Real-time spatial audio rendering is increasingly prevalent in applications such as virtual reality (VR), augmented reality (AR), gaming and live sound.

Creating accurate room acoustic models in these applications is challenging, but essential for achieving high levels of immersion and presence.

Wave-based models provide high physical accuracy but are computationally expensive and do not scale well to high frequencies or large environments.

As a result, geometric acoustic models are commonly used for real-time applications due to their computational efficiency.

However, these models are based on high-frequency approximations of sound propagation and often neglect important wave phenomena such as diffraction (bending of sound waves around obstacles).

As virtual environments become increasingly complex, there is a need for models that can accurately simulate acoustics as receivers and sound sources move through the environment.

A key scenario occurs when a sound source is occluded from the receiver, removing the direct sound path.

In such cases, the first arriving sound is due to diffraction which strongly influences perceptual attributes such as sound localisation and colouration.

Models that only consider reflections by surfaces, and neglect diffraction, can therefore produce highly inaccurate and implausible results in these scenarios.

Although a number of diffraction models aimed at real-time applications have recently been proposed, few have been assessed using perceptual evaluations.

In particular, only a limited number of listening experiments have included comparisons between diffraction models or evaluated path-finding algorithms.

The first objective of this thesis is to conduct a perceptual evaluation of existing diffraction models, assessing their perceived naturalness and determining whether efficient infinite impulse response (IIR)-based models can achieve similar perceptual performance to physically accurate but more computationally expensive models.

A second objective is to address the limitations of existing diffraction models by developing a novel diffraction model that is both computationally efficient and perceptually plausible.

This model trains a neural network to fit 2nd-order IIR filters to data generated by a physically accurate diffraction model.

The final objective is to develop and apply a methodology and framework for evaluating the perceptual plausibility of room acoustic models in dynamic VR and AR scenarios, where users can move and rotate freely.

This investigation examines the influence of late reverberation on the perception of diffraction and identifies which diffraction paths are important for perceptual plausibility.

The results indicate that diffracted reflection paths are important when the sound source is heavily occluded and that increased levels of late reverberation reduce perceptual differences between simulations with and without diffraction.

The proposed diffraction model achieves perceptual plausibility comparable to the physically accurate model, while being significantly more computationally efficient.

In the AR scenario, simulations including diffraction are rated as more plausible than those without diffraction, although all simulations are rated as less plausible than the real sound source.