Abstract
Current music generation models exhibit a profound Western-centric bias, with non-Western traditions accounting for less than 6% of training data. This study quantifies this representational gap and evaluates the effectiveness of Parameter-Efficient Fine-Tuning (PEFT) in adapting state-of-the-art models like MusicGen and Mustango to Hindustani Classical and Turkish Makam music. Our findings reveal that while lightweight adapters can improve cultural adherence, architectural differences significantly impact a model's capacity for cross-cultural transfer.
Introduction
While AI music generation has reached impressive milestones, its progress is unevenly distributed. A staggering 94% of training data in leading models is sourced from Western traditions. This paper explores the "representational gap" and investigates whether lightweight fine-tuning can bridge this divide for underrepresented genres like Hindustani Classical and Turkish Makam music.
The Representational Gap
Music-Language Models (MLMs) suffer from a systemic bias. Most are trained on datasets where non-Western music is a marginal presence. This is not merely a social issue but a technical one: non-Western traditions often utilize entirely different "grammars," such as the complex melodic frameworks of ragas or the intricate rhythmic cycles of talas and usuls.
When a model trained predominantly on Western pop attempts to generate an Indian classical piece, it often forces the output into Western tonal and rhythmic structures. This results in "musical hallucinations" that sound culturally inauthentic to trained ears.
Methodology
To address this, researchers employed Parameter-Efficient Fine-Tuning (PEFT) using Bottleneck Residual Adapters. Instead of retraining billions of parameters, they inserted small, trainable modules that account for only about 0.1% of the total parameter count. They tested this on two models:
- MusicGen: A transformer-based model.
- Mustango: A model integrating diffusion and transformer architectures.
Evaluation moved beyond standard automated metrics to a novel arena-style human evaluation framework based on Bloom’s Taxonomy. Experts ranked clips based on Recall (reproducing known elements), Analysis (forming novel combinations), and Creativity (blending genres).
Data and Meta-Analysis
The authors conducted a meta-analysis of 152 papers and over 1 million hours of music data to quantify global bias. For adaptation, they utilized the Dunya corpus:
- Hindustani Classical Music: ~23 hours from the MTG Saraga dataset.
- Turkish Makam Music: ~121 hours.
These datasets are critical because they include rich metadata for melodic lines, rhythmic patterns, and specific instrumentation.
Experimental Results
Fine-tuning yielded significant improvements, particularly for the diffusion-based Mustango model, which saw an 8% improvement in ELO ratings for Hindustani Classical and 4% for Turkish Makam. Fine-tuned models also achieved lower Fréchet Audio Distance (FAD) scores, indicating audio statistically closer to the target cultural distribution.
Interestingly, MusicGen's performance was less consistent; for Turkish Makam, the fine-tuned version actually regressed in some analytical tasks, suggesting that simple adapters are not a universal solution for all architectures.
Discussion: Insights and Trade-offs
- Architecture and Adaptability: Diffusion-based models like Mustango appear more flexible for cross-cultural transfer than pure transformer models.
- Knowledge Attrition: A trade-off was observed where specialization in a new genre (improving Recall) sometimes hindered a model's "Creativity" or ability to generalize.
- Subjectivity: Human annotators agreed on objective elements (e.g., instrument identification) but varied significantly on subjective qualities like creativity, highlighting the cultural dependency of musical quality.
Conclusion
This study highlights the urgent need for more diverse training data and specialized fine-tuning techniques to create truly global music AI. Future work will explore more advanced PEFT techniques like LoRA and attempt to bridge the gap for even lower-resource traditions from Africa and Southeast Asia, which remain largely absent from the current AI landscape.
Summary updated