Rejection using rank statistics based on HMM state shortlists

We study a measure of confidence in the speech recognizer output based on a rank-order probability model of HMM state likelihoods. The motivation for rank models is based on the conjecture that statistics based on ranks are likely to be more robust than those based on the likelihood values, especially when the test and training distributions are mismatched. We investigate a number of different issues that arise in the development of rank models. We test the proposed rank-order model on two ASR rejection tasks: a combination of the log-likelihood ratio and rank order probability yields relative reductions of the equal error rates of 31% and 8% (for the two tasks, respectively) over the log-likelihood ratio alone.