Blind reverberation time estimation using a convolutional neural network

The reverberation time of an acoustic environment is a useful parameter for applications including source localisation, speech recognition and mixed reality. However, estimating the reverberation time blindly and on the fly remains a challenge. Here we propose formulating the estimation as a regression problem and using a convolutional neural network (CNN) to estimate the reverberation time directly from a four second long single-channel recording of reverberant speech in noise. Evaluation on the ACE Challenge data corpus suggests that the proposed method is computationally efficient and outperforms state-of-the-art methods.

Block diagram of convolutional neural network architecture

Confusion matrices of ground truth and estimated T60 for training set (left) and evaluation set (right). Results are binned by T60 with a resolution of 0.1 s.