Hidden Markov models estimation via the most informative stopping times for Viterbi algorithm

We propose a sequential approach for studying the Viterbi algorithm via a renewal sequence of the most informative stopping times which allows us in particular to obtain new asymptotic "single-letter" decoding conditions of equivalency between the Baum-Welch, segmental K-means and vector quantization algorithms of the hidden Markov models parameters estimation which have important applications in speech recognition.