Single-Channel Maximum-Likelihood T60 Estimation Exploiting Subband Information
This work addresses incremental improvements in audio signal processing for applications like acoustic characterization, focusing on reducing estimation variance and computational trade-offs.
The authors tackled the problem of estimating reverberation time (T60) from single-channel audio by developing four algorithms that improve upon a baseline maximum-likelihood approach, resulting in lower variance in estimation error for fullband T60 and competitive error for subband T60 compared to fullband estimators.
This contribution presents four algorithms developed by the authors for single-channel fullband and subband T60 estimation within the ACE challenge. The blind estimation of the fullband reverberation time (RT) by maximum-likelihood (ML) estimation based on [15] is considered as baseline approach. An improvement of this algorithm is devised where an energy-weighted averaging of the upper subband RT estimates is performed using either a DCT or 1/3-octave filter-bank. The evaluation results show that this approach leads to a lower variance for the estimation error in comparison to the baseline approach at the price of an increased computational complexity. Moreover, a new algorithm to estimate the subband RT is devised, where the RT estimates for the lower octave subbands are extrapolated from the RT estimates of the upper subbands by means of a simple model for the frequency-dependency of the subband RT. The evaluation results of the ACE challenge reveal that this approach allows to estimate the subband RT with an estimation error which is in a similar range as for the presented fullband RT estimators.