A new recognizer for isolated and connected words based on second-order hidden Markov models is presented. This system uses a modified formulation of the Baum-Welch algorithm and an extension of the Viterbi algorithm. A comparative study between first-order (HMMl) and second-order (HMM2) Markov models is carried out. In order to evaluate the performances of both models, the TI-NIST database of isolated and connected digits was used.