arxiv arXiv cs.LG · 1h ago · research

MulTTiPop: multitrack transcription dataset for pop music

from English

The authors present MulTTiPop, a dataset of pop music segments and their associated multitrack MIDI recordings designed for evaluating automatic music transcription models.

The dataset contains 572 segments totaling 3.5 hours of audio from the 1930s to the 2000s.
Data collection involved metadata-based matching on Lakh MIDI and TheoryTab, anchor beat identification, and tempo warping.
Evaluation shows substantial room for improvement, with the best model achieving only 38% Onset F1.

Importance 1/3 arXiv cs.LG Evaluation & benchmarks Training data

Read original