Auditory representations of speech sounds in a neural model: the role of peripheral processing

The categorization of speech sounds by the auditory system has been a subject of intense attention. In the author's previous work it has been established that a two-stage computational model can mimic important aspects of the speech categorization behavior of human and animal listeners. The first stage employs a biologically motivated 'front-end' for modeling the peripheral auditory system, and the second stage is a trainable artificial neural network 'back-end' for modeling more central processes. The behavior is emergent in that it was not explicitly programmed into the model. A software model can be interrogated to find out the contribution of its component parts to the overall behavior. Replacing the auditory front-end by a more prosaic fast Fourier transform analyzer allows one to focus on the contribution of the acoustic in auditory transformation to categorization. We find that the front-end processor is not essential to category formation but plays an important part in the boundary-movement phenomenon, by emphasizing important time frequency of the speech signal.