Learning to estimate reverberation time in noisy and reverberant rooms

The reverberation time, T₆₀, is an important indicator of the reverberation strength in a room and has many applications in speech processing, such as dereverberation. However, the T₆₀ must be blindly estimated if only reverberant speech is available. In this paper, we provide a learning based approach for T60 estimation. We treat the T₆₀ estimation as a classification problem by dividing the T₆₀ range into countable bins (e.g. 19 bins covering 0:1s to 1s with a bin width of 0:05s) and the estimation becomes predicting which bin the true T₆₀ falls into for a given speech. We use deep neural networks (DNN) to learn such a mapping from speech to the T₆₀. The DNN is trained on a large amount of reverberant and noisy speech signals generated from various simulated rooms with known reverberations. After training, we observe that the DNN can learn highly sensible features for the T₆₀ estimation task. Experimental results on the data from both simulated rooms and real rooms confirmed the effectiveness of the DNN learning based approach. In all the test cases, the DNN method significantly outperforms the state-of-the-art SDD T₆₀ estimation method.