Improved modelling of attack transients in music analysis-resynthesis
Paul Masri, ANDREW BATEMAN · Bristol Research (University of Bristol) · 1996
Current music analysis-resynthesis models represent sounds through a set of features, which are extracted from a time-frequency representation. So that each time-frame can present a good approximation to the instantaneous spectrum, it is necessary to analyse the waveform in short segments. This is achieved with a window function whose position is advanced by a fixed amount between frames. When the window encompasses a transient event, such as the percussive onset of a note, it contains information both before and after the event. These partiallycorrelated spectra often become confused during analysis and cause audible ‘diffusion’ upon resynthesis. This paper presents a simple, novel technique to avoid the problem, by synchronising the analysis window to transient events. Event locations are identified by observing short-term changes in the spectrum. Thereafter the position of the analysis window is constrained, to prevent it capturing the signal both sides of an event simultaneously. This method, which has been automated, yields an improvement that is clearly audible, particularly for percussive sounds which retain their ‘crispness’.