Bidirectional truncated recurrent neural networks for efficient speech denoising

We propose a bidirectional truncated recurrent neural network architecture for speech denoising. Recent work showed that deep recurrent neural networks perform well at speech denoising tasks and outperform feed forward architectures [1]. However, recurrent neural networks are difficult to train and their simulation does not allow for much parallelization. Given the increasing availability of parallel computing architectures like GPUs this is disadvantageous. The architecture we propose aims to retain the positive properties of recurrent neural networks and deep learning while remaining highly parallelizable. Unlike a standard recurrent neural network, it processes information from both past and future time steps. We evaluate two variants of this architecture on the Aurora2 task for robust ASR where they show promising results. The models outperform the ETSI2 advanced front end and the SPLICE algorithm under matching noise conditions.