For illustration, we created an MR-HMM with three sub-classes:
background noise, low-frequency (LF) burst, and high-frequency (HF) burst.
Feature extraction consisted of autoregressive (AR) analysis
with model order depending on the sub-class
(See Section 9.4.10).
In Figure 14.3 (top), we see a spectrogram
exhibiting an LF and an HF noise burst.
In the center of the figure are the a posteriori
proxy state probabilities
These are summed, collecting all wait states in each sub-class,
to obtain the sub-class probabilities (bottom).
Note that at some times, there is considerable uncertainty about
which trellis path is in effect as evidenced by
probability sharing. But, the path probabilities sum
up to provide a crisp decision about the sub-class.
Illustration of MR-HMM using synthetic data.