Improved utterance rejection using length dependent thresholds

In this paper, we propose to use an utterance length (duration) dependent threshold for rejecting an unknown input utterance with a general speech (garbage) model. A general speech model, comparing with more sophisticated anti-subword models, is a more viable solution to the utterance rejection problem for low-cost applications with stringent storage and computational constraints. However, the rejection performance using such a general model with a fixed, universal rejection threshold is in general worse than the anti-models with higher discriminations. Without adding complexities to the rejection algorithm, we propose to vary the rejection threshold according to the utterance length. The experimental results show that significant improvement in rejection performance can be obtained by using the proposed, length dependent rejection threshold over a fixed threshold. We investigate utterance rejection in a command phrase recognition task. The equal error rate, a good figure of merit for calibrating the performance of utterance verification algorithms, is reduced by almost 23% when the proposed length dependent threshold is used.