A two-layer Kohonen neural network using a cochlear model as a front-end processor for a speech recognition system

The authors describe a two-layer neural network speech recognition system based on Kohonen's algorithm. A cochlear model is used as a front-end processor for the system. The basilar membrane is represented by a cascade of 128 digital filters, of which 90 filters fall within the speech bandwidth of 250 Hz to 4 kHz. The outputs of these 90 filters are presented as the input vector to the first layer of the Kohonen net every 16 ms. The input to the second layer consists of a concatenated vector, created from a trajectory of successively excited neurons, firing on the first layer. Sammon's nonlinear mapping algorithm was used as an analysis tool for measuring the effectiveness of different parts of the recognition process. The system was first simulated and later implemented on Inmos transputers.<>