Abstract

This paper describes the use of word triggers in the context of statistical language modeling for speech recognition. It consists of two parts: First we describe the use of trigram models and smoothing in language modeling; smoothing techniques are necessary due to unseen events in training data. In the second part we consider the use of word triggers in language modeling to capture longdistance dependencies. The experimental results presented are based on the Wall Street Journal task. 1. INTRODUCTION The need for a stochastic language model in speech recognition arises from Bayes' decision rule for minimum error rate [1]. The word sequence w 1 :::w N to be recognized from the sequence of acoustic observations x 1 :::x T is determined as that word sequence w 1 :::w N for which the posterior probability P r(w 1 :::w N jx 1 :::x T ) attains its maximum. This rule can be rewritten in the form: arg max w 1 :::w N fP r(w 1 :::w N ) \Delta P r(x 1 :::x T jw 1 :::w N )g ; (1) where P ...

Citations