Something about HMM　training, say, Baum Welch algorithm.

If the HMM may have 1000 possible values for observations, but my training data may only contain 800 different values of observations (although I may have 2000 observations in total, but 1200 observations are equal to at least 1 observation of the 800 observations)

which means, the training might be insufficient.

What can I do in such cases?

Best Regards
JIA Pei