Users of digital mobile phones have access to an increasing number of text-based services but most of them can only use a very reduced keyword for text entry. Traditionally, this has been solved by means of multiple tapping and delays or next-character keys (e.g. 66(pause)666 for no), which is slow (two taps per letter on average), inconvenient, and prone to error. Recently, some mobile terminals offer dictionary-based predictive text-entry schemes which, in almost all cases, allow for one-tap-per-letter text entry (e.g. 367 for for). This paper shows how text corpora can be used to build stochastic finite-state automata which may in turn be easily used to implement predictive text entry for medium-sized languages such as Catalan, not currently supported by major predictive-text-entry companies.