Although both Block segmentation and Hanning-3 segmentation
topologies allow integration of multile segment sizes into
a single classifier, the methods may be inadequate to
repredsent data in which multiple sub-event resolutions
co-exist in a single time-serires. For such cases, we need
to consider ``On-the-fly Segmentation".
There are two possibilities for On-the-fly Segmentation.
First, the time-series can be segmented by some kind
of segmentation algorithm into
spectrally-similar segments of various size .
It even be possible to use more than one
algorithm, or one algorithm with multiple
settings, then consider each segmentation as another branch in the
block-segmentation topology (Figure 12.1).
So, ``On-the-fly Segmentation" is really a special case
of block segmentation.
The only problem is the varying segment sizes, and how to pose
the PDF estimation problem for varying-size segments.
One solution to this is to use a segment size independent spectral
representation such as autoregressive (AR) or MFCC.
We touch on this topic again in Section 14.2.8.