In Level building, unlike in Viterbi search, HMM for each speech unit
is considered separately [, ]. The search is performed at
various levels for each HMM using the Viterbi algorithm, where
a level corresponds to the position of the speech unit within the
possible sentence. After searching at each level, we find the maximum
Viterbi score over all speech unit models for every time frame t.
The search in next level starts with the winning score of the previous
level at previous time frame. After having completed searches in
levels, the sequence of winning speech unit model at levels
in that order represents the possible sentence
which is speech units long. In order to find the optimum
sentence a maximization is done over within a reasonable range.