Current state-of-the-art speaker-independent continuous speech recognizers are able to achieve word recognition rates in excess of 94 percent with lexicons of 1000 words of less using grammars or language models with perplexity 60 or less. Performance of these systems decreases rapidly as the perplexity of the grammar increases. As we allow users more flexibility in interacting with speech recognition and understanding systems, the lexicon size and perplexity increases more than an order of magnitude. Allowing spontaneous speech further increases perplexity and introduces many additional sources of variability current speech systems are unaccustomed to dealing with. However, natural language knowledge sources can be used to dynamically decrease perplexity and allow more natural interaction given the current state of the technology.