Estimating Onset and Offset Asynchronies in Polyphonic Score-Audio Alignment

Audio-to-Score Alignment Using Deep Automatic Music Transcription

“…We stopped the agglomerative procedure when a certain minimum distance t among clusters was reached. We have randomly chosen such threshold in [0.03, 0.07] seconds, representing broad interval around 0.05 seconds that is assumed as upper-bound of usual chord asynchronies [3]. Subsequently, we set the onsets of the notes in each cluster equal to the average onset time of that cluster so that the final misaligned note sequence contains chords made by notes having the same onset.…”

“…There is a limited number of works that have faced the problem of note-level alignment, mainly with the objective of music performance analysis. Indeed, it is known that subtle asynchronies are generated during a human performance: notes in the same chord are written in musical scores as events having the same onset -and possibly the same offset -, but music players always introduce asynchronies of less than 0.05 seconds among the timings of such notes [3]. Other discrepancies between score and performance note order are related to the phrasing and articulation practices; for instance, the legato articulation consists in a slight overlap between two successive notes, even if in the musical score they are notated with no overlap.…”

“…These almost imperceptible timing effects are considered to be responsible of the incredibly various expressiveness of music performances and are consequently of crucial importance in music performance analysis studies [4]. Methods used for note-level alignment so far include HMM, DTW, NMF and blob recognition in spectrograms [3], [5]- [8] The rise of Artificial Neural Networks in their Deep Learning (DL) paradigm has led several researchers to exploit DL models for feature learning tailored to DTW. Two methods are particularly noteworthy: one [9] employs Siamese Networks for learning features that can be used for some distance function in DTW; the second method relies on the improvements made in the field of AMT for converting the sequences to a common space -the space of the symbolic notation [10].…”

Encoding music performance data in Humdrum and MEI

“…AMPACT [8] is a suite of tools for estimating performance data in monophonic or polyphonic contexts for which a symbolic representation of the musical content is also available. In situations where a full musical score is available, AMPACT uses a score-audio alignment method that identifies distinct onsets and duration for all notes, including notated simultaneities, such as chords, in polyphonic contexts [9].…”

Blend in Singing Ensemble Performance: Vibrato Production in a Vocal Quartet

“…In the simultaneous recording of several voices, even using individual microphones situated close to the mouth, a certain amount of bleed from other voices is inevitable, and is likely to interfere with robust analysis of the individual voices. Various techniques have been employed to account for this issue and allow assessment of the ensemble singing voice, including polyphonic acoustic analysis toolkits such as AMPACT for MATLAB [5].…”