Blind Estimation of Spatial Room Impulse Responses Using a Pseudo Reference Signal

Paper i proceeding, 2024

In auditory augmented reality applications, virtual sound sources can be added to a real-world acoustic environment by processing each source signal with a spatial room impulse response (SRIR) to render acoustic characteristics of the environment, and with a set of head-related transfer functions to create binaural headphone signals.
The SRIR of a user's environment is typically unknown and needs to be estimated. We propose a method to estimate the SRIR blindly from speech signals captured with a microphone array. The blind estimation task is transformed into a non-blind one using a pseudo reference signal that is obtained from the array signals via dereverberation and beamforming. The SRIR is then estimated using a frequency-domain multichannel Wiener filter with the pseudo reference as the input and the array signals as the desired signals.
In contrast to conventional methods, the proposed method is able to successfully estimate SRIRs of realistic lengths at a sampling rate that supports the entire audible frequency range.
Results from 200 simulated and 16 measured SRIRs show that the estimates from the proposed method reproduce the reverberation time and the direct-to-reverberant energy ratio with low error, outperforming a baseline method that does not use dereverberation.