MulTTiPop: A Multitrack Transcription Dataset for Pop Music

Summary

MulTTiPop is a new dataset introduced by Nathan Pruyne, Benjamin Stoler, and William Chen for AI music research, released on arXiv (2507.08753) in July 2025. It pairs pop music excerpts with multitrack MIDI recordings, enabling research in automatic music transcription at the multitrack and instrument-level level rather than just single-stem transcription. By providing aligned audio and per-instrument MIDI for pop songs, the dataset supports training and benchmarking models on tasks such as multitrack transcription, source separation, and music generation. This post on zhichai.net shares the paper's overview and arXiv link for researchers working in music AI and computational musicology.