Comparing Voice and Stream Segmentation Algorithms

Nicolas Guiomard-Kagan (1, 2) , Mathieu Giraud (3, 2) , Richard Groult (1, 2) , Florence Levé (1, 3, 2)

1 MIS - Modélisation, Information et Systèmes - UR UPJV 4290
2 Algomus
3 CRIStAL - Centre de Recherche en Informatique, Signal et Automatique de Lille - UMR 9189

Voice and stream segmentation algorithms group notes from polyphonic data into relevant units, providing a better understanding of a musical score. Voice segmentation algorithms usually extract voices from the beginning to the end of the piece, whereas stream segmentation algorithms identify smaller segments. In both cases, the goal can be to obtain mostly monophonic units, but streams with poly-phonic data are also relevant. These algorithms usually cluster contiguous notes with close pitches. We propose an independent evaluation of four of these algorithms (Tem-perley, Chew and Wu, Ishigaki et al., and Rafailidis et al.) using several evaluation metrics. We benchmark the algorithms on a corpus containing the 48 fugues of Well-Tempered Clavier by J. S. Bach as well as 97 files of popular music containing actual polyphonic information. We discuss how to compare together voice and stream segmen-tation algorithms, and discuss their strengths and weaknesses.

Mots clés: stream segmentation; voice segmentation; music information retrieval; melody

Domaines: Musique, musicologie et arts de la scène; Son [cs.SD]; Algorithme et structure de données [cs.DS]

Nicolas Guiomard-Kagan, Mathieu Giraud, Richard Groult, Florence Levé. Comparing Voice and Stream Segmentation Algorithms. International Society for Music Information Retrieval Conference (ISMIR 2015), Oct 2015, Malaga, Spain. pp.493-499. ⟨hal-01246693⟩