Spectral Modelling Synthesis of Vehicle Pass-by Noise

Spectral Modelling Synthesis (SMS) is a sound synthesis technique that models time-varying spectra of given sounds as a collection of sinusoids plus a filtered noise component. Although originally utilized to produce musical sounds, this technique can also be extended for analysis, transformation and synthesis of a wide range of environmental sounds, such as traffic noise. Simplifications based on psychoacoustic analysis can be conducted during the modelling process to avoid redundant data, which leads to perceptual similarity between synthesized sounds and the original recordings of vehicle pass-by noise. In this paper, we investigate if this perceptual similarity can be described by objective metrics, and how to improve the synthesis by tuning the parameters in the SMS algorithm. The results showed that vehicle pass-by sounds characterized by tyre and engine noise can be well synthesized with different parameter sets in the SMS algorithm. Furthermore, it is found that Zwicker Roughness is a sensitive metric for measuring the perceptual similarity between original recordings and synthesized sounds as it varies significantly when tuning SMS parameters.