Speech Research, Problem with continuous Hidden Markov Model

Hi everyone,

I'm creating Hidden Markov Models for musical instrument recognition,
I used the "famous" rabiner paper mainly, and usually I would say,
I'm finished with building the algorithms. Only problem is, when I try
to train my model with a matrix of collected features of test samples,
after 2 to 3 succesful Expectation->Maximization-steps with good
learning, the results of the gaussian densities all go zero, which
stops the baum-welch algorithm.
And also I realized, that the mixture coefficients don't seem to change
in the reestimation process, although I checked all my lines a thousand
times and compared to the written formulas.
Can anyone give me ANY hint for my problem? If some code is needed, no
problem, just mail me or ask here. :)
Thanks in advance for any answer!!

Christian

"kox" < XXXX@XXXXX.COM > writes:


You mean the emission probabilities? What do the means and variances
look like?


Do you understand the theory behind EM reestimation? If so, what do the
posterior probabilities of each component look like?

However, if all your emission probs are zero then the mixture coefs
aren't going to change...


Are you taking steps to avoid overflow, and if so are you sure that they
are working?

When I write related algorithms to this I start with a very small toy
problem that I can write down the answer to, then see if it converges to
the known answer. Have you done that?

Is this commercial or academic? Have you looked at HTK?

This stuff is hard to get right - but once you can do it then you can
solve a load of other problems in a efficient and elegant way, so I
think it's a great tool to have under your belt.

Tell us how you get on.



Tony Robinson