Parallel training of MLP probability estimators for speech recognition: a gender-based approach

Explores the averaging of mixtures of multiple neural network probability estimators in speech recognition. The authors experiment with different ways of dividing up the speaker space. A division based on gender seems to be the most important. The division based on a priori knowledge (in this case, rate of speech) resulted in lower error rates than the use of k-means clustering. The overall accuracy of the Parallel Net architecture is about: the same as the monolithic probability estimator, but communication costs on parallel machines can be expected to be significantly lower. Additionally, the overall product of patterns times parameters is lower with such a partitioning, resulting in reduced training time even on serial machines