I have implemented Segmental K-means algorithm for better HMM parameter
initialization (Rabiner: A tutorial on Hidden Markov models and
selected applications in speech recognition 1989). But I have some
1- After some iteration of running Viterbi algorithm for each training
sample of the word and segmenting it according to the optimal path,
observation vectors of some states are less than number of predefined
Gaussian mixture models. So the algorithm can not reestimate mean
vector, covariance matrix and mixture gains of these states.
2-During HMM learning process with Segmental K-means, some times log of
likelihood is decreased suddenly and the algorithm can not converge.
Any comments would be helpful.