This paper presents our automatic speech recognition engine ESPERE and several results obtained from experiments on telephone speech. ESPERE (Engine for SPEech REcognition) is a HMM-based toolbox for speech recognition allowing the user to choose the modeled unit (word, phone, triphone), define the topology of every Hidden Markov Model, train the models with the Baum-Welch algorithm and evaluate the recognition accuracy on speech databases. To validate the ESPERE toolbox, we have conducted tests on real world data: the recognition of a three-digit code to access a call center. We have investigated the influence of some parameters and some preprocessing algorithms. Finally, combining the best parameters, the recognition score reaches 96.4% at the word level and 92.1% at the sentence level.