Musical offset detection of pitched instruments: The case of violin
Journal
Proceedings of the 16th International Society for Music Information Retrieval Conference, ISMIR 2015
ISBN
9788460688532
Date Issued
2015-01-01
Author(s)
Abstract
Musical offset detection is an integral part of a music signal processing system that requires complete characterization of note events. However, unlike onset detection, offset detection has seldom been the subject of an in-depth study in the music information retrieval community, possibly because of the ambiguity involved in the determination of offset times in music. This paper presents a preliminary study aiming at discussing ways to annotate and to evaluate offset times for pitched non-percussive instruments. Moreover, we conduct a case study of offset detection in violin recordings by evaluating a number of energy, spectral flux, and pitch based methods using a new dataset covering 6 different violin playing techniques. The new dataset, which is going to be shared with the research community, consists of 63 violin recordings that are thoroughly annotated based on perceptual loudness and note transition. The offset detection methods, which are adapted from well-known methods for onset detection, are evaluated using an onset-aware method we propose for this task. Result shows that the accuracy of offset detection is highly dependent on the playing techniques involved. Moreover, pitch-based methods can better get rid of the soft-decaying behavior of offsets and achieve the best result among others.
Type
conference paper