Automatic speech recognition error/no decision tradeoff curves

This paper analyzes the tradeoff between ASR recognition error probability (PE) and percent of no decisions (ND). The tradeoff characteristic of PE versus ND is first studied by means of a simple model of the use of the ratio of the best to the next-to-best scores as decision variable. The robustness of the tradeoff curves with respect to modeling assumptions is displayed. More insight into the tradeoff curves is obtained when we compare using (as decision variable) the ratio of the best to the next-to-best scores (the relative case) with using the best score alone (the absolute case). It is seen that the tradeoff curves for the absolute case depend on more details about the scores while the curves for the relative case are parameterized by a single parameter. Thus, there is a set of canonical curves for the relative case and not for the absolute case. Also, the significant role of correlation between scores of different words is shown.