A Metric for Music Notation Transcription Accuracy

Automatic music transcription aims at transcribing musical performances into music notation. However, most existing transcription systems only focus on parametric transcription, i.e., they output a symbolic representation in absolute terms, showing frequency and absolute time (e.g., a pianoroll representation), but not in musical terms, with spelling distinctions (e.g., A[ versus G]) and quantized meter. Recent attempts at producing full music notation output have been hindered by the lack of an objective metric to measure the adherence of the results to the ground truth music score, and had to rely on time-consuming human evaluation by music theorists. In this paper, we propose an edit distance, similar to the Levenshtein Distance used for measuring the difference between two sequences, typically strings of characters. The metric treats a music score as a sequence of sets of musical objects, ordered by their onsets. The metric reports the differences between two music scores based on twelve aspects: barlines, clefs, key signatures, time signatures, notes, note spelling, note durations, stem directions, groupings, rests, rest duration, and staff assignment. We also apply a linear regression model to the metric in order to predict human evaluations on a dataset of short music excerpts automatically transcribed into music notation.