Also included is an example based on the one from the Wikipedia page for a simple 2-state model with 4 observations. To paraphrase:
Bob tells Alice his daily activities (observations) and Alice wants to determine the most likely weather each day (states). Since Alice lives far away, the weather is unknown to her (hidden). Assuming that the states behave as a Markov process, and the observations have a statistical dependency on the states, the Viterbi algorithm can find the most likely weather pattern (path).
foward_viterbi.m contains a bug not in the original python version. If the output is longer than 3 states then this line throws an error. The 2nd index to emit_p should be the position of the output token in the vector of possible output states, not the index into the observation vector.
p = emit_p(source_state,output) * trans_p(source_state,next_state);
Updates
12 Jan 2010
Fixed the problem found by Justin. To make things simpler, observations are now numbered, rather than allowing strings.