Blind Directional Room Impulse Response Parameterization from Relative Transfer Functions

Abstract

Acquiring information about an acoustic environment without conducting dedicated measurements is an important problem of forthcoming augmented reality applications, in which real and virtual sound sources are combined. We propose a straightforward method for estimating directional room impulse responses from running signals. We adaptively identify relative transfer functions between the output of a beamformer pointing into the direction of a single active sound source and the complete set of spherical harmonics domain signals, representing all directions. To this end, estimation is performed with a frequency domain recursive least squares algorithm. Then, parameters such as the directions of arrival of early reflections and the reverberation time are extracted. Estimation of the direct-to-reverberant ratio requires dedicated processing. We show examples of successful estimation from speech signals, based on a simulated and a measured response.

Spatial Subtraction of Reflections from Room Impulse Responses Measured with a Spherical Microphone Array

“…The herein proposed method extends the work presented in [6], [7]. It uses a pseudo reference signal obtained by beamforming and dereverberation from a few seconds of captured speech and then identifies a multichannel RIR which forms the basis for the subsequent binaural rendering.…”

“…In this work, we perform batch processing using the MWF, assuming static sources and time-invariant conditions. Equivalently, a recursive least squares (RLS) filter can be employed as in [7] if time-variant conditions are assumed. 1 IV.…”

“…The WAE [7] compares the directional energy distribution of the RIR estimate to the one of the ground truth and is calculated by representing the multichannel RIR via circular harmonics coefficients calculating the mismatch of the pseudo intensity vectors of the estimate and the ground truth, and weighting the directional error by the total energy of the RIR. As the perceptual importance of directional energy is particularly high for early reflections, the WAE was calculated for the early part of the RIRs up to 50 ms after the direct sound.…”

“…However, the authors showed in [6] that cross-relation-based methods fail to converge if acoustic RIRs of realistic lengths (multiple hundreds of milliseconds) are estimated. Thus, new methods that converge quicker have been proposed [6], [7], [12]. They aim at transforming the blind identification task into a non-blind one by applying dereverberation and/or beamforming to estimate a pseudo reference signal.…”

“…Estimates of the transfer functions between the pseudo reference signal and the array microphone signals are then converted to RIR estimates. Promising results have been obtained for reverberation time (RT) estimation using simulated omnidirectional RIRs [12] and multichannel RIRs [7] in the spherical harmonics (SH) domain. The generalized approach from [6] does not require SH-domain processing.…”