In this paper a hierarchical classification of different levels of phonetic information is proposed in order to improve phone recognition. In this paradigm several intermediate classifiers give posterior probability predictions for broad phonetic classes, achieving phone detail in the last layer. Class membership probabilities are weighted and combined in order to get a more robust phoneme prediction. A method for finding the best set of weights is also proposed based on discriminative training in a hybrid MLP/HMM system. Ex-periments show that the use of broad-class information en-hances phone recognition. Relative improvements of 8% in Correctness and 5% in Accuracy were achieved in phoneme recognition on the TIMIT database compared to a baseline system.