MulTTiPop Releases a Multitrack MIDI Dataset for Evaluating Pop-Music Transcription

arXiv·low signal

MulTTiPop is a dataset of pop-music segments paired with multitrack MIDI recordings for evaluating automatic music-transcription models. It contains 572 segments totaling 3.5 hours of audio spanning genres and decades from the 1930s to the 2000s, built by metadata-matching song segments from the Lakh MIDI and TheoryTab datasets and manually identifying an anchor beat to align audio and MIDI. A narrow but concrete new evaluation resource for the music-AI/audio community.