1 of 12

1

Between the AI and Me:

Analysing Listeners' Perspectives on AI-and Human-Composed Progressive Metal Music

Pedro Sarmento, Jackson Loth, Mathieu Barthet

​

C4DM, Queen Mary University of London, United Kingdom

ISMIR 2024 @ San Francisco, USA

UKRI Centre for Doctoral Training (CDT) in Artificial Intelligence and Music (AIM)

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

1

2 of 12

2

  • Online listening test + post-task questionnaire
  • AI-generated symbolic music (originally in tablature format)

​

  • Investigates the effects of:
    • Generation type (human vs. AI)
    • Genre (progressive metal vs. rock)
    • Curation process (random vs. cherry-picked)

​

  • Statistical analysis of results with Friedman and post-hoc pairwise Wilcoxon tests, plus thematic analysis

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

2

3 of 12

3

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

What is progressive metal? 🤘

  • Subgenre of metal
  • Builds on progressive rock’s complex phrasing and odd time signatures
  • Incorporates a heavier focus on guitars and metal influences
  • Example bands:
    • Dream Theater, Between the Buried and Me, Periphery, Protest The Hero, Opeth, Meshuggah, Animals As Leaders;

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

3

4 of 12

4

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

4

5 of 12

5

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

5

6 of 12

6

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

6

7 of 12

7

H1: Human-composed music obtains better scores than AI-generated music in terms of preference, creativity, consistency, playability and repeatability.

​

H2: AI- and human-composed music can be distinguished - Musical Turing Test.

​

H3: AI-generated music matches the genre used for model conditioning (namely progressive metal).

​

H4: Cherry-picked AI-generated music is preferred over randomly chosen AI-generated music.

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

7

8 of 12

8

  • Stimuli (60 examples, 10 per sub-category)
    • Categories
      • Generated progressive metal (cherry-picked vs random) (progcp, progrand)
      • Generated rock (cherry-picked vs random) (rockcp, rockrand)
      • Human-composed progressive metal and rock (proghum, rockhum)
    • All stimuli rendered with the same synthesiser
    • Generated using Transformer-based models ProgGP[1] and Genre-CTRL[2]

​

​

​

  • Participants (32)
    • Progressive metal fans recruited from communities on specialized Reddit and Discord channels

​

​

[1] J. Loth, P. Sarmento, C. Carr, Z. Zukowski, and M. Barthet, “ProgGP: From GuitarPro Tablature Neural Generation To Progressive Metal Production,” CMMR, 2023.

[2] P. Sarmento, A. Kumar, Y.-H. Chen, C. Carr, Z. Zukowski, and M. Barthet, “GTR-CTRL: Instrument and Genre Conditioning for Guitar-Focused Music Generation with Transformers,” EvoMUSART, 2023.

​

​

​

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

8

9 of 12

9

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

9

10 of 12

10

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

  • Human-composed music obtains better scores than AI-generated music in terms of preference, but not necessarily other dimensions (H1).

​

  • AI-generated and human-composed music can sometimes be distinguished (H2):
    • Participants struggled to determine if excerpts were AI-generated in the progressive metal (cherry-picked) and rock (random) groups

​

  • The ProgGP model seems to excel more at generating music in its target genre than the rock model (H3).

​

  • Cherry-picked AI-generated progressive metal songs tend to be preferred over randomly generated ones (H4).

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

10

11 of 12

11

  • What features made you identify excerpts as progressive metal?
    • Complexity (40), Composition/style (38), Instrumentation (7);

​

  • What features made you identify excerpts as rock?
    • Musical structure/composition (24), Simple/straightforward (23), Guitar techniques (14), Instrumentation (11);

​

  • What made you identify excerpts as being composed using AI?
    • Something “off” about the composition (40), Repetition (14), Uninteresting/simple (8), Melody (7);

​

  • What made you identify excerpts as being composed by humans?
    • Well-composed (36), Human-qualities (10);

()indicates number of occurrence of each theme

Overview / Prog Metal / Methodology / Results / Analysis & Findings

​

​

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

11

12 of 12

12

UKRI Centre for Doctoral Training (CDT) in Artificial Intelligence and Music (AIM)

This work is supported by the EPSRC UKRI Centre for Doctoral Training in Artificial Intelligence and Music (Grant no. EP/S022694/1) and by UKRI - Innovate UK (Project number 10102804).

\m/ Thank you for your attention \m/

@umpedronosapato

@jackjamesloth

{p.p.sarmento, j.j.loth}@qmul.ac.uk

​

🤘

12