Problem with scaling in HMM alg.

Hi everyone. I am working on voice recognition sistem on TMS320LF2407 DSP.
I have implemented a Baum-Welch Algorithm using papers: 
    Rabiner, L., �Tutorial on HMMs and Selected Application in Speech
Recognition� and 
    Trebbe, H., �Correction of Errata in Rabiner�s Tutorial on HMM� 

as references and I still have a problem with scaling. I wrote code for
DSP in Code Composer Studio and in C for PC. PC version is working but
same code dont work on DSP. Some probability values in observation matrix
appear to be larger than one. That is probably because underflow or
overflow but I don't know what to do more. 
   


		
This message was sent using the Comp.DSP web interface on
www.DSPRelated.com

Hi,

I not sure if you have done this already, but I the Rabiner's paper
that you mentioned above he goes over a procedure for scaling to avoid
underflow of the alpha and beta coefficients (pg. 272). If you have not
implemented that I would suggest you consider it.

-Ikaro