We present a rejection method of out-of-vocabulary words assessed on two representations of extraneous speech inputs. First, we consider a garbage template to match the non keywords, named the ergodic 7-garbages model, which is directly built from the context independent (CI) phoneme models used to represent the keywords. In a second step, we do not attempt to explicitly model the extraneous inputs : this method refers to the ergodic best-fit model. The main motivation of using these two garbage models is that they do not require any specific training. We report results on rejection performance for speaker independent isolated words recognition task (perplexity 54) with standard discrete density HMM and hybrid HMM/MLP based recognizers.
Bibliographic reference. Accaino, Sari / D'hoore, Bart / Vantieghem, Johan / Compernolle, Dirk Van (1995): "Rejection capabilities for HMM-based speech recognizers", In EUROSPEECH-1995, 2115-2118.