical modules is proposed which defines the motion of the bass voice according to several statistical aspects: (a) melody voice contour and previous bass line motion with the utilisation of a Hidden Markov Model, (b) statistics regarding the absolute difference between the bass voice and the melody and (c) statistics regarding inversions and note doublings in chords. The experimental results on diverse music idioms indicated that the proposed methodology captures the characteristics of each idiom efficiently, thus leading to the conclusion of the effectiveness to generate the bass voice-first during melodic harmonisation. - An innovating Deep Learning Architecture called Conditional Neural Sequence Learners (CNSL) for rhythm (drums) generation. Considering music as a sequence of events with multiple complex dependencies on various levels of a composition, the Long Short-term Memory-based (LSTM) architectures have been proven to be very efficient in learning and reproducing musical styles. However, they are hardly useful for tasks that incorporate human input or generally constraints. To address this problem, a novel solution is presented that harnesses the LSTM module with a Feed-Forward (FF) part, which is called the Conditional Layer. The main contribution of this architecture relies on the ability that CNSLs are able to combine implicitly learnt information (LSTM module) and explicitly defined conditions (FF module). Three variations are presented highlighting different representations and feature selection on LSTM and FF modules, respectively. The experimental results indicated the effectiveness of the CNSL architecture in producing drums’ sequences that (a) resemble a learnt style (implicit learning) and (b) keep the entire network “on-track” and enable it to respond to constraints that were not encountered during training (explicit learning).

Degree

thesis:*
Grantor dc:publisher
Ionian University
Year dc:date
2020

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Makris, Dimosthenis
  • Μακρής, Δημοσθένης

Subjects

dc:subject × 20

Rights

dc:rights
Statement dc:rights
  • BY
Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/47656

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Makris, Dimosthenis; Μακρής, Δημοσθένης. Conditional machine learning architectures for harmonic and rythm music generation. Ionian University, 2020. http://hdl.handle.net/10442/hedi/47656