Online reverberation time and clarity estimation in dynamic acoustic conditions

Previously proposed methods for estimating acoustic parameters from reverberant, noisy speech signals exhibit insufficient performance under changing acoustic conditions. A data-centric approach is proposed to overcome the limiting assumption of fixed source–receiver transmission paths. The obtained solution significantly enlarges the scope of potential applications for such estimators. The joint estimation of reverberation time RT60 and clarity index C50 in multiple frequency bands is studied with a focus on dynamic acoustic environments. Three different convolutional recurrent neural network architectures are considered to solve the tasks of single-band, multi-band, and multi-task parameter estimation. A comprehensive performance evaluation is provided that highlights the benefits of the proposed approach.