The present paper describes an incremental approach to the optimization of an HMM recognizer for a dialling-by-name task in a telephone environment. A Bayesian adaptive learning scheme was used to obtain domain-adapted models of sub-word speech units deriving from general, application independent models. The Maximum A Posteriori (MAP) theory was exploited in order to combine two sources of information, the Domain-Independent (DI) and the Domain-Dependent (DD) speech corpora. This approach was tested on a Voice Dialling-by-Name recognition task over the telephone in a speaker-independent mode. A nearly 92% recognition score was obtained for the best hypothesis using the optimal model adaptation procedure, whereas a 97% score has resulted by inclusion of the second best hypothesis.