A Research on Deep Learning Approaches for Denoising of Audio Signals

Abstract

This Article is going to compare the denoising performance of proposed architecture with the existing ones. Till now [5] CNN and its derivatives like CNN.0, CNN.1 (fully connected layer/without skip connection) and CNN.2 (with skip connections) were developed with and without Single skip connections. These architectural results were compared in terms of [5] PESQ (Perceptual Evaluation of speech quality), [5] LSD (Log Spectral Distance) and [5]SNR. The latest model CNN.2 provides worse performance than CNN.0 in PESQ scores at some SNR levels (-15dB, -10d Bandi-5dB) and also better performance was achieved for CNN.2 in contrast with wiener filtering. Once again CNN.2 is going Ito reconstruct with the help of pooling layers and skip connections. Even though it provides additional burden to the Network parameters, better immune levels are achieved.