Editorial for TISMIR Special Collection: Cultural Diversity in MIR Research

Zhiyao Duan, Peter van Kranenburg, Juhan Nam, Preeti RaoTrans. Int. Soc. Music. Inf. Retr 2023 · DOI

Title card for the TISMIR Special Collection on Cultural Diversity in MIR Research.

Introduction: Breaking the Western-Centric Bubble

For decades, the field of Music Information Retrieval (MIR) has operated largely within a Western-centric bubble, focusing predominantly on Western popular and classical music. This bias is more than just a matter of representation; it creates significant technical "blind spots." Algorithms trained on Western scales, rhythms, and instruments often fail when applied to the diverse musical traditions of the world. This editorial introduces a special collection of research aimed at dismantling these biases by diversifying datasets and rethinking the philosophical frameworks that underpin music AI.

The core challenge is twofold: the scarcity of machine-readable data for non-Western genres and the need for a "philosophical transformation" in how researchers approach interdisciplinarity and ethics. Music cannot be treated as a mere signal in a vacuum; both the music itself and the "enculturated listener" are shaped by specific cultural contexts. To address this, the collection advocates for "Responsible MIR," a framework that respects the nuances of global music traditions through collaboration with ethnomusicologists and critical theory.

Methodological Innovations in Global MIR

This special collection synthesizes methodologies from five featured studies, each tackling a different aspect of cultural diversity in MIR:

  • Community-Based Digitization: Researchers utilized a community-driven approach to transform historical, handwritten manuscripts of Klezmer music into machine-readable formats. This ensures that the digitization process is grounded in the community's own understanding of its heritage.
  • Precision Annotation for Folk Traditions: New software and methodological frameworks were developed to annotate complex metrics and note onsets in Norwegian folk traditions, specifically where standard Western-trained beat-trackers fail.
  • Evaluating Transfer Learning: The limits of "pre-train and fine-tune" strategies were tested by applying models trained on Western drums to Indian Tabla strokes, revealing the depth of cultural differences in sound.
  • Repertoire-Specific Synthesis: To bypass the need for massive amounts of manually labeled "clean" data, researchers created a methodology to synthesize realistic training data for Carnatic vocal pitch analysis, tailored to specific melodic nuances.
  • Critical Theory and Ethics: By applying epistemological and axiological frameworks, authors examined the ethical implications of generative AI on traditional music ecosystems, such as Irish traditional music.

Specialized Datasets and Technical Benchmarks

The collection introduces several specialized datasets that provide a foundation for future research in computational ethnomusicology:

  • Kiselgof-Makonovetsky Digital Manuscript Project (KMDMP): Includes approximately 1,300 melodies covering Jewish dance genres, liturgical chants, and European dance music.
  • Hardanger Fiddle Dataset: 18 recordings of Norwegian folk music with precise, manually annotated note onsets, offsets, and beat positions.
  • Tabla Stroke Dataset: Audio recordings categorized into four specific stroke types for transcription tasks.
  • Carnatic Vocal Pitch Data: A dataset generated through synthesis specifically for the melodic nuances of Carnatic music.

Key Insights: The "Transfer Learning Wall" and Domain Drift

One of the most striking insights from this collection is the identification of the "Transfer Learning Wall." Researchers found that transfer learning is not a universal fix for data scarcity. Models pre-trained on Western drums and fine-tuned on Tabla only outperformed "from-scratch" models in one out of four categories. This proves that cultural differences in sound are deep and non-trivial; a model's understanding of "rhythm" or "timbre" is often inextricably linked to the data it was raised on.

Furthermore, "domain drift" is a significant hurdle. Standard MIR tools, like state-of-the-art beat trackers, failed to reach even moderate success on the Norwegian Hardanger fiddle dataset due to its metric complexities. This suggests that many "universal" models in MIR are actually "Western-specific" models in disguise. Conversely, the collection shows that repertoire-specific data synthesis can outperform models trained on massive but culturally irrelevant datasets, setting new benchmarks for genres like Carnatic music.

Toward Responsible MIR

The authors advocate for a shift toward "Responsible MIR," which moves beyond simply "adding more data." This involves "agonistic interdisciplinarity"—working closely with ethnomusicologists to ensure technical solutions are culturally grounded—and considering the ethical impact of AI on traditional music ecosystems.

While this collection is a significant step forward, the authors acknowledge it is only a starting point. Thoroughly addressing cultural diversity across the entire MIR pipeline remains a massive undertaking. Future work must focus on scaling these philosophical frameworks and deepening the collaboration between engineers and cultural experts to ensure that the future of music AI is truly global and inclusive.

Summary updated

BibTeX
@article{duan2023editorial,
  author = {Zhiyao Duan and Peter van Kranenburg and Juhan Nam and Preeti Rao},
  title = {{Editorial for TISMIR Special Collection: Cultural Diversity in MIR Research}},
  journal = {Trans. Int. Soc. Music. Inf. Retr},
  year = {2023},
  doi = {10.5334/tismir.179},
}
Figure