Beyond a Western Center of Music Information Retrieval: A Bibliometric Analysis of the First 25 Years of ISMIR Authorship

2.1 ISMIR on ISMIR

Numerous ISMIR publications have focused on the work published in and around the conference. At the 10‑year mark of ISMIR, Downie et al. (2009) and Lee et al. (2009) provided perspectives on the community to date, highlighting growth in papers and author teams, the formation of the ISMIR society, and aspects of disciplines and paper topics. Even at this relatively early milestone of the community, Downie et al. (2009) issued a call to action for the field to ‘expand its musical horizons.’ Later, Sordo et al. (2015) analyzed the evolution of research groups and topics over 15 years, noting interactions between research group size, membership turnover, and breadth of topics. Hu et al. (2016) addressed the gender of ISMIR authors, reporting on extents of equal or unequal representation of women in the context of authorship and citations. ISMIR authors have also published tools for navigating the conference’s own proceedings. Grachten et al. (2009) introduced the ISMIR Cloud and ISMIRviewer applications for navigating early years of ISMIR proceedings, while the ISMIR Paper Explorer—first introduced by Stober et al. (2015) as a late‑breaking/demo and later published by Low et al. (2019)—allows users to browse the ‘locally restricted neighborhood’ of a query paper with pin, star, and history functions. While these resources have offered insights into the evolution of ISMIR research, few have directly addressed the field’s well‑known cultural or geographic biases.

2.2 Western‑centered MIR

Understanding the geographic and institutional distribution of ISMIR authorship is critical for several reasons. Research produced primarily in Global North contexts with Western‑centric methodologies, and evaluation metrics may fail to capture the nuances of diverse musical practices and audiences (Born, 2020; Holzapfel et al., 2018; Huang et al., 2023). In turn, this may result in limited diversity in datasets, models, and algorithms, which can then limit the generalizability of MIR findings and further exclude musical traditions and listening experiences from other cultures. For instance, algorithms trained on Western music often underperform when applied to underrepresented genres (Cano et al., 2021; Gómez‑Marín et al., 2024; Maia et al., 2022). The growing body of research on ethics in MIR (Holzapfel et al., 2018; Huang et al., 2021; Morreale et al., 2023) has also highlighted how Western‑dominated research can reinforce global inequalities in access to funding, infrastructure, and visibility. Researchers from the Global South may face systemic barriers to contributing to and benefiting from the field.

A clearer picture of the field’s composition may also support more equitable and sustainable international collaborations (Savage et al., 2023). A lens for examining these dynamics is the WEIRD framework—Western, Educated, Industrialized, Rich, and Democratic—originally developed in psychology to critique claims of universality based on narrow population samples (Henrich et al., 2010). This framework has since informed broader critiques of inclusivity across academic disciplines (Barrett, 2020; Jakubowski et al., 2025; Muthukrishna et al., 2020). Following Born (2020), applying this lens to MIR raises questions about who conducts the research, what music is studied, and whose interests the field serves. Scholarship across geography of science, science and technology studies, and decolonial theory offers richer conceptual tools for analyzing knowledge asymmetries. For example, Livingstone (2003) and Meusburger et al. (2010) emphasize how scientific knowledge is shaped by spatial and institutional geographies—who gets to produce knowledge, where, and under what conditions. Scholars such as Quijano (2000) and Mignolo (2011) argue that modernity is inseparable from coloniality—Western knowledge systems dominate not by chance, but through centuries of political, cultural, and epistemic imposition. Southern theory by Connell (2020) builds on this critique by documenting how social theory itself has been shaped by metropolitan centers, systematically ignoring or appropriating intellectual contributions from the Global South. Grosfoguel (2022) builds on this by calling for epistemic decolonization—challenging the categories, methods, and institutions that uphold global inequalities in knowledge production. In the current work, we argue that elucidating the geographic distribution of institutional affiliations offers a concrete way to engage with these questions.

