In this letter, we propose a new algorithm for nonparametric estimation of
hidden Markov models (HMM's), The algorithm is based on a "wavelet-shrinkag
e" density estimator for the state-conditional probability density function
s of the HMM's. It operates in an iterative fashion similar to that of the
EM reestimation formulae used for maximum-likelihood estimation of parametr
ic HMM's. We apply the resulting algorithm to simple examples and show its
convergence. The proposed method is also compared to classical nonparametri
c HMM estimation based on quantization of observations ("histograms") and d
iscrete HMM's.