If a HMM is trained with sucient data which has been
drawn from all potential operating environments, it should
be able to make optimal decisions. A system operating in
an unfamiliar environment is in essentially the same situation
as a system which has simply been undertrained. In
both cases, the probability models used to classify utterances
poorly represent the rue" distribution which characterizes
the operating environment. If discrete probability models are
used, the result of this is simply that probability estimates
are incorrect | the condence associated with the estimate is
generally unaected. If continuous or semicontinuous models
are used, there will be generally lower probability scores since
the observations are likely to dier from those encountered
during training.