Real-Time Zero-Phase Digital Filter Using Recurrent Neural Network

Abstract

This paper proposes a method to design and implement a zero-phase digital filter that can run in a real-time system. Generally, zero-phase filters are designed for non-causal systems only as the time-reversal operations are required. Thus, the typical usage of these filters is for offline applications. For this reason, we propose a real-time zero-phase digital filter that is designed based on a recurrent neural network model, particularly the gated recurrent units. The model learns to perform zero-phase filtering by using training data made from the filtered signals that are generated by using the conventionally designed zero-phase filter. The original digital filter used to create the dataset is an IIR filter performing forward-backward filtering. The best trained model yields the mean absolute loss values at approximately 0.001 and can process at least 30 times faster than real-time. Furthermore, the trained model was implemented as a 3-band zero-phase graphic equalizer to exhibit one of its applications.