While working on the ChoraleBricks Dataset using the libf0.yin implementation with a low tolerance of 10 cents, I observed that the pitch estimation trajectories have certain boundaries, like a local upper bound, which I cannot currently explain.
Below is documentation for the bug fix process.
- Estimates show discrete values (like a plateau).
- There is a 20-cent gap between the reference and estimates.
- Thus, the 10-cent tolerance performs worse than the 25-cent or 50-cent tolerance.
Figures
Zoomed-In
While working on the ChoraleBricks Dataset using the
libf0.yinimplementation with a low tolerance of 10 cents, I observed that the pitch estimation trajectories have certain boundaries, like a local upper bound, which I cannot currently explain.Below is documentation for the bug fix process.
Figures
Zoomed-In
