For improving the efficiency of a speech recognition system based on a vocabulary of statistical word models, initially coarse word preselections are made for utterances, and are marked in a training session to indicate whether the selections were correct or wrong. Furthermore, each utterance is matched against each word preselected for it to obtain a probability measure for that combination as in a usual reccgnition process. Based on this knowledge of the correct or erroneous result of the coarse selection plus the probability measure, a discriminant analysis is made to determine, for the whole population of word models, how each phone in each word model should be weighted so that an optimum discrimination between similar words is achieved. The weighting coefficients thus obtained are stored with the word models and are later used, during actual speech recognition, to weight the probabilistic contribution of each phone for the final word selection.