palonso
commented
6 years ago Great, seems very useful. We can implement the proposed aggregated confidence measure and compare it with the current strength output (relative difference between the first and second best key candidates) to find out which one is more informative about difficult tracks.
The paper "A confidence measure for key labelling" proposes two confidence measures, keyness, and stability, both are useful to identify difficult tracks for key estimation.
Both measures are based on chroma and should be straightforward to implement in Essentia. We can also consider adding a frame-wise keyness curve too (may be useful to describe the structure of the track).
We should be able to compute these measures using the raw non-normalized HPCP vectors.