The renewed interest in grapheme to phoneme conversion (G2P), due to the need of developing multilingual speech synthesizers and recognizers, suggests new approaches more efficient than the traditional rule&exception ones. A number of studies have been performed to investigate the possible use of machine learning techniques to extract phonetic knowledge in a automatic way starting from a lexicon. In this paper, we present the results of our experiments in this research field. Starting from the state of art, our contribution is in the development of a language-independent learning scheme for G2P based on Classification and Regression Trees (CART). To validate our approach, we realized G2P converters for the following languages: British English, American English, French and Brazilian Portuguese.