2.3 Neighboring fields

Other fields have evaluated their authorship and the music studied, highlighting concerns about diversity, inclusion, and the importance of tracking representation over time. In music psychology, Jakubowski et al. (2025) analyzed journal publications from the past decade and found a strong reliance on WEIRD populations, even in studies focused on non‑Western contexts. This pattern matters for MIR, as psychological research underpins many music technologies (Gómez et al., 2024): Biased sampling in harmony, rhythm, perception, and emotion can directly shape the development of tools and models, reinforcing the need for broader diversity in research subjects and settings. Recent work in music cognition has highlighted concrete steps to improve representation—through cross‑cultural research factors (Jacoby et al., 2020), anti‑racism initiatives (Baker et al., 2020; Eagle et al., 2023), and inclusive mentorship (Agrawal et al., 2024)—addressing both the research content and community composition.

Neighboring audio and computing communities have undertaken similar self‑assessments, offering shared concerns and methods. DAFx has analyzed co‑authorship networks to understand collaboration patterns and contributor influence (Wilson, 2017), while Affective Computing has examined the composition of its conferences, journal IEEE Transactions of Affective Computing (TAFFC),⁵ and society Association for the Advancement of Affective Computing (AAAC)⁶ (Hupont et al., 2021, 2023). Geographic disparities have been reported in International Conference on New Interfaces for Musical Expression (NIME)⁷ (Avila et al., 2022) and RecSys (Porcaro et al., 2023), where a handful of countries dominate participation and leadership roles. Studies across International Conference on Auditory Display (ICAD),⁸ Audio Engineering Society (AES),⁹ and NIME have documented gender imbalances (Andreopoulou and Goudarzi, 2017; Buckingham and Ronan, 2019; Mathew et al., 2016; Xambó, 2018; Young et al., 2018). Similarly, artificial intelligence venues such as NeurIPS,¹⁰ RecSys,¹¹ and International Conference on Machine Learning (ICML)¹² have been analyzed for gender and geographical diversity using quantitative metrics like Shannon’s and Pielou’s indices (Freire et al., 2021). Several fields have also developed public tools to track trends in authorship, topics, and affiliations over time (Goode and Fasciani, 2022; Mauro et al., 2020). These efforts point to a growing cross‑disciplinary recognition of structural inequities.

3.1 Data collection

We implemented a web scraper to collect data from the ‘Conferences’ page of the ISMIR website,¹³ linking each paper to its corresponding entry on Zenodo.¹⁴ The data‑collection process consisted of the following steps:

1. Gather all available metadata from each Zenodo entry.

2. Download each available PDF and extract text from the first page, as Zenodo provides abstract metadata only from 2018 onward. We applied the same process to all papers to be consistent with our method across all conference years.

3. Use the OpenAI GPT‑4o‑mini API¹⁵ to analyze the extracted text, identifying abstracts, author affiliations, and the country and entity type of each institution (see Appendix 1 for the prompt).

4. Manually standardize institutional affiliations (e.g., converting ‘Music Technology Group, UPF’ to ‘Universitat Pompeu Fabra’).

5. Classify each institution’s country as a developed economy, economy in transition, or developing economy, following United Nations classifications.¹⁶

6. Manually validate the dataset to ensure consistency and accuracy. At least 14% of all entries were verified manually for errors in transcription of authors, titles, and abstracts, corresponding to the 2000–2005 period. The visualization tool also later helped identify other anomalous/missing entries, which were then manually corrected (7%).

In total, we obtained the following information for 2,458 published papers: authors, title, year, Digital Object Identifier, abstract, affiliations, institution names, institution types, and institution countries. While some early ISMIR papers featured dead links or embedded fonts that hindered reliable text extraction, we successfully retrieved usable titles and abstracts for 2,441 papers—covering more than 99% of the full corpus. For the remaining 17 papers with embedded fonts, we used optical character recognition from Apple’s Preview app to manually input the abstracts. For further analysis, we grouped research institutions into three main categories following the Research Organization Registry¹⁷ taxonomy: Education (academic institutions), Company, and Facility (e.g., research centers, governmental institutions, non‑profit associations, libraries). To foster reproducibility and transparency, ISMIR25Meta (in CSV format) and its analysis is available on GitHub under an MIT license.¹⁸ ISMIR25Viz is openly available to the research community as a website.¹⁹

3.2 Data analysis

We conducted the following analyses to investigate the geographical distribution of authorship and institutional affiliations.

