Editorial for TISMIR Special Collection: Cultural Diversity in MIR Research

In this special collection of papers on Cultural Diversity in MIR Research, we present five articles that also illustrate the diversity in research focus under this broad theme. Below we provide a summary of each of the articles.

One starting point of improving cultural diversity in MIR research is to diversify the music corpora in machine-readable formats. Work of this kind takes a tremendous amount of effort but can have a lasting impact in our community. In “Community Based Music Information Retrieval: A Case Study of Digitizing Historical Klezmer Manuscripts from Kyiv”, Malin et al. present the Kiselgof-Makonovetsky Digital Manuscript Project (KMDMP), a community-based project to digitize historical handwritten music manuscripts held by the Vernadsky National Library of Ukraine. This collection contains a total of around 1,300 melodies, combining typical Jewish dance and non-dance genres, European dance music, and songs and liturgical chants. In addition to presenting the challenges and methods of this digitization project, the paper also presents its implications for MIR and computational ethnomusicology and reflections on the project in the larger context of MIR and diversity.

Another contribution to the diversification of data is provided in “A Dataset of Norwegian Hardanger Fiddle Recordings with Precise Annotation of Note and Beat Onsets” by Lartillot et al. The authors address the specificities of Hardanger fiddle music, which is a folk music tradition from the western and central part of southern Norway. In 18 recordings of fiddle pieces, note onsets and offsets, and beat positions are precisely annotated. The methodological approach and the newly developed annotation software have the potential to be applied for the annotation of recordings from other musical traditions. The beat tracking is challenging due to the metric complexities of the style. The authors show that existing beat tracking methods fail to reach even moderate success on this data set. For the annotation of beats, a new approach is proposed in which those notes that have their onset on the beat are annotated. The resulting data set is a valuable contribution to MIR as well as to ethnomusicology. It provides new and challenging ground truth for various MIR tasks, and it allows study of the peculiarities of this intriguing folk music style.

Along the same line of addressing the data scarcity issue in MIR tasks for less investigated music genres, in “Four-way Classification of Tabla Strokes with Transfer Learning Using Western Drums”, Ananthanarayana et al. explore the idea of transfer learning for tabla stroke transcription. Specifically, they pre-train a model on similarly sounding Western drums and fine tune it on tabla data for the transcription of four categories of tabla strokes from audio recordings. Experiments show that while fine-tuning significantly improves transcription accuracy over pre-trained models as expected, it only surpasses the model trained on tabla data from scratch on one of the four investigated stroke categories. This result shows the challenge of transfer learning from culturally different music data.

The superiority of deep learning models over heuristics-based algorithms in MIR tasks, such as predominant melody extraction from polyphonic audio, has been amply demonstrated in recent years. As is also well known, such data-driven approaches are prone to performance degradation arising from domain drift or the mismatch between training and test data. In “Repertoire-Specific Vocal Pitch Data Generation for Improved Melodic Analysis of Carnatic Music”, Plaja-Roglans et al. propose a methodology for the synthesis of realistic training data that corresponds exactly with provided melodic pitch annotations, where the latter annotations are previously estimated from cleaned up close-mic recordings available for vocal Carnatic concerts. On Carnatic music test data, the resulting trained state-of-the-art model is shown to outperform a similar model trained with Western music data, thus providing a valuable methodology for similar resource-constrained repertoires and MIR tasks.

Collecting underrepresented music corpora and framing them as MIR tasks is a prevalent strategy to promote cultural diversity within the MIR community. However, in “Beyond Diverse Datasets: Responsible MIR, Interdisciplinarity, and the Fractured Worlds of Music”, Huang et al. call for a more profound philosophical transformation in MIR beyond the current practice. This transformation entails addressing epistemological, ontological, methodological, and axiological considerations. The authors illustrate this perspective through two insightful case studies. Firstly, they explore the ethical implications of generative music AI and emphasize the importance of integrating axiological dimensions to encourage a comprehensive examination of music ecosystems within the MIR community. Secondly, they re-contextualize the study of Irish traditional music, demonstrating an agonistic interdisciplinarity approach that promotes responsible engineering within the realm of traditional music.