Blind reverberation time estimation using multi-channel speech recordings

Reverberation time is an objective index to characterize the reverberation and is also one of the most important parameters in room acoustics. Numerous methods for single-channel blind reverberation time estimation have been proposed and have achieved promising results. However, estimating reverberation time using single-channel speech only utilizes the spectro-temporal characteristics. As a matter of fact, the direct and early reflected sounds and the late reverberation have quite different spatial properties: the former is directional, while the latter is more diffused. Therefore, estimating reverberation time blindly using multichannel speech recorded with microphone arrays can utilize the information in the spatial domain, which can be expected to improve the performance of the reverberation time estimation. In this work, a multi-channel blind reverberation time estimation method was proposed. The complex spectra of multi-channel speech recordings were used as the input features of the neural network and the framework of neural network was designed to extract the spatial information. The estimation methods using single-channel, 2-channel and 3-channel speech recordings were evaluated in noisy environments at different SNRs. Experimental results showed that multi-channel estimation methods outperformed single-channel methods.