Hybrid language models and spontaneous legal discourse

The idea of using multiple speech recognizers and alternate language models in a spoken language task is a familiar one. The problems which arise are deciding which language model(s) and/or recognizer(s) to use, or when to change language models. We use the idea of local perplexity of a corpus to determine when to change to an alternate language model and compare this strategy to a number of other strategies in the context of court interactions