Yearly affiliation: First, we examined the evolution of institutional affiliations within the research community by categorizing publications into single‑country and multi‑country collaborations. We quantified the proportion of papers originating from international collaborations and analyzed the most frequently contributing countries. Furthermore, we assessed whether these leading contributors belonged to the category of developed nations, as defined by the United Nations.

Temporal trends in diversity: Second, to measure the distribution of research output across countries, we calculated the Gini and Pielou indices yearly. The Gini index captures inequality in the number of research contributions per country (Dixon et al., 1987; Gini, 1921). The Pielou index quantifies how evenly research activity is distributed as a normalized version of Shannon entropy to measure diversity (Pielou, 1966). From these indices, we extracted linear trends to assess whether research output has become more equitably distributed over time. Then, we mapped fluctuations in these indices to identify specific years that exhibited significant shifts in diversity and concentration.

Collaboration networks: Finally, we constructed bipartite graphs that link authors with their affiliated countries to explore collaboration patterns. Using the Louvain community detection algorithm, we identified clusters of interconnected researchers and institutions (Blondel et al., 2008). This allows us to determine the extent of cross‑regional collaboration and assess whether certain research clusters include affiliations from the Global South, providing insights into the inclusivity of research networks.

3.3 Data visualization

In an effort to extend the work of the ISMIR explorer (Low et al., 2019), we used the collected data to generate a visualization of the publications (see Figure 1). We used the pre‑trained Sentence‑BERT module (Reimers and Gurevych, 2019) to obtain embeddings from all titles and abstracts and then used t‑stochastic neighbor embedding (t‑SNE) (Maaten and Hinton, 2008) and uniform manifold approximation projection (UMAP) (Healy and McInnes, 2024) for dimensionality reduction. Sentence‑BERT was applied to the entire abstract as a single input. Since Sentence‑BERT is designed to handle variable‑length text and uses mean pooling internally, it produces a fixed‑size embedding that captures the overall semantic content without needing to average individual sentence embeddings. Since we currently do not focus on a deep analysis of topic distribution like Sordo et al. (2015) or Low et al. (2019) do, we offer this as a tool for MIR researchers to find similar articles based on distance.

Figure 1

Figure 1 shows a screenshot of ISMIR25Viz, an interactive web‑based visualization system presenting available conference papers across four integrated views:

1. Paper embeddings: Papers are visualized using t‑SNE or UMAP dimensionality reduction applied to title and abstract embeddings, with interactive filtering by year, author, and research topics derived from n‑gram analysis of titles. We extracted bi‑grams and tri‑grams exclusively from titles, removing stop words and phrases, ultimately selecting the top 60 most frequent n‑grams to serve as research topics for readability. These n‑grams capture coherent topics and support embedding filtering, temporal evolution, and network construction through frequency and co‑occurrence patterns. Selecting a paper from the embedding plot displays full metadata and automatic translation of abstracts using Google Translate for widely spoken languages from the Global South.

2. Research evolution: Temporal trends of n‑grams are tracked as a timeline view. Along with the most frequent topics, we calculate recent trends through a linear regression on the yearly frequency and choose the n‑grams with highest slope.

3. Countries: A global collaboration network presents country‑level co‑authorship patterns with nodes sized by publication volume and edges weighted by collaboration frequency. Countries are filtered by minimum collaboration thresholds and colored by geographic region or United Nations development categories, enabling analysis of cross‑national research partnerships and identification of collaboration hubs within the global MIR community.

4. Authors: A bipartite network connecting authors to topics based on n‑gram co‑occurrence, an author collaboration network where edges represent co‑authorship relationships weighted by shared publications, and a topic co‑occurrence network displaying topics that frequently appear together in paper titles.

The tool supports literature reviews and facilitates exploration of the state of the art. In addition, it enables users to examine how specific research topics have evolved across institutions and regions worldwide.

