Resolving octave ambiguities: A cross-dataset investigation
Journal
Proceedings - 40th International Computer Music Conference, ICMC 2014 and 11th Sound and Music Computing Conference, SMC 2014 - Music Technology Meets Philosophy: From Digital Echos to Virtual Ethos
ISBN
9789604661374
Date Issued
2014-01-01
Author(s)
Abstract
Octave dual-tone is one of the most difficult patterns to identify in multipitch estimation (MPE), as the spectrum of the upper note is almost masked by the lower one. This paper investigates the potential for a supervised binary classification framework to address this issue, and whether such a framework is adequate for diverse real-world signals. To this end, a new dataset comprising of 3,493 real single notes and octaves recorded by two pianists and guitarists are constructed to facilitate an in-depth analysis of this problem. The dataset is available to the research community. Performance of synthetic and real-world octave dualtones using various spectral-, cepstral- and phase-based features are studied systematically. Our experiments show that the instantaneous frequency deviation (IFD) represents the most reliable feature representation in discriminating octave dual-tones from single notes. Based on this new dataset and the RWC dataset, we present a series of experiments to offer insights into the performance difference between synthetic and real octaves, piano and guitar notes, as well as studio recordings and home recordings. As the proposed method holds the promise of resolving octave dualtone, we envision that it can be an important module of a multipitch estimation system.
Type
conference paper