Optimization of Input Parameters for the Real-time Simulation of Room Acoustics – Revisited Dirk Schröder 1 , Björn Starke 1 , Hans-Joachim Maempel 2 1 Institute of Technical Acoustics, RWTH Aachen University, 52066 Aachen, E-Mail: dsc@akustik.rwth-aachen.de 2 Audio Communication Group, TU Berlin, 10623 Berlin, E-Mail: hans-joachim.maempel@tu-berlin.de Introduction Immersive environments mostly aim to simulate opto- acoustical scenes in a plausible way. For real-time auralization, methods of geometrical acoustics [1] provide quite accurate results within a reasonable computation time. The best results are achieved by combining deterministic methods for the computation of early specular reflections with stochastic approaches, such as ray tracing, for the computation of the reverberant sound field. The computations must be performed at interactive rates, thus their costs have to be minimized. This is typically done by reducing the order of image sources and/or the number of traced rays, as these parameters strongly influence both, the computational costs and the perceptual accuracy of the simulation. Since the degradation caused by the reduction does not have to be audible, a parameter threshold (PT) is a measure for an optimum parameterization since no perceptible improvement can be achieved above this threshold. The PTs of the number of traced rays have been experimentally determined for three orders of image sources as well as for an acoustical and an opto-acoustical stimulus generated in a CAVE-like environment. Therefore, a listening test applying a criterion-free 3AFC-paradigm with two different assessment methods and with the participation of expert listeners has been performed. The 3×2-design reveals the interaction of relevant technical and perceptual conditions. Results show that a lowered accuracy, hence faster computation time of the simulation, is not noticeable when a convergent opto-acoustical stimulus is presented. Listening Tests Due to the promising results of the preliminary studies presented at NAG/DAGA 2009 [2], several refined listening test series were carried out for further investigations [3]. The results presented here include additional listening tests that were accomplished in collaboration with the Audio Communication Group (ACG) at TU Berlin. In this joint series, the same geometrical room model of a concert hall Figure 1: Test subject doing listening tests inside the CAVE-like environment at RWTH Aachen University (volume 14372m 3 , T(Sabine) 1.3 sec) was used as in the preliminary study, but with different auditory stimuli and a different receiver position. 21 Stimuli were created from convolving differently simulated impulse respones with a dry recording. With regard to the facilitation of the detectability of potential artefacts, a slowly picked acoustic guitar was chosen as audio content. For the simulation parameterization, the settings of the preliminary study were refined with focus on the PTs that were obtained at that time. Thus, models based on 3 orders of image sources and, at a time, 7 numbers of particles per simulated frequency band (100, 4000, 7000, 9000, 12000, 20000, 40000) were applied for the production of the stimuli. The numbers range from a level causing an easily detectable distorted sound (100 rays) to a state of the art simulation, which was defined as reference (40000 rays). The listening tests were performed according to the method of constant stimuli following a 3AFC-paradigm. The task was to discriminate between the test stimulus and the reference by determining the position of the differing stimulus. In order to improve its reliability, the test (see Table 1) was repeated once. A GUI guided the subjects through the test. Factor I (repeated measures) Fixed image source order Measure: PT to the reference [No. of particles] 1 2 3 acoustical (Lab) Na=19 Factor II (grouping variable) Stimulus modality opto-acoustical (CAVE) Noa=14 Table 1: Overview of the test design. N denotes the number of subjects. DAGA 2010 - Berlin 711