4.1 Data and topics

The ISMIR25Meta dataset contains metadata for the 2,458 ISMIR papers published between 2000 and 2024, including titles, authors, Digital Object Identifiers, abstracts (when available), institutional affiliations, and geographic classifications. Our analysis of temporal trends in ISMIR research shows that the most frequently occurring title n‑grams include polyphonic music, genre classification, source separation, singing voice, and music similarity. The top trending topics according to the highest slopes in the linear regression fit include music generation, symbolic music, representation learning, source separation, and self‑supervised learning. Together, these trending topics account for a total of 135 papers, 78 of which were published in the past five years, indicating growing momentum. With respect to contributions from developing economies and economies in transition, we found that: (1) in Asia (313 papers), top topics include singing voice, music generation, symbolic music, and Carnatic music, reflecting both technical and culturally specific themes; (2) in Latin America (48 papers), frequent topics include genre classification, rhythmic pattern, and drum transcription; (3) Southeastern Europe is represented by a single paper on glass ceiling; and, (4) in the Commonwealth of Independent States and Georgia (3 papers), recurring topics are content separation and single tracks.

4.2 Western institutions dominate ISMIR

The first aspect that emerges from our analysis is the dominance of American institutions in ISMIR research (Figures 2 and 3). This is understandable when considering that the organization was incorporated in Canada on July 4, 2008, and three of the first four conferences were held in the United States while it was still in its early stage. Accounting for almost one‑quarter of all published papers, the United States is also the only country where ISMIR has been held more than twice (seven times in total). Following this, we find mostly European countries, with only Japan, China, Taiwan, Korea, India, and Brazil—incidentally, locations of past ISMIR conferences (Table 1)—representing other regions of the world. Using the United Nations country classification by regions, we find that Europe accounts for the largest number of affiliations (3,793), followed by Northern America (2,088), Asia (929), Developed Asia and the Pacific (659), Latin America (142), the Commonwealth of Independent States and Georgia (8), and Southeastern Europe (2).²⁰ As expected, most of the authorship is produced by what is considered Western affiliations (Euro‑American).

Figure 2

Figure 3

4.3 Underrepresented but emerging

Figure 2 provides a global perspective on the countries contributing most to ISMIR, as well as those with little to no representation. Africa remains the least‑represented continent, along with the Middle East, the Caribbean, and Central and South America. Similarly, institutions in Western and Southeast Asia are largely absent, with only a few exceptions. This disparity is further highlighted when categorizing contributions by developing versus developed economies.

When analyzing the countries of institutional affiliation, we find that 85.27% of papers reflect developed economies, 14.64% of papers reflect developing economies, and only the remaining 0.09% reflect economies in transition. Among developing and transition economies, the most active contributors include Taiwan (252 papers), China (207), South Korea (164), India (121), Brazil (113), Singapore (77), Hong Kong SAR (53), Israel (36), Malaysia (13), and Uruguay (13). Encouragingly, the number of papers from institutions in these countries has shown marked growth since 2012 (see Figure 4). Notably, ISMIR editions hosted online or in developing countries (e.g., Brazil and China) saw increased participation from first authors affiliated with these regions. Our analysis shows that 72.5% of Brazil’s publications and 74.2% of China’s publications in ISMIR occurred after the conference was hosted in each country, respectively, although the present data do not allow us to establish a causal relationship between ISMIR conference location and the participation of local institutions.

Figure 4

4.4 Broader collaboration patterns

In terms of research institutions, we may observe from the results shown in Figure 5 that ISMIR has been and is still an academically‑driven conference, with between 70%–80% of the papers in every edition coming from this kind of institution. Facilities seem to have maintained a constant presence over the course of the years, while industry papers have declined over time. Nevertheless, we observe a slow but steady increase in papers reflecting collaboration between different types of research organizations, arriving at almost 20% of papers published at ISMIR 2024.²¹ These numbers—perhaps highlighting the interest of the community in fostering collaboration between different entities—may indicate a positive trend in translating academic knowledge into practical applications or reflect new academia‑industry funding models.

Figure 5

Network analysis and Louvain community detection reveal that only a few research affiliations extend beyond single‑country boundaries. Following the United Nations Conference on Trade and Development (UNCTAD) (2018), we use Global South to include countries that belong to developing economies or economies in transition. Based on this Global North–South categorization, 17 countries (38.6%) were identified as part of the Global South. In terms of collaboration structure, 2,003 papers (81.5%) involved authors from a single country, while 452 papers (18.4%) included international collaborations. Among these, the majority were collaborations between developed countries (69.0%), followed by North–South collaborations (28.5%) and a small number of South–South collaborations (2.4%).

