Hello,
Thanks for this great work. I find these proposed metrics interesting, and I have been evaluating them.
For very short segments, or short pieces, the cloud-based results can return nan or be indeterminate. Fixing this involves varying the window size to a small number (which makes sense intuitively). Are there any heuristics for reasonably choosing this (alongside the step size) especially since the length of any given piece can be anything from a couple of seconds to more than 5 minutes.
Hello,
Thanks for this great work. I find these proposed metrics interesting, and I have been evaluating them.
For very short segments, or short pieces, the cloud-based results can return nan or be indeterminate. Fixing this involves varying the window size to a small number (which makes sense intuitively). Are there any heuristics for reasonably choosing this (alongside the step size) especially since the length of any given piece can be anything from a couple of seconds to more than 5 minutes.