Training about HMM with more than one observations #21

In wiki, it was wrote:
    * Hidden Markov Model learning algorithms
    * The learning is based on a set of observation sequences (not only one).
        # K-Means learning;
        # Baum-Welch learning with scaling (meaning that a learning based on long observations sequences don't generate underflows). 

Can you give some tips about how use those two methods(K-Means learning and 
Baum-Welch learning with scaling )to learn parameters in hmm; or you can give 
me some meterials about it.

Thanks.