Speech Modeling with Linear Auto Regressive and Nonlinear Hammerstein Models

This paper aims to explore speech modeling through the utilization of both a Linear Autoregressive (AR) model and a Nonlinear Hammerstein model. A comparative analysis was conducted, encompassing both clear and noisy speech samples, to assess the appropriateness of the linear autoregressive model for speech by examining the prediction errors it generates. This evaluation involved measuring the performance of the linear autoregressive model in comparison to the Hammerstein model for speech, using the error signal and the application of the Akaike Information Criterion (AIC). The findings from this study indicate that the proposed approach results in enhancements in terms of error reduction, particularly in scenarios characterized by low signal-to-noise ratios (SNR). Additionally, the obtained AIC measurements also suggest that the nonlinear model is better suited for modeling audio signals when contrasted with the AR model.