A seminal study on data augmentation wins Test-of-Time honors from the International Society for Music Information Retrieval

Juan Pablo Bello — a professor in NYU Tandon’s Department of Computer Science and Engineering and Department of Electrical and Computer Engineering, faculty member at the Center for Urban Science + Progress (CUSP), and director of the NYU Steinhardt Music and Audio Research Lab (MARL) — has been named a runner-up in this year’s International Society for Music Information Retrieval (ISMIR) Test-of-Time Award for his influential 2015 paper, “A Software Framework for Musical Data Augmentation.”

Bello co-wrote the paper with lead author Brian McFee, a core member of MARL and associate professor at NYU’s Center for Data Science, and Eric Humphrey, then a MARL doctoral candidate. The team received the honor at the 26th annual ISMIR conference, held last month in Daejeon, Korea.

Music information retrieval — the science of teaching computers to understand and organize music — underpins many important tasks. From genre classification and automatic transcription to algorithmic song recommendations on platforms like Spotify, the field helps make sense of the world’s vast and growing collections of digital music.

In their pioneering paper, the authors tackle a central challenge: musical audio files are extraordinarily complex, and training machine-learning models to interpret them requires massive, carefully annotated datasets — a resource that’s often scarce. Their solution was to use data augmentation, a technique that enhances training data by introducing subtle, controlled variations, helping models learn to recognize patterns more robustly.

The team developed a general software framework for applying this approach to musical datasets, allowing researchers to easily expand training sets with musically meaningful perturbations of both audio and annotations, and their work has since become a cornerstone of contemporary music information research.

In announcing the Test-of-Time honorees, ISMIR described “A Software Framework for Musical Data Augmentation” as “a truly foundational paper,” noting that “it’s hard to imagine training a state-of-the-art audio model now without data augmentation.”

“It seems obvious in hindsight, but the beauty of Brian McFee’s vision at the time was in using domain-knowledge to understand how specific perturbations affect music labels,” Bello explains. “This enabled augmentation to be deployed across music datasets while going beyond label-preserving perturbations. I am honored that, more than a decade later, the community still finds these ideas useful and valuable.”