Transient and Steady-state Component Extraction Using Nonlinear Filtering

Recently, a fast and simple method for separation of the transient and steady-state components of a music signal was proposed in [4]. The technique involves the application of two median filters to the spectrogram: one along the frequency bins to eliminate steady-state components, and the other along the time frames to eliminate transient components. In the present paper, a modification of the filtering stage is proposed; the resulting algorithm is evaluated both regarding the perceptual quality of the attained separation and its use as a preprocessing stage for improving the performance of a beat-tracking technique. The results obtained for a reference data set of beat-annotated music excerpts are promising.