Computing P(W)

What kind of language models are possible for continuous speech?