Reproduce “Twenty-Five Years of MIR Research: Achievements, Practices, Evaluations, and Future Challenges” on OpenResearch

About this paper

Geoffroy Peeters, Zafar Rafii, Magdalena Fuentes, Zhiyao Duan, Emmanouil Benetos, Juhan Nam, Yuki Mitsufuji

Geoffroy Peeters, Zafar Rafii, and coauthors trace the development of Music Information Retrieval (MIR) from early content-based retrieval and handcrafted signal features to deep, self-supervised, multimodal, and generative models. The review also assesses benchmarking, open science, industrial adoption, representation gaps, environmental costs, and legal questions, arguing that future progress requires broader evaluation and more culturally inclusive data and practices.

Key insights

• MIR has shifted from knowledge-driven pipelines toward data-driven and end-to-end systems, while the boundaries between analysis, processing, and generation are increasingly blurred by foundation models.
• Shared infrastructure such as MIREX, HEAR, and MARBLE, together with tools including librosa, Essentia, mir_eval, and music21, has helped establish common practices, although fixed benchmarks and restricted datasets can constrain progress.
• The field’s technical advances coexist with uneven performance across genres and cultures, underrepresentation of non-Western music, substantial computational demands, and unresolved questions about generated music and training-data rights.

Summary by alphaXiv.