Evaluating Reverberation Models for Augmented Reality

The ultimate goal of the reverberation model for augmented reality is to create an auditory illusion, making simulated sound sources indistinguishable from real, measured ones. However, existing evaluation methods are not tailored to achieve this objective. This paper adopts the evaluation paradigm of auditory illusion tests to evaluate reverberation models under two distinct tasks: authenticity and transferring. The listening test uses the three-alternative forced-choice design, where subjects are asked to detect the speech signal processed with a model-generated room impulse response (RIR) among the signals processed with measured RIRs. For the authenticity task, the three signals contain the same speech sample, while for the transferring task, they contain different samples from different speakers. A Bayesian analysis shows that detecting model-generated RIRs is significantly more challenging in the transferring task than in the authenticity task across all models. Additionally, while the listening test results positively correlate with the selected objective metrics, the reliability and generalizability of these correlations for predicting listening test outcomes remain uncertain. The proposed evaluation framework for reverberation models can serve as a precursory analysis for developing dynamic, binaural rendering for augmented reality applications.