Improving perplexity measures to incorporate acoustic confusability

Abstract

Traditionally, Perplexity has been used as a measure of language model performance to predict its goodness in a speech recognition system. However this measure does not take into account the acoustic confusability between words in the language model. In this paper, we introduce Equivocality - modification of the perplexity measure for it to incorporate the acoustic features of words in a language. This gives an improved measuring criterion that matches much better with the recognition results than conventional Perplexity measure. The acoustic distance is used as a feature to represent the acoustic characteristic of the language model. This distance is measurable only with the acoustic model parameters and does not require any experimentation. We derive the Equivocality measure and calculate it for a set of grammars. Speech recognition experiments further justify the appropriateness of using Equivocality over Perplexity.