In this paper, we propose a new active/non-active word control algorithm, which we call unknown word re-evaluation algorithm. In the algorithm, we incorporate an unknown word detection using Garbage Models into a speech recognizer which dynamically changes its vocabulary. When the Garbage Models are perceived, non-active words suitable for the situation are activated and compared with the unknown part of the speech. Information for the re-evaluation is obtained from the speech containing the unknown part. Supplementary information may be obtained in the succeeding discourse in some cases. The algorithm makes a dynamic vocabulary recognizer more useful decreasing the risk of misprediction against the next utterance without spoiling its advantages in recognition speed and rate. In our preliminary experiment, 86.4% of proper nouns were correctly re-evaluated using the Garbage Models, while the false alarms of 5% were found.