DEMUCS-Mobile : On-device lightweight speech enhancement

As the importance of speech enhancement for real-world application increases, the compactness of the model is also becoming a crucial study. In this paper, we present compression techniques to reduce the model size and applied them to the state-of-the-art real-time speech enhancement system. We successfully reduce the model size by actively applying channel pruning while maintaining performance. In particular, we propose a method to prune more channels of convolutional neural networks (CNN) by utilizing gated linear unit (GLU) activation. In addition, lower-bit-quantization is applied to reduce model size, while minimizing performance degradation caused by quantization. We show the performance of our proposed model on a mobile device where computing resources are limited. In particular, it is implemented to enable streaming, and speech enhancement works in real-time.