Real time speech enhancement using densely connected neural networks and Squeezed temporal convolutional modules

In this paper, we present a fully convolutional neural network for enhancing real-time speech in the time domain. Skip connections are included in the architecture of the proposed encoder-decoder network. The layers in the decoder and encoder consists of densely connected blocks (DCB) with causal and dilated convolutions. These dilated convolutions facilitate the aggregation of contextual information across multiple resolutions. The network is ideal for real-time applications due to the causal convolutions' utilization of information inflow prevention from subsequent frames. Additionally, we propose employing up sampling in the decoder with sub-pixel convolutional layers. We also proposed a Squeezed temporal convolutional network (STCNs) after every dense block in encoder and decoder. According to experimental outcomes, the suggested model greatly surpasses previous state-of-the-art models in quality scores as well as objective intelligibility in real-time scenarios.