Network centrality measures reveal a concentration of connectivity and influence among institutions based in developed economies. The top 10 countries by degree centrality—indicating the number of direct collaborative ties—are all from the Global North, with the United States, United Kingdom, Germany, Spain, and Austria occupying the highest ranks. Similarly, betweenness centrality—reflecting a country’s role as a bridge in the collaboration network—further highlights the structural dominance of these countries. Brazil, as Global South’s highest‑ranking country, still keeps the 10th position in the world. Average centrality measures reveal a marked disparity in network positioning: Global North countries exhibit an average degree centrality of 0.250, more than twice that of Global South countries (0.115). This imbalance is even more evident in betweenness centrality (0.037 vs. 0.002), suggesting that institutions from the Global South play a comparatively limited role in connecting otherwise unlinked parts of the collaboration network.

Within the Global South, Asia emerged as the most active region, with 12 countries represented. This region also exhibited the highest intra‑regional collaboration (19 links), whereas Latin America accounted for only three such collaborations. Cross‑regional collaboration within the Global South was limited, with a single Asia–Latin America collaboration observed. The most central Global South countries, as determined by degree centrality, include Taiwan, China, Singapore, Hong Kong SAR, and South Korea. These countries serve as regional hubs but remain peripheral when compared to their Global North counterparts in terms of overall network influence. These findings suggest that the Global South is increasingly represented in ISMIR collaborations, though its institutional connectivity and structural positioning within the research network remain limited.

4.5 Progress made, challenges remain

To assess the distribution of research contributions worldwide, we computed the Gini and Pielou indices on a yearly basis. When restricting the analysis to only the 49 countries that have historically participated in the conference, the Gini index consistently fell between 0.7 and 0.9. However, to acknowledge that many countries have yet to contribute research to the field, we considered all countries listed in the ISO 3166 standard.²² As a result, our approach inherently pushes the Gini index closer to 1, reflecting a high degree of inequality. We conducted a linear regression to examine relationships between research diversity (Pielou index), or inequality (Gini index), and time. Figure 6 shows that the inequality has slowly decreased since the beginning of the conference, with a small but significant negative slope ($slope=-0.0004$, $R^2=0.18$, $t=-2.256$, $p=0.03$), while we see no significant linear change in Pielou coefficients over time ($slope=-0.0008$, $R^2=0.04$, $t=-0.979$, $p=0.337$). Despite modest changes in the Gini and Pielou indices, structural imbalance persists. Increased non‑Western participation may be offset by parallel growth elsewhere, keeping inequality measures relatively static.

Figure 6

Analyzing simultaneous local minima/maxima of these indices helps reveal periods of higher concentration (less diversity) and greater inclusivity (more diversity). Notably, years with higher research concentrations— indicating that contributions were dominated by a smaller subset of countries—align with conferences held in the United States (2001), Austria (2007), Portugal (2012), and online (2021). In the case of the United States (2024), data for 2025 are needed to assess if it was a local maxima for Gini. Conversely, years with increased diversity—where research contributions were more evenly distributed across countries—correspond to conferences held in Spain (2004), Taiwan (2014), Paris (2018), Canada–online (2020), and Italy–hybrid (2023). These findings could suggest that the geographical location of a conference may influence participation diversity, with certain regions fostering broader international involvement—likely influenced by the attractiveness or visa‑friendliness of the host country. However, the variation between local minima/maxima is very small, and a more detailed analysis is needed. Additionally, the hybrid format appears to have supported a more inclusive research environment—likely by lowering barriers to travel and funding—although for the fully online 2021 conference, diversity by this measure seems to decrease.

5.1 Design suggestions

Building on our analysis, we offer suggestions to support geographically diverse authorship and broader collaborative networks. While some of these recommendations are already in progress thanks to the ISMIR community, we find it important to present them together to illustrate their potential and impact.

