A Comparative Analysis of Progressive Loss Functions with Multi Layered Resnet Model

Abstract

In this paper, one of the proposed ResNet model is used for denoising of RB noise. In fact, ResNet is one of the advanced deep learning methods for analysing and improving various 1D and 2D signals. Accuracy decreases due to the vanishing gradients in plain networks. The model Mozilla common speech data set is used. These are 48kHz recordings of all short sentence speaking subjects. They are all fixed at the same length and the same sampling frequency. The training course for this model uses an Adam optimizer/solver. This model is implemented in scheduling the learning rate “with a division” of 0.9 drop factor and a period of one. About 50 noise samples are available in the data set. Similarly, noise signals are acquired under various environmental conditions. Therefore, one separate data set is prepared for the T&T of the signal. When the T&T data set is small, the problem of overcompliance arises. In other words, since we are only trying to collect all data points from our dataset, we have used one proposed model to manage this dataset more efficiently. In the RMSE and precision validation values, you can feel the over- compliance issues here. Overfitting means that by 1 point of travel, the learning plot starts to deteriorate after loss and an increase in accuracy in terms of the identification. Similarly, if we are trying to pick a simple model for denoising, i.e. there is another problem - underfitting. Underfitting means that the model is either oversized or this model is oversized so that it doesn't learn enough about the dataset using that model. Each time various types of noises tries to rip off the amount added to the voice signal. Improvements in terms of denoising, RMSE and validation precision with the help of this model was given in the following sections.