Subjective and Objective Evaluation of Procedurally-Generated Audio for Soft-Body Interactions

Abstract

Procedurally-generated audio has proven to be an effective solution to synthesize complex sound phenomena such as soft-body interactions in computer animations and games. However, the quality of synthesized audio varies depending on the kind of methods or parameters chosen. As a consequence, it is often necessary to constantly evaluate the output sound quality as it is produced, which can be a difficult task. In this paper, we address this issue by taking both subjective and objective approaches, and with a focus on synthesized soft-body audio. In particular, our subjective evaluation consists of a three-part perceptual study, where we explore the recognisability, quality, and synchronization of the simulated sound. For objective evaluation, we adapt the metrics from generative adversarial networks (GANs) that also measure the recognisability as well as quality of the sound from a different angle. Our results suggest that while both evaluation criteria are largely independent of each other in assessing the recognisability of the sound, objective evaluation tends to be a more efficient alternative to measure the output sound quality. In addition, we provide several findings from our results that can guide sound designers in synthesizing higher quality audio for soft bodies.