The aim of the this paper is to examine the extent to which field data can improve speech recognition performance when included into the training procedure of the model parameters. The approach used herein was twofold. First of all, a thorough error analysis of the misrecognized words was carried out before and after the introduction of field data into the training procedure. Such an analysis enabled the detection of major perturbation effects on word recognition, as well as the possibility of modelling them using Hidden Markov Models. Secondly, garbage models were trained using non-speech signals and out-of-vocabulary words from the field data, together with garbage models trained on laboratory data. These garbage models showed improved rejection performance when compared to garbage models trained exclusively with laboratory data.