ASA 127th Meeting M.I.T. 1994 June 6-10

4aMU5. Derivation of a synthesis model from time-variant spectral analysis.

James Beauchamp

School of Music and Dept. of Elec. and Comput. Eng., 2136 Music Bldg., 1114 W. Nevada, Univ. of Illinois at Urbana--Champaign, Urbana, IL 61801

Andrew Horner

Hong Kong Univ. of Sci. and Technol., Clear Water Bay, Kowloon, Hong Kong

An effective spectrum-based musical instrument synthesis model requires that time-variant spectrum analysis (TVSA) be performed on a large group of tones representing the entire pitch and dynamic range of the instrument. Once enough common properties in the TVSA results are identified, a synthesis model can be devised, implemented, and tested. When the synthesis model is sufficiently complete, it can be employed in music performance or timbral research. It is important that a synthesis model's physical parameters strongly correlate with perceived timbral features. Our current model employs three time-variant parameters: rms amplitude, spectral centroid, and fundamental frequency, which strongly correlate with loudness, brightness, and pitch. In addition, a centroid-controlled transfer function, preconstructed by averaging TVSA data from the original set of analyzed tones, is used to generate the detailed time-variant spectra. When the three time-variant parameters are matched to those measured in particular instrument tones, global contours and micro-details emerge that contribute to synthesis realism. Small perceptual differences between original and synthetic sounds, which some listeners perceive as ``roughness,'' may be caused by smoothing of asynchronous spectral fluctuations in the synthesis. Current research is attempting to data-reduce both the time-variant control parameters and the centroid-controlled transfer function without sacrificing quality.