Utterance verification using word-dependent thresholds based on probabilistic distributions of phone-level log-likelihood ratio

Abstract

This paper suggests word voiceprint models and word-dependent thresholds using distributions of phone-level log-likelihood ratio and duration to verify the recognition results obtained from a speech recognition system. Word voiceprint models have word-dependent information based on the distributions of phone-level log-likelihood ratio and duration. Thus, we can obtain a more reliable confidence score for a recognized word by using its word voiceprint models that represent the more proper characteristics of utterance verification for the word. There are many conditions to affect the decision of thresholds in utterance verification system. In this paper, we propose an algorithm to generate the threshold for each word using distributions of phone-level log-likelihood ratio. For each word, confidence measure obtained from phone-level log-likelihood ratios has different distribution, so we need to adapt the different threshold for recognized word. The algorithm using word voiceprint models shows that relative reduction in equal error rate is 16.9% compared to the baseline system using simple phone log-likelihood ratios. And the word-dependent thresholding method shows that the relative reduction in equal error rate is 14.6% compared to the baseline system using one global threshold.