Perceptually-Based Evaluation of the Errors Usually Made When Automatically Transcribing Music

This paper investigates the perceptual importance of typical errors occurring when transcribing polyphonic music excerpts into a symbolic form. The case of the automatic transcription of piano music is taken as the target application and two subjective tests are designed. The main test aims at understanding how human subjects rank typical transcription errors such as note insertion, deletion or replacement, note doubling, incorrect note onset or duration, and so forth. The Bradley-Terry-Luce (BTL) analysis framework is used and the results show that pitch errors are more clearly perceived than incorrect loudness estimations or temporal deviations from the original recording. A second test presents a first attempt to include this information in more perceptually motivated measures for evaluating transcription systems.