Syllables duration distribution problem

[tommigara]

Hi, after the model training I check the syllable duration distribution and it has a weird shape even if the median duration matches the expected length (around 8 frames for recording at 20hz). It seems that the majority of the syllables are very short and the probability for longer syllables dramatically drops, this doesn't happen in the example presented in your guide. Am I doing something wrong? Thank you in advance.
Best,
Tommaso
fitting_progress.pdf

[calebweinreb]

Hi Tommaso,

Such single-frame syllables are usually the result of jitter in keypoint coordinates. The full modeling step is designed to effectively smooth over this jitter, but its possible the hyperparameters are forcing the model to take the keypoint coordinates 'too literally', i.e. not allowing the model to treat them as noise. Did you do the calibration step of the pipeline? I know its currently not working for many users. It might also be useful to identify where the single-frame syllables and occur and note that they keypoint coordinates are doing at those timepoints.

[tommigara]

Ok thank you. Yes, I did the calibration step and it worked. Is there a way to overcome the problem changing appropriately the hyperparameters?

[calebweinreb]

It's possible but it would be nice to have more insight into what's happening. Are you able to plot the keypoint coordinates alongside the syllables to see what happens during those super short syllables?