Joint Estimation Of Acoustic Parameters From Single-Microphone Speech Observations

Key parameters that characterise the degree of degradation in an acoustic environment are reverberation time (RT), direct-to-reverberant ratio (DRR) and signal-to-noise ratio (SNR). To address the inherent interplay that exists between these parameters, which can hinder existing methods designed to estimate only a single parameter, we propose a data-driven solution to jointly estimate all three parameters using a convolutional neural network. To facilitate robustness to unseen acoustic conditions, the method is trained using a large set of simulated acoustic impulse responses that have been carefully selected so as to mitigate interplays that exist between RT and DRR. In this work we evaluate the performance of the proposed estimator with respect to reverberation only. Results show the estimator compares favourably with respect to the state-of-the-art for unseen and real acoustic scenarios.