Framework for automatic singing transcription: Database and Software.

Abstract

This Research data consists of a cross-annotated dataset of 1154seconds and a novel set of evaluation measures, able to report thetype of errors made by the system. Both the dataset, and a Matlabtoolbox including the presented evaluation measures, are freely available. In order to show the utility of the work presented in thispaper, we have performed an detailed comparative study of threestate-of-the-art singing transcribers plus a baseline method, lead-ing to relevant information about the performance of each method.