Generating Room Impulse Responses Using Neural Networks Trained with Weighted Combinations of Acoustic Parameter Loss Functions

This paper investigates loss functions based on room acoustic parameters for training conditional generative adversarial networks (CGANs) used for generating room impulse responses (RIRs) at specific locations within a real room. The CGANs are trained on RIRs recorded at multiple positions in the room and then used to generate RIRs at new positions. The study evaluates the effectiveness of adaptive and fixed weightings, applied to various combinations of acoustic parameter-based loss functions, including those based on reverberation time (RT) and early decay time (EDT), along with the time-domain mean squared error (MSE) and frequency-domain multi-resolution short-time Fourier transform (MRSTFT). Results from reconstructing RIRs in two real rooms show that adaptive weightings are more effective than fixed weightings. MRSTFT surpasses MSE in accurately reconstructing RIRs across all frequency bands. Additionally, the loss functions using combinations of MRSTFT, RT30 and EDT with adaptive weightings significantly enhance the accuracy of RIR generation.