Estimation of Reverberation Time from Live Recordings Using Deep Learning

Silent, continuous measurement of reverberation time from live recordings of speech or music performances in a room has several advantages over the standard measurement techniques relying on deterministic sequences or sound impulses. It is suitable for online monitoring of room acoustics and for the control of systems for artificial reverberation during events. However, a direct, blind estimation of reverberation time and other room acoustic parameters, which is based exclusively on the recordings of sources in the acoustic far field, presents a non-trivial problem. Even with the aid of modern machine learning algorithms, this problem has not been solved adequately in general, meaning for different or multiple sources of sound, such as musical instruments, and for varying room acoustic conditions. In this work, we investigate a three-step approach. In the first step, the direct sound is extracted using any available technique (near-field recording, beamforming...). Second, the room impulse response is estimated in the far field of the sources by means of the deconvolution of the recordings with the direct sound. In the final step, a deep learning model is used for further estimation of reverberation time from the obtained impulse response. The first tests of the method provide promising results.