Abstract

To measure the reliability of recognized words in an ASR, we propose a generalized word posterior probability (GWPP) as the sole confidence measure. This measure is computed efficiently via a word graph with the forwardbackward algorithm or directly with the generalized string likelihoods of N-best strings from the recognizer. The GWPP is a modified word posterior probability where a word event, given all the acoustic observations of an utterance, is measured as a conditional probability. Time registration of the starting and ending frames of a hypothesized word is relaxed, similar to the Baum-Welch model training algorithm, and acoustic and language model weights are optimally adjusted to accommodate instrumental but inaccurate modeling assumptions used in implementing those two models. When tested on the ATR Japanese BTEC speech database, the confidence error rates are significantly reduced as much as 25 % at various operating points. 1

Citations