Abstract-Hybrid speech recognizers, where the estimation of the emission pdf of the states of Hidden Markov Models (HMMs), usually carried out using Gaussian Mixture Models (GMMs), is substituted by Artificial Neural Networks (ANNs) have several advantages over the classical systems. However, to obtain performance improvements, the computational requirements are heavily increased because of the need to train the ANN.Departing from the observation of the remarkable skewness of speech data, this paper proposes sifting out the training set and balancing the amount of samples per class. With this method the training time has been reduced 18 times while obtaining performances similar to or even better than those with the whole database, especially in noisy environments.However, the application of these reduced sets is not straightforward. To avoid the mismatch between training and testing conditions created by the modification of the distribution of the training data, a proper scaling of the a posteriori probabilities obtained and a resizing of the context window need to be performed as demonstrated in the paper.