Automatic Parameters Tuning of Late Reverberation Algorithms for Audio Augmented Reality
Riccardo Bona, Davide Fantini, Giorgio Presti, Marco Tiraboschi, Juan Isaac Engel Alonso-Martinez, Federico Avanzini · 2022
The matching of reverberation features between real sound sources and virtual ones is a key task in Audio Augmented Reality. An adequate matching provides a proper auditory immersion to the user. In this paper, we propose a method for reverb matching. The method automatically optimizes the parameters of an artificial reverberator to match a target Room Impulse Response (RIR). We used a Bayesian optimization procedure using a Gaussian Process as a prior distribution. This procedure iteratively tunes the artificial reverberator parameters to match its output, i.e. the approximated RIR, with the target one. The matching between the approximation and the target is implemented with a perceptually motivated loss function. Before this parameters optimization, the target early reflections are approximated through an autoregressive model. The method has been implemented with two artificial reverberation algorithms: a Feedback Delay Network (FDN) and an implementation of the Schroeder-Moorer reverberator (Freeverb). We evaluated the method with a listening test to assess the similarity with target reverberation. Our method yields overall statistically significant higher scores with respect to other anchor conditions. Further, subjective differences between FDN and Freeverb are not significant.