Abstract

Acoustic model estimation using the Bayes risk criterion involves computation of the minimum edit (or Levenshtein) error between the correct sequence of words (or phonemes) and each member of a set of alternative hypotheses. In the context of large vocabulary speech recognition, the set of alternative hypotheses is large, rendering the computational side of this process prohibitively expensive. So practical implementations of Bayes risk minimization use two approximations.

The first approximation is to use a pruned set of alternative hypotheses instead of the full hypothesis space. The second is to use a computationally-inexpensive approximation to the Levenshtein error for each member of this pruned set.

This talk focuses on the latter approximation. A previously introduced approach (D. Povey, PhD thesis, 2003) uses alignments of the correct and alternative sequences to approximate the Levenshtein error between the sequences. We examine the accuracy of this approach, highlight some limitations, and propose alternative alignment-based approximations. We then present experimental work quantifying the impact of these alternatives upon MBR parameter re-estimation.