LOW-LATENCY APPROXIMATION OF BIDIRECTIONAL RECURRENT NETWORKS FOR SPEECH DENOISING

The ability to separate speech from non-stationary background disturbances using only a single channel of information has increased significantly with the adoption of deep learning techniques. In these approaches, a time-frequency mask that recovers clean speech from noisy mixtures is learned from data. Recurrent neural networks are particularly well-suited to this sequential prediction task, with the bidirectional variant (e.g., BLSTM) achieving strong results. The downside of bidirectional models is that they require offline operation to perform both a forward and backward pass over the data. In this paper we compare two different low-latency bidirectional approximations. The first uses block processing with a regular bidirectional network, while the second uses the recently proposed lookahead convolution layer. Our results show that using just 1000 ms of backward context can recover approximately 75% of the performance improvement gained from using bidirectional as opposed to forward-only recurrent networks.