A Deep Learning Approach to Digital Filter Parameter Estimation Based on Amplitude Responses

This paper presents our attempt to tackle the problem of digital filter type and parameter estimation given a set of points sampled from a filter frequency response. We compared results from various multilayer perceptron and convolutional neural network configurations. The results suggest that a convolutional neural network generally produces faster convergence with a lower loss at the same number of epochs than a multilayer perceptron network. However, the maximum amplitude response error, which is the true performance metrics, can be comparable in some cases. A combination of multiple best-performing configurations for different tasks is used to assemble the final model.