July 18, 2006
state sequences
These plots are designed to help clarify what MFCC frames are being captured by the small-variance components. The top pane is the cepstragram, and the bottom is the maximum activation state sequence. Not that this does not take into account the prior, so it's not the same as the maximum likelihood state sequence. The states are sorted by the determinant of the component, which is plotted along the left edge.
The first plot is for the second-highest hub, and the next plot is taken from the middle of the hub distribution. I chose this pair because they seem fairly similar in that the smallest component is modeling silence at the beginning and ending of the track (which is pretty typical, but not universal).
Posted by madadam at July 18, 2006 07:09 PM