Although there is a significant body of work on the classification of
time-series [67,68],
existing methods such as HMM [65]
and dynamic time-warping (DTW) [69]
mainly address variations in
the time axis between samples, and rely on a fixed
front-end processing (segmentation and
feature extraction), so do not adequately
address the variations in temporal and spectral
resolution. The fields of time-frequency distributions and wavelet analysis
address these issues for analysis, but are not statistical models.
Although the segmental HMM [70]
is a step in the right direction, it still is based
on fixed front-end processing.
Now, with the introduction of the class-specific feature theorem
[71], [10], [9],
and later the probability density function (PDF) projection theorem (PPT)
[6,3], the freedom now exists
to use different feature sets together in a common
statistical model. The class-specific HMM allows each state in a HMM
uses a different feature [4].
The multi-resolution HMM (MR-HMM) goes even further,
allowing different analysis window lengths
to be used simultaneously
[72,73,74,20] .