Recently, deep learning-based methods for blind reverberation time estimation have been proposed, and outperform those based on conventional signal processing. The signal processing approaches extract the reverberant environmental features of… Click to show full abstract
Recently, deep learning-based methods for blind reverberation time estimation have been proposed, and outperform those based on conventional signal processing. The signal processing approaches extract the reverberant environmental features of sound by statistical analysis, while deep learning approaches train a network to capture the relationship between acoustic features and the reverberation time. In this letter, we propose a method for blind reverberation time estimation that explicitly reflects physical properties of reverberation by combining the deep learning approach, attentive pooling, and statistical characteristics of reverberant speech obtained using a signal processing method, i.e., spectral decay rates (SDRs). The results obtained with the proposed blind reverberation time estimation method are superior to the previously published state-of-the-art results for the EVAL dataset of the ACE Challenge. This work can be considered a good example of the collaboration between signal processing expertise and deep learning approach.
               
Click one of the above tabs to view related content.