Computing P(W)

What kind of language models are possible for continuous speech?

Conditional independence of observations

The HMM assumes each observation is independent of all the others, given the state that emitted it.