evan_tech -- paper: a tutorial on hidden markov models

07:47 pm, 10 Dec 06

paper: a tutorial on hidden markov models

Google Scholar reports that A tutorial on hidden Markov models and selected applications in speech recognition has 4661 citations -- and deservedly so. I've read presentations on what hidden Markov models are a bunch of times, but I hadn't quite grasped it until I read (and thought hard about) this paper. I eventually had a real aha! moment and it now (as many things do in CS) seems kinda obvious. The bulk of the paper is on speech recognition but I wouldn't expect a paper from the 80s to be too relevant to today's state of the art in that field, so I skipped it. But the math behind hidden Markov models doesn't change, so you can read the first nine pages and get quite a bit out of it.

Definitely recommended, though expect (at least if you're like me) to devote a bit of time to understanding it.

(I thought I posted about this earlier but I apparently haven't?)

Anonymous

I think Rabiner's isn't a great intro, compared, e.g., to Jelinek's 1997 book. Something whick essentially no HMM intros say is that you can have posterior input probability distributions associated with arcs or nodes, and a priori distributions with the other. Most intros show one way and practically imply that it is universal, causing much confusion. Modern speech recognition systems have several layers of HMM networks, often using both methods on different layers.

reply to this

evan	What do you mean by "arcs or nodes"? In Rabiner's presentation an HMM concretely is (1) a state transition probability distribution matrix, (2) an observation probability distribution for each state, (3) an initial state distribution. Could you describe what you mean in terms of those parameters? (This stuff is still all pretty new to me.)
reply to this

trochee

guessing here: I think the first commenter in this thread probably means "transition probabilities" and "emission/observation probabilities" for "arcs" and "nodes" respectively. Some discussions of HMMs put the transition probabilities "in" the hidden states, but properly it makes more sense to put all the probabilities on the arcs, as I see it. (the matrices and distributions you mention correspond there).

reply to this