Adaptive filter based on NARX model for recorded audio noise removal

This paper presents system identification-based approach to create a Non-linear Auto-Regressive model with Exogenous (NARX)-based adaptive noise filter to remove noise from recorded audio signals. The NARX model was trained with noisy recorded signal as inputs, and clean signal (from the MP3 audio file) as the output. The system identification process then tries to relate between the input and the output so that the noise component from the input is removed in the output stage. The binary Particle Swarm Optimization (BPSO) algorithm was used to perform model structure selection (selection of input and output lagged signals that best explains the future values of the data). Parameter estimation of the NARX model was done using Householder Transform-based QR factorization. Fitting and residual tests results show that the NARX model was successful in estimating the model, and filtering out noise well.