Rodrigo reported something similar in 2022. I’m now trying to segment a monodic melody with no background noise and clear pitch changes.
I’m trying to make sense of the noveltyslice parameters with the pitch algorithm.
These settings generate zilions of markers.
moving the threshold just a tiny bit and increasing minslicelength suddenly generates only one marker at the beginning of the file. I also tried the other algorithms. I’m keeping filtersize high to avoid too many small segments.
@jamesbradbury mentioned that it might have to do with NOT normalizing pitch and confidence against each other.
It is not broken, it is just not good. If you take the patch that automatically finds threshold you get it to work… but I never managed in context. For examples like yours I use chroma to very good results (and polyphonic as a bonus 
)
there is one sentence that worries me though:
I suggest to look at noveltyfeature with the same settings to see how that might not do what you might think. In other words, if noveltyfeature doesn’t have a spike where you hear a change, then it won’t work as you want… and the inverse is also true.
I hope this helps.
From speaking with @weefuzzy about this back then, I think this was the main culprit.
In general the ranges are all over the place with the novelty slice algorithms, which is a shame.
If you have SP-Tools or Data Knot installed there’s a visualizer in dk.novelty~ that lets you view the generated novelty envelopes relative to your threshold:
