Synthesized Polyphonic Music Database with Verifiable Ground Truth for Multiple F0 Estimation

To study and to evaluate a multiple F0 estimation algorithm, a polyphonic database with verifiable ground truth is necessary. Real recordings with manual annotation as ground truth are often used for evaluation. However, ambiguities arise during manual annotation, which are often set up by subjective judgements. Therefore, in order to have access to verifiable ground truth, we propose a systematic method for creating a polyphonic music database. Multiple monophonic tracks are rendered from a given MIDI file, in which rendered samples are separated to prevent overlaps and to facilitate automatic annotation. F0s can then be reliably extracted as ground truth, which are stored using SDIF.