Parallel training algorithms for continuous speech recognition, implemented in a message passing framework

A way of improving the performance of continuous speech recognition systems with respect to the training time will be presented. The gain in performance is accomplished using multiprocessor architectures that provide a certain processing redundancy. Several ways to achieve the announced performance gain, without affecting precision, will be pointed out. More specifically, parallel programming features are added to training algorithms for continuous speech recognition systems based on hidden Markov models (HMM). Several parallelizing techniques are analyzed and the most effective ones are taken into consideration. Performance tests, with respect to the size of the training data base and to the convergence factor of the training algorithms, give hints about the pertinence of the use of parallel processing when HMM training is concerned. Finally, further developments in this respect are suggested.