We employed the implementation of hidden Markov models provided by HTK
(the HMM Tool Kit) [YKO$^+$]. Each class is modelled using the
same topology. The following number of states are considered: 3, 5,
10, 20; and the following topologies: left-right, left-right with 1
skip allowed and ergodic. Each state in a left-right HMM allows
transitions only to itself or the next state - hence each state must
be visited. to A 1-skip model is a left-right model that also allows
the transition from state 
to state 
, effectively allowing to
skip certain states. Ergodic HMMs allow all possible transitions.
HMMs that include the raw data as well as both the raw data and the
first derivative are considered. As with the naive segmenter, unless
otherwise indicated the best results, over all the topologies,
states and both raw data and raw data with first derivative (a total
of 32 possibilities) were reported. It should be noted this is biased
in favour of hidden Markov models.