In this work, we introduce a Neural 3D Audio Renderer (N3DAR) - a conceptual solution for creating acoustic digital twins of arbitrary spaces. We propose a workflow that consists of several stages including: 1. Simulation of high-fidelity Spatial Room Impulse Responses (SRIR) based on the 3D model of a digitalized space, 2. Building an ML-based model of this space for interpolation and reconstruction of SRIRs, 3. Development of a real-time 3D audio renderer that allows the deployment of the digital twin of a space with accurate spatial audio effects consistent with the actual acoustic properties of this space. The first stage consists of preparation of the 3D model and running the SRIR simulations using the state-of-the-art wave-based method for arbitrary pairs of source-receiver positions. This stage provides a set of learning data being used in the second stage - training the SRIR reconstruction model. The training stage aims to learn the model of the acoustic properties of the digitalized space using the Acoustic Volume Rendering approach (AVR). The last stage is the construction of a plugin with a dedicated 3D audio renderer where rendering comprises reconstruction of the early part of the SRIR, estimation of the reverb part, and HOA-based binauralization. N3DAR allows the building of tailored audio rendering plugins that can be deployed along with visual 3D models of digitalized spaces, where users can freely navigate through the space with 6 degrees of freedom and experience high-fidelity binaural playback in real time. We provide a detailed description of the challenges and considerations for each of the stages. We also conduct an extensive evaluation of the audio rendering capabilities with both, objective metrics and subjective methods using a dedicated evaluation platform.
