Towards on-line hidden Markov signal processing

TLDR: The authors propose causal schemes with delay that asymptotically achieve signal model identification and optimal signal statistics and online re-estimation formulae that reduce memory requirements and improve computational processing speed and the adaptive capabilities of hidden Markov model estimation schemes.

Abstract: A commonly used hidden Markov model signal processing scheme that obtains certain optimal signal statistics and estimates is the forward-backward algorithm. This is a noncausal fixed-interval scheme. Repeated application of this algorithm, along with the Baum Welch re-estimation formulae, allows optimal estimation of the signal model parameters, including signal levels, level transition probabilities, and noise statistics. The authors propose causal schemes with delay that asymptotically achieve signal model identification and optimal signal statistics. The key features of these schemes are sawtooth processing and online re-estimation formulae. The intention is to significantly reduce memory requirements and improve computational processing speed and the adaptive capabilities of hidden Markov model estimation schemes. >