Improving Task Independent Utterance Verification Based on On-line Garbage Phoneme Likelihood

Utterance verification based on on-line garbage (OLG) models is often adopted as the benchmark method. However, we find its performance can be remarkably improved by fine-tuning. In this study, OLG phoneme likelihood is proposed. It achieves much better performance and efficiency for task independent utterance verification to reject mis-recognition and OOV utterances than the OLG frame likelihood, which is usually employed. A keyword spotter and a voice confirmation system are adopted to evaluate the performance of verifiers on utterances of different natures. We find it is important to evaluate a verifier with OOV utterances of different natures. Instead of using the whole utterance for computing confidence measure, we only use the hypothesized part (i.e. the part recognized as a vocabulary word) of the utterance. Also, researches have been conducted to improve the verification performance and optimize the competing set in computing OLG likelihood. With 5.4% of the within-domain utterances rejected (FRR= 5.4%), the best verifier boosts accuracy of the keyword spotter for Private Branch Exchange from 90.3% to 94.9% and at the same time reject 99.5% of the irregular utterances and 97.4% of the irrelevant sentences (FAR below 2.6%).