Exponential Auto Regressive Model for Nonlinear Modeling of Noisy and Clean Speech

Abstract

This paper proposes a method for non-linear modeling of speech signals using Exponential Auto-Regressive (ExpAR) model. This paper shows the non-suitability of linear Auto-Regressive (AR) model during noisy speech using prediction error from the model. The error signal and Akaike’s Information Criterion (AIC) measures are used to compare the performance of ExpAR model with respect to AR model for speech. The proposed method gives more than 2X improvement in terms of error with respect to the linear auto-regressive model in low SNR situations. Analysis of AR and ExpAR is done for clean speech as well; the proposed ExpAR model attains approximately 2X factor better performance with respect to the linear AR model. The AIC measure obtained also suggests the better suitability of ExpAR with respect to AR for modeling speech signals.