This paper deals with the problem of acceptance/rejection of recognition hypotheses for continuous speech utterances. Two different techniques are investigated to improve the rejection of out-of-vocabulary (OOV) words. A combined approach is first proposed which uses two garbage models (a trained one and an on-line garbage model). The second method uses the trained garbage model and consists in post-processing the recognizer hypotheses by computing for each of them a confidence measure. Both approaches are evaluated in the context of a stock exchange application through the telephone for French. The parameters of the two approaches are studied to improve recognition accuracy.