Source-filter modeling in sinusoid domain

摘要

This paper presents the theory and implementation of source-filter modelling in sinusoid domain and its applications on timbre processing. The technique decomposes the instantaneous amplitude in a sinusoid model into a source part and a filter part, each capturing a different aspect of the timbral property. We show that the sinusoid domain sourcefilter modelling is approximately equivalent to its time or frequency domain counterparts. Two methods are proposed for the evaluation of the source and filter, including a least-square method based on the assumption of slow variation of source and filter in time, and a filter bank method the models the global spectral envelope in the filter. Tests show the effectiveness of the algorithms for isolation frequency-driven amplitude variations. Example applications are given to demonstrate the use of the technique for timbre processing.