Estimating probability of correctness for ASR N-best lists

For a spoken dialog system to make good use of a speech recognition N-Best list, it is essential to know how much trust to place in each entry. This paper presents a method for assigning a probability of correctness to each of the items on the N-Best list, and to the hypothesis that the correct answer is not on the list. We find that both multinomial logistic regression and support vector machine models yields meaningful, useful probabilities across different tasks and operating conditions.