Estimating Onset and Offset Asynchronies in Polyphonic Score-Audio Alignment
Johanna Devaney · Journal of New Music Research · 2014
In order to use score-audio alignment to study expressive performance or guide signal processing algorithms, it is necessary to identify the onsets and offsets of all of the notes in sounded simultaneities (e.g. chords). This paper describes an algorithm to improve the accuracy of dynamic time warping-based score-audio alignment for monaural polyphonic recordings. The algorithm uses a multi-pass approach, where an initial dynamic time warping alignment is refined by a hidden Markov model to allow the identification of asynchronies between musical events that are notated as simultaneities in the score. By providing estimates of individual onsets and offsets in notated simultaneities, the multi-pass algorithm improves the median accuracy of the DTW alignment for individual onsets by 41 ms on average, from 118 to 77 ms, and for individual offsets by 6 ms on average, from 75 to 69 ms.