Hello everybody.

I have been using HTK for a while so far, but I never exploited much of its capabilities.

In these days I wanted to try to use full covariance matrixes for my HMMs. However, I couldn't find any reference regarding the way to initialize them in the prototype.

I tried simply to initialize the prototypes with diagonal covariance matrixes and then use HHEd with the option FC to convert the states into full covariance matrix states, but after doing that, if I use HCompV to initialize the model it returns the error 

ERROR [+5220] CovInvert: [......] not invertible

which according to the HTK book (and the error itself xD) means that the matrix is not invertible.

I think that the error is due to the fact that the covariance matrixes created by HHEd have just the 1st column with values different from 0. So, the information within the matrixes are pretty useless.

So, I'd like to ask you if you know how to initialize the prototype in order to be able to use HcompV afterwards.

Thanks!

Giuliano

--- (Edited on 6/25/2011 9:29 am [GMT-0500] by gbernardi) ---