Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Music for All: Representational Bias and Cross-Cultural Adaptability of Music Generation Models

Atharva Mehta, Shivam Chauhan, Amirbek Djanibekov, Atharva Kulkarni, Gus Xia, Monojit Choudhury

TL;DR

The paper systematically analyzes representational bias in music generation datasets, revealing that non-Western genres constitute only a small fraction of hours and exhibit disparate model performance across genres. It introduces a cross-cultural adaptation approach using parameter-efficient fine-tuning with Bottleneck Residual Adapters on two open-source models (MusicGen and Mustango) for Hindustani Classical and Turkish Makam, employing a novel arena-style Bloom's taxonomy evaluation. Results show that PEFT can improve generation quality for underrepresented genres, but effectiveness is highly model- and genre-dependent, underscoring the non-triviality of cross-cultural transfer. The work calls for more inclusive musical datasets and cross-cultural transfer learning baselines to prevent Western-centric homogenization of AI-generated music.

Abstract

The advent of Music-Language Models has greatly enhanced the automatic music generation capability of AI systems, but they are also limited in their coverage of the musical genres and cultures of the world. We present a study of the datasets and research papers for music generation and quantify the bias and under-representation of genres. We find that only 5.7% of the total hours of existing music datasets come from non-Western genres, which naturally leads to disparate performance of the models across genres. We then investigate the efficacy of Parameter-Efficient Fine-Tuning (PEFT) techniques in mitigating this bias. Our experiments with two popular models -- MusicGen and Mustango, for two underrepresented non-Western music traditions -- Hindustani Classical and Turkish Makam music, highlight the promises as well as the non-triviality of cross-genre adaptation of music through small datasets, implying the need for more equitable baseline music-language models that are designed for cross-cultural transfer learning.

Music for All: Representational Bias and Cross-Cultural Adaptability of Music Generation Models

TL;DR

The paper systematically analyzes representational bias in music generation datasets, revealing that non-Western genres constitute only a small fraction of hours and exhibit disparate model performance across genres. It introduces a cross-cultural adaptation approach using parameter-efficient fine-tuning with Bottleneck Residual Adapters on two open-source models (MusicGen and Mustango) for Hindustani Classical and Turkish Makam, employing a novel arena-style Bloom's taxonomy evaluation. Results show that PEFT can improve generation quality for underrepresented genres, but effectiveness is highly model- and genre-dependent, underscoring the non-triviality of cross-cultural transfer. The work calls for more inclusive musical datasets and cross-cultural transfer learning baselines to prevent Western-centric homogenization of AI-generated music.

Abstract

The advent of Music-Language Models has greatly enhanced the automatic music generation capability of AI systems, but they are also limited in their coverage of the musical genres and cultures of the world. We present a study of the datasets and research papers for music generation and quantify the bias and under-representation of genres. We find that only 5.7% of the total hours of existing music datasets come from non-Western genres, which naturally leads to disparate performance of the models across genres. We then investigate the efficacy of Parameter-Efficient Fine-Tuning (PEFT) techniques in mitigating this bias. Our experiments with two popular models -- MusicGen and Mustango, for two underrepresented non-Western music traditions -- Hindustani Classical and Turkish Makam music, highlight the promises as well as the non-triviality of cross-genre adaptation of music through small datasets, implying the need for more equitable baseline music-language models that are designed for cross-cultural transfer learning.

Paper Structure

This paper contains 32 sections, 7 equations, 5 figures, 8 tables.

Table of Contents

  1. Introduction
  2. The Disparity in Music Generation Research
  3. Data Collection
  4. Dataset Papers
  5. Findings
  6. Genre Adaptation: Data, Models and Experiments
  7. Dataset Creation
  8. Prompt Generation
  9. Models
  10. MusicGen
  11. Mustango
  12. Training settings
  13. Results
  14. Automatic Metrics
  15. Human Evaluation
  16. ...and 17 more sections

Figures (5)

  • Figure 1: The bottom left piechart shows the global distribution of genre. Each piechart in the map shows the distribution of genres in different regions with the size of each piechart being proportional to their contribution to the data corpus.
  • Figure 2: Mustango & MusicGen settings for low-resource fine-tuning.
  • Figure 3: Inter Annotator Agreement metrics for Hindustani Classical & Turkish Makam.
  • Figure 4: Scaled ELO ratings for each model in Hindustani Classical and Turkish Makam music. Categories include OA: Overall Assessment, Inst.: Instrument Accuracy, MC: Melody Capture, RC: Rhythm Capture, and CR: Creativity. Ratings are provided for all query types and individual query categories.
  • Figure 5: Screenshots of Label Studio, annotation tool for Inter Annotator Agreement and ELO rating comparison