Speech Enhancement Using Residual Convolutional Neural Network
Harinder Singh Mashiana, Abhishek Salaria, Kamaldeep Kaur · 2019 International Conference on Smart Systems and Inventive Technology (ICSSIT) · 2019
This paper deals with Deep Speech Enhancement which uses Deep Neural Networks to denoise a noisy speech to produce a clean speech thus removing almost all of the background noise. The method proposed here uses two dimensional convolutional neural networks with residual connections which take advantage of two key facts i.e the non linear functions learned by convolution neural network and the linearity introduced due to residual networks. These factors help remove noise even in low SNR conditions where usually most of the other speech enhancement models fail. The model proposed is fast, robust and lightweight and requires less training as compared to other deep learning models employed for the same task. The model is not expected to produce state of the art results but produces reasonable results keeping in mind the resource limitations and the small size of the model. The model is built upon previous research done in the field and main aim is the reduction in model size while simultaneously preserving the performance.