Semi-Blind Estimation of Direct-to-Reverberant Energy Ratio Using Residual Energy Test Statistics

Abstract

We propose a method for semi-blind estimation of direct-to-reverberant energy ratio (DRR) using a number of statistical features of the residual energy test (RENT), a recently proposed direct-path metric that evaluates the dominance of a direct or diffuse component in an ambisonics recording. Our method is based on the observations and correlations between these features of this test and various recording scenarios. This method requires a procedure resembling a calibration step to tune the parameters of a linear regression model. Recordings from various locations within a room are used to extract the statistics that connect the residual energy test values to direct-to-reverberant energy ratio value estimates. For rooms with simpler geometry and volume, our method gives a median mean absolute error of 0.5 dB. For rooms featuring various materials and coupled volumes, errors can grow with increasing distances, with a median mean absolute error of 2.5 dB. The performance remains the same for speech or music recordings and is found to improve when a wider variety of sounds are used in the calibration step.