Physics-Informed Neural Networks for Binaural Room Impulse Response Prediction

Laursen, Mads Lang

4. semester

Electronic Systems, Master

2025

2025-06-04

64

This report examines the use of physics-informed neural networks (PINNs) for the prediction of binaural room impulse responses (RIRs). It investigates blind estimation of binaural RIRs for spatial audio rendering in virtual meeting scenarios, with the aim of reconstructing realistic sound fields using only the microphones in a user’s headset and laptop. As the project progresses, the scope is narrowed to focus on PINNs constrained by the three-dimensional acoustic wave equation, estimating RIRs based on room-specific information. A large-scale data synthesis pipeline is developed using PyRoomAcoustics, generating training data across diverse yet realistic room geometries, absorption coefficients, and source/receiver placements, while ensuring variation in interaural time differences. The proposed PINN shows promising performance, requiring less training and achieving more accurate predictions in simplified tasks, in this case predicting the direct sound propagation and early reflections between a source and a receiver, compared to a standard neural network. Nevertheless, it struggles to generalise to complete RIRs, and while the results highlight the potential of PINNs for sound-field reconstruction from minimal input, the model ultimately fails to function as intended. The report concludes by analysing the causes of model failure and outlining directions for future work.

BRIR ; PINN ; physics-informed ; neural ; network ; RIR ; room impulse response ; estimation

View record in AAU Student Projects