5.2 Centering pluriversal knowledge systems in MIR

Beyond strictly geographical considerations, the ISMIR community can consider other facets of implicit representations underlying the field’s research. For example, CompMusic, while broadening cultural focus, focused primarily on certain ‘classical’ traditions, reflecting academia’s biases about which music is considered ‘worthy.’ In contrast, popular styles like funk carioca, cumbia, and reggaetón—rich in cultural meaning and social history (Sneed, 2008; Tamayo‑Duque, 2025)—remain underexplored due to perceptions of aesthetic legitimacy. This bias stems in part from prioritizing Western classical paradigms over traditional regional music, marginalizing both musical traditions and the scholarly imagination shaped by colonial hierarchies. Existing decolonial frameworks mentioned in § 2 advocate for centering pluriversal knowledge systems—epistemologies rooted in diverse local ontologies, lived experiences, and cultural practices, particularly those shaped by colonial histories.

In MIR, this shift toward more inclusive research requires more than expanding datasets or adding musical styles from the Global South. It demands a fundamental rethinking of the conceptual and methodological assumptions behind tasks, raising questions that are not just technical but epistemological: What does segmentation mean for musical traditions characterized by syncopation—how well do our algorithms handle non‑metrical or polyrhythmic music? Can music similarity systems trained on Western taxonomies account for non‑Western styles—are all cumbias identical? And, more fundamentally, what do we mean when we label ‘world music’ in a genre classifier—whose world, and according to whose classifications? These questions expose the cultural and ideological boundaries that MIR inherits and reproduces. Addressing them also means rethinking who is recognized as an expert—elevating the voices of local musicians and scholars working outside institutions. A pluriversal approach offers MIR a path beyond token inclusion toward a deeper transformation of its research priorities, collaborative practices, and technological visions. This shift is needed not only for epistemic justice but for building systems that reflect and engage with the world’s diverse musical cultures.

5.3 Limitations and future work

ISMIR25viz is subject to several limitations. (1) Large language models like ChatGPT were used only for metadata extraction and organization, not interpretation, but can still produce errors and hallucinations. (2) Using Sentence‑BERT on titles and abstracts alone oversimplifies the full content of academic papers. (3) Dimensionality‑reduction techniques (t‑SNE and UMAP) prioritize local over global structure, causing distortions on distances, variability across runs, and limiting reproducibility, which may affect grouping articles in the paper embeddings. Finally, (4) extracting n‑grams solely from paper titles may miss complex or emerging topics, potentially underrepresenting the research landscape.

Moreover, several critiques highlight the limitations of the WEIRD framework, as follows: (1) it oversimplifies the distinction between the West and the rest, neglecting regional complexities; (2) there is an apparent correlation between the Industrialized and Rich attributes to describe a society; and (3) the category of Democratic governance appears questionable, raising questions about whether it serves an analytical purpose or merely reinforces a provocative acronym.

A key limitation is using authors’ institutional affiliations as a proxy for geographic diversity. This approach does not account for authors’ countries of origin, cultural backgrounds, or transnational academic trajectories, providing an incomplete picture of who contributes to the MIR community and from where. Much could be gained from demographic information like gender, nationality, native languages, and their intersection. Thus, echoing past calls for action (Born, 2020; Hu et al., 2016), we identify the need to collect and understand additional information about authors, as well as non‑authors who participate through reviewership, conference attendance, or community initiatives (Young et al., 2018). Importantly, the collection and potential release of personal or sensitive data will require further discussion around ethical research conduct and privacy issues. Future research could explore author locations in greater detail, examining their relationship to authorship position (Andreopoulou and Goudarzi, 2017; Xambó, 2018), the impact of conference location on representation (Avila et al., 2022), and patterns of recurring authorship over multiple years (Andreopoulou and Goudarzi, 2017; Mauro et al., 2020; Sordo et al., 2015). Researchers, conference organizers, and society leadership can work together to identify best practices around requesting and handling this type of data in the context of community research. Future work can also consider whether different geographical or cultural designations can provide additional insights into authorship patterns (for example, addressing population imbalances such as that of the United States in comparison to single European countries). Finally, regarding the visualization tool itself, future user studies can clarify and improve usability aspects and the benefits of the tool to future ISMIR research.