Nonlinear models of the cochlea are best implemented in the time domain, but their computational demands usually limit the duration of the simulations that can reasonably be performed. This letter presents a modified state space method and its application to an example nonlinear one-dimensional transmission-line cochlear model. The sparsity pattern of the individual matrices for this alternative formulation allows the use of significantly faster numerical algorithms. Combined with a more efficient implementation of the saturating nonlinearity, the computational speed of this modified state space method is more than 40 times faster than that of the original formulation.

Models of the mammalian cochlea have been proposed in a number of ways and they have varying degree of realism and complexity. The transmission-line (TL) models are faithful to the physiology, particularly in terms of cochlear nonlinearity, but are computationally demanding. The pole-zero filter cascade (PZFC) model, however, is much more efficient to implement, but the nonlinearity is included implicitly, using an automatic gain control network. In this study, the connection between the linear responses of these two models is first discussed, followed by a comparison of their nonlinear responses in terms of self-suppression and two-tone suppression on the level of the basilar membrane. Both models are capable of simulating dynamic range compression as measured on the cochlear partition, but the TL model is more reasonable in representing twotone suppression, with the PZFC having lower suppression thresholds and over-predicting the suppression due to high-side suppressors. Further tuning of its parameters and structure, especially the automatic gain control (AGC) network may be possible to make it more compatible with these experimental observations. After adapting the PZFC model to have a more realistic nonlinear behavior, its use for investigating auditory signal processing such as masking effects, and hence as a front-end processor for acoustic signals can be enhanced, while retaining its computational efficiency.