In this paper we make first a comparative study of the recognition performance of different HMM training algorithms in Spanish: Discrete context-independent phone models, Discrete context-dependent (aglomerative-clustered generalized triphone) models and Semieontinuous context-independent and context-dependent models. We also propose two alternatives to improve the performance of the systems, the first one by using phone-class dependent modelling. The second one by preprocessing the training sentences in order to separate interword pauses and consequently train better contextual models. Preliminary experiments on a speaker dependent database, 1000 words vocabulary show good improvement of both systems compared to the baseline system.