The filter diagonalisation method for music signal analysis: frame-wise vibrato detection and estimation

Abstract

We present a novel approach to frame-wise vibrato detection and estimation in music signals using the Filter Diagonalisation Method (FDM). In contrast to conventional fast Fourier transform-based methods, the FDM's output remains robust over short time frames, allowing frame sizes to be set at values small enough for accurately identifying local vibrato characteristics and pinpointing vibrato boundaries. FDM decomposes the local fundamental frequency into sinusoids and returns their frequencies and amplitudes, which the system uses to determine vibrato presence and vibrato parameter values. We test two decision mechanisms – the decision tree and Bayes' Rule – for vibrato detection. The systems are tested against state-of-the-art techniques on monophonic datasets consisting of string, woodwind, brass, and voice excerpts. In addition to using existing datasets, we have created a new monophonic dataset consisting of performances of an entire music piece on erhu and violin, with annotations of vibrato presence and parameters. We show that FDM-based techniques consistently yield the best results in both frame-level and note-level evaluations. Furthermore, FDM with Bayes' Rule leads to better F-measure results – 0.84 (frame-level), 0.41 (note-level) – than FDM with decision tree – 0.80 (frame-level), 0.31 (note-level). FDM's accuracy for determining vibrato rates is above , and for vibrato extents is about.