Hidden Markov model - zero probability problem

Started by pete...@st.fri.uniza.sk December 7, 2010
I am implementing a speech recognition system based on Hidden Markov models. I have finished the Baum-Welch algorithm for model training - and tested it on some primitive models.

I work with discrete HMM, so I am using a code book with 50 output symbols. For speech parametrization I am using 12 MFCC coefficients for every block of sound(512 sound samples - 30 ms).

Now when I am training the model on several speech samples (10 times the same word is recorded). I am getting the emission probability of some of the symbols from code book equal ZERO in all model states (after training). Thant wouldn't be a problem. But sometimes during recognition process I get a symbol, which has for this particular model probability of emission in all states equal 0. (Even when it is the same word). That's why, the total probabilty for this word to be generated by this model is equal ZERO too. (and that is obviously wrong)

Is there any posibility to avoid this situation? (Maybe more sound samples during model training - to cover all the posibilities how the word can be spoken by the speaker). I haven't seen any note about this problem in articles about HMM and speech recognition.

I hope I explained my situation clear enough.

Thanks in advance.