This paper describes the development of a grapheme (letter) to phoneme (sound) conversion system for European Portuguese. The system uses the combination of the two established approaches: one based on phonological rules (rule-driven system) and other based on stochastic models (data-driven system). The final system also includes a module to deal with inflected terms (nominal, verbal and adjectival) in order to increase the coverage of the vocabulary and reduce de pronunciation lookup table. The results show that the inclusion of some phonological rules strongly improves the conversion performance. The system evaluation on a dictionary with 40,000 most frequent words (from the CETEMPúblico corpus) shows a pronunciation word error rate of 3.65 %. Problems of multiple pronunciations of heterophonic homographs words were also addressed by giving additional information of part of speech or by specifying the stressed vowel.