As part of an exercise/demonstration for a course, last night I ran Neville Ryant's second-best speech activity detector (SAD) on Barack Obama's Weekly Radio Addresses for 2010 (50 of them), and George W. Bush's Weekly Radio Addresses for 2008 (48 of them). The distributions of speech and silence durations, via R's kernel density estimation function, look like this:
Then I wondered what the 2D distributions would look like. So courtesy of R's two-dimensional kernel density estimation, here they are:
The durations of speech segments are on the horizontal axis, and the durations of the immediately following silence segments are on the vertical axis.
Or maybe they're interstellar gas clouds?
Anyhow there's clearly some structure there. And it's neat that it just took a few minutes of computer time and a few lines of R to create the pictures.
Update — contour plots make it a little clearer what's going on:
See also "Poetic sound and silence", 2/12/2016