DrCoffey
commented
4 years ago Usually, the better the contouring, the better the clustering. The contour is based on the tonality slider, so padding doesn't change anything. Unfortunately, you can only pick 1 value for tonality for the entire dataset. Whatever the slider is set to when you hit the clustering button will be used. I would open a representative file, set the slider, and then cluster everything.
We wanted to add file or call based tonality, but didn't have time to add the feature. If anyone on your team can code and wants to add it, feel free to send us a pull request!
More thoughts: Are pup calls less tonal than adult calls? Does the constant 58 kHz bad mess with the contour?
Newer clustering method (I am not sure if they clustered appropriately with DeepSqueak, so it may be better or not depending): https://www.biorxiv.org/content/10.1101/811661v1.full.pdf
LWilkinKrug
Hi Kevin,
I've been trying to cluster about 55,000 rat pup calls. All calls were hand-checked in Raven. Initially, I ran unsupervised clustering with the standard settings (Weight: 1/1/1, max clusters: 100, reps: 3), but have since tried several other combinations without much luck. While some clusters look pretty good, others seem to be a mix of e.g. step up, step down, multi-step, or upward ramp, downward ramp, inverted U, short.
I used k means clustering in R, but the clusters didn't make sense either; plus R doesn't provide a visual cluster overview i.e. no way to easily check the clusters.
Because I have a permanent noise line from the air vent system in the lab at about 58 kHz, I deduced that perhaps this was interfering with the contour output, so I manually adjusted the contour threshold for calls in 40 files, thinking that this should allow me to get clean clusters—but it didn't.
I tried the WrightClassifier, but that wasn't usable either.
I was wondering if DS adds padding around the calls which might change the contour and subsequently the slope?
Thanks for any ideas and help you can offer.
Warm regards
Linda