SCI★FI★HI★FI

Machine Learning Remix

Owen Green (Developer), Matt Brennan (Composer)

Research output: Non-textual formComposition

Abstract

A machine learning remix of the "Build A Thing Of Beauty" album that is generative, unstoreable, and infinite in length.

The process works by constructing a naïve model of music as rhythmic patterns of similarity and difference at different time scales, albeit one that doesn't yet have any notion of interdependency between different voices in a song (such as pitch relationships, or rhythmic counterpoint). Each voice in the multi-track of a song is independently analysed using a small selection of machine learning (i.e. pattern recognition) and machine listening techniques to construct a 'map' that tries to estimate where the major sections, sub-sections, phrases and individual events (notes, drum hits) may be. This takes a while, so is done offline. At each of these musical time scales, the program makes a two-dimensional map that shows how ‘similar’ one chunk is to another. The algorithmic ‘remixes’ are then generated by taking the original song to be represented as particular paths taken through these maps. The original paths are warped and redrawn to produce new sequences based on the patterns of rhythm and similarity in the original.
Original languageEnglish
Publication statusPublished - 2019

Fingerprint

Song
Remix
Machine Learning
Time Scales
Rhythmic Patterns
Music
Generative
Drum
Rhythm
Length
Counterpoint
Interdependencies
Chunk
Musical Time
Pattern Recognition
Nave
Albums

Cite this

Green, O. (Developer), & Brennan, M. (Composer). (2019). SCI★FI★HI★FI: Machine Learning Remix. Composition
Green, Owen (Developer) ; Brennan, Matt (Composer). / SCI★FI★HI★FI : Machine Learning Remix. [Composition].
@misc{c584adb3f7ed4b0c9a1f49b242c3431d,
title = "SCI★FI★HI★FI: Machine Learning Remix",
abstract = "A machine learning remix of the {"}Build A Thing Of Beauty{"} album that is generative, unstoreable, and infinite in length. The process works by constructing a na{\"i}ve model of music as rhythmic patterns of similarity and difference at different time scales, albeit one that doesn't yet have any notion of interdependency between different voices in a song (such as pitch relationships, or rhythmic counterpoint). Each voice in the multi-track of a song is independently analysed using a small selection of machine learning (i.e. pattern recognition) and machine listening techniques to construct a 'map' that tries to estimate where the major sections, sub-sections, phrases and individual events (notes, drum hits) may be. This takes a while, so is done offline. At each of these musical time scales, the program makes a two-dimensional map that shows how ‘similar’ one chunk is to another. The algorithmic ‘remixes’ are then generated by taking the original song to be represented as particular paths taken through these maps. The original paths are warped and redrawn to produce new sequences based on the patterns of rhythm and similarity in the original.",
author = "Owen Green and Matt Brennan",
year = "2019",
language = "English",

}

Green, O & Brennan, M, SCI★FI★HI★FI: Machine Learning Remix, 2019, Composition.
SCI★FI★HI★FI : Machine Learning Remix. Green, Owen (Developer); Brennan, Matt (Composer). 2019.

Research output: Non-textual formComposition

TY - ADVS

T1 - SCI★FI★HI★FI

T2 - Machine Learning Remix

A2 - Green, Owen

A2 - Brennan, Matt

PY - 2019

Y1 - 2019

N2 - A machine learning remix of the "Build A Thing Of Beauty" album that is generative, unstoreable, and infinite in length. The process works by constructing a naïve model of music as rhythmic patterns of similarity and difference at different time scales, albeit one that doesn't yet have any notion of interdependency between different voices in a song (such as pitch relationships, or rhythmic counterpoint). Each voice in the multi-track of a song is independently analysed using a small selection of machine learning (i.e. pattern recognition) and machine listening techniques to construct a 'map' that tries to estimate where the major sections, sub-sections, phrases and individual events (notes, drum hits) may be. This takes a while, so is done offline. At each of these musical time scales, the program makes a two-dimensional map that shows how ‘similar’ one chunk is to another. The algorithmic ‘remixes’ are then generated by taking the original song to be represented as particular paths taken through these maps. The original paths are warped and redrawn to produce new sequences based on the patterns of rhythm and similarity in the original.

AB - A machine learning remix of the "Build A Thing Of Beauty" album that is generative, unstoreable, and infinite in length. The process works by constructing a naïve model of music as rhythmic patterns of similarity and difference at different time scales, albeit one that doesn't yet have any notion of interdependency between different voices in a song (such as pitch relationships, or rhythmic counterpoint). Each voice in the multi-track of a song is independently analysed using a small selection of machine learning (i.e. pattern recognition) and machine listening techniques to construct a 'map' that tries to estimate where the major sections, sub-sections, phrases and individual events (notes, drum hits) may be. This takes a while, so is done offline. At each of these musical time scales, the program makes a two-dimensional map that shows how ‘similar’ one chunk is to another. The algorithmic ‘remixes’ are then generated by taking the original song to be represented as particular paths taken through these maps. The original paths are warped and redrawn to produce new sequences based on the patterns of rhythm and similarity in the original.

UR - http://citizenbravo.com/research.html

M3 - Composition

ER -