Experimental evaluation of structure of garbage model generated from in-vocabulary words

This paper proposes an effective garbage model for rejection of out-of-vocabulary words (OOV) in a word recognition system. Many methods for rejecting OOV and generating garbage models based on registering OOV have been proposed. However, they could not have sufficient rejection capability. To solve them, we propose a new garbage model generated from only in-vocabulary words (IV) without information about OOV. The proposed garbage model employs 2 parameters, i.e., the number of states (NS) and the number of mixtures (NM). From a large amount of word recognition simulations, we derived the best values of the parameters. We found that the equal error rate of the proposed garbage model was saturated around NM = 120 at any NS, and no significant change was observed with increasing NM more. In addition, the best value of NS was 18 at NM = 120. The results of simulations showed that the best parameters of the proposed garbage model (i.e., NS and NM) were 18 and 120, respectively.