Using confidence measures and domain knowledge to improve speech recognition

In speech recognition domain knowledge is usually implemented by training specialized acoustic and language models. This requires large amounts of training data for the domain. When such data is not available there often still exists external knowledge, obtainable through other means, that might be used to constrain the search for likely utterances. This paper presents a number of methods to exploit such knowledge; an adaptive language model and a lattice rescoring approach based on Bayesian updating. To decide whether external knowledge is applicable a word level confidence measure is implemented. As a special case of the general problem station-to-station travel frequencies are considered to improve recognition accuracy in a train table dialog system. Experiments are described that test and compare the different techniques.