This paper investigates Hidden Markov Models (HMMs) in which the observatio
ns are generated from an autoregressive (AR) model. The overall model perfo
rms nonstationary spectral analysis and automatically segments a time serie
s into discrete dynamic regimes. Because learning in HMMs is sensitive to i
nitial conditions, we initialize the HMM model with parameters derived from
a cluster analysis of Kalman filter coefficients. An important aspect of t
he Kalman filter implementation is that the state noise is estimated on-lin
e. This allows for an initial estimation of AR parameters for each of the d
ifferent dynamic regimes. These estimates are then fine-tuned with the HMM
model. The method is demonstrated on a number of synthetic problems and on
electroencephalogram data. (C) 1999 Academic Press.