Blind estimation of the direct-to-reverberant ratio using a beta distribution fit to binaural coherence

Abstract

Knowledge of the direct-to-reverberant ratio (DRR) between an acoustic source and a listening position or receiver can be useful for various acoustic and audio applications including dereverberation, source distance estimation, automatic speech recognition, and binaural synthesis. While the DRR can be computed easily from a room impulse response (RIR), blind estimation using acoustic sources of opportunity is necessary when such RIRs are not available. In this presentation, we describe an approach for blind estimation of the DRR which uses the magnitude-squared coherence (MSC) between the two channels of a binaural signal. The method involves fitting a beta distribution to the MSC values, aggregated over time and frequency, and deriving the estimate from the relationship between the DRR and the shape parameters of that distribution. Validation experiments utilizing speech convolved with binaural RIRs collected from a variety of publicly available datasets yield DRR estimates that are within the just-noticeable difference for DRRs in the range -15 to + 18 dB.