Parallel implementation of time delay neural networks for phoneme recognition

Large training problems using the backpropagation (BP) as a training algorithm are very computationally demanding when used for multilayer perceptrons (MLPs). In speech recognition applications, this is especially true. In order to overcome these problems, implementation issues are addressed. The algorithm is parallelized by splitting the training database across many processors and updating the network weights using an arithmetic average of the weight updates for each transputer microprocessor. The parallel implementation achieves a speedup factor of 8.8 over a VAX 3600 and scales well up to 16 transputers in a hypercube configuration. Phoneme recognition accuracy for Afrikaans fricatives, nasals, stop consonants and glides is given