Implementation of Audio De-noising using Convolutional Neural Networks on Edge Device

Babble noise causes major discomfort to the users of hearing aids. It is a challenging task to remove the babble noise from background without disturbing the human speech present in the signal especially in low SNR environment. The architecture used in this paper is based on learning the matching between clean speech signal and noise speech signal as seen in some of the works done on speech enhancement. The model used is CR-CED network which is made of repeated redundant convolutional encoder decoder network and achieves better performance with reduced time for convergence. An important attribute of CR - CED network is that convolution is carried out in only 1-Dimension with input spectrum shape of 129*8 performed in frequency axis. This makes the frequency axis to remain constant towards the propagation. Overall, the union of small number of training parameters make the model lightweight with fast execution especially on edge devices [2].