Sparrow Classification

41 views
Skip to first unread message

Peter Keyel

unread,
Sep 30, 2020, 9:52:50 PM9/30/20
to birdvox
I think BirdVox Detect and Classify are really neat, and really helpful for someone just starting NFCs like myself. Since I am very unfamiliar with NFCs, how do I verify NFCs? I'm familiar with the OldBird species glossary, but I can't find matches to any of the calls that I hear. Is there a way to see likely runners-up or potential for uncertainty? I presume if there's a hit with 63% confidence and one with 60% confidence, it only lists the top hit? 

I think most of the calls have been CCSP, based on feedback from the Facebook NFC group. BirdVox gave me slightly different results. I tried out BirdVox Detect on a short (~13 min) recording taken earlier this week (checklist: https://ebird.org/checklist/S74263231 ). When I ran BVD on the whole unprocessed clip (threshold =10), I got nothing. Running BVD on the clip filtered (3k cutoff) to eliminate worst of the traffic and other noise and amplified gave me 3 hits, 2 of which were real. It IDed the clips as SAVS and WTSP with 58% and 68% confidence. Running BVD on the isolated (filtered) clips gave me WTSP for both, with 54% and 64% confidence, respectively. Feedback from the NFC Facebook group suggests CCSP for at least one of the clips, which are *way* more abundant on the ground in Lubbock now. That's similar to feedback I've gotten on other clips with what look to be similar calls (https://ebird.org/checklist/S73794113 ).

Benjamin Van Doren

unread,
Oct 5, 2020, 3:14:55 PM10/5/20
to Peter Keyel, birdvox
Hi Peter,

Thanks for getting in touch with us. It’s great to hear that you are getting started with NFCs. The current versions of BirdVox Detect and Classify do not include Clay-colored Sparrow, so unfortunately this species is not an available option for the classifier at this time. One of our top priorities is to add more species to the classifier, so this should change in a future release. I should also mention that the recordings on your eBird checklist do not strike me personally as typical CCSP calls, but I may not be familiar with the full breadth of variation in this species.

I believe at the moment that it is not easy to extract “runners-up” from the output, but this is something we have also been discussing, and it is very helpful to know that it would be useful to users. Others on the team may be able to chime in with additional input and advice related to threshold-setting.

Thanks again for your valuable feedback - I hope this helps!
Benjamin


Benjamin Van Doren, Ph.D.
Postdoctoral Associate, Cornell Lab of Ornithology
Cornell Presidential Postdoctoral Fellow

On Sep 30, 2020, at 9:52 PM, Peter Keyel <labtro...@gmail.com> wrote:

I think BirdVox Detect and Classify are really neat, and really helpful for someone just starting NFCs like myself. Since I am very unfamiliar with NFCs, how do I verify NFCs? I'm familiar with the OldBird species glossary, but I can't find matches to any of the calls that I hear. Is there a way to see likely runners-up or potential for uncertainty? I presume if there's a hit with 63% confidence and one with 60% confidence, it only lists the top hit? 

I think most of the calls have been CCSP, based on feedback from the Facebook NFC group. BirdVox gave me slightly different results. I tried out BirdVox Detect on a short (~13 min) recording taken earlier this week (checklist: https://ebird.org/checklist/S74263231 ). When I ran BVD on the whole unprocessed clip (threshold =10), I got nothing. Running BVD on the clip filtered (3k cutoff) to eliminate worst of the traffic and other noise and amplified gave me 3 hits, 2 of which were real. It IDed the clips as SAVS and WTSP with 58% and 68% confidence. Running BVD on the isolated (filtered) clips gave me WTSP for both, with 54% and 64% confidence, respectively. Feedback from the NFC Facebook group suggests CCSP for at least one of the clips, which are *way* more abundant on the ground in Lubbock now. That's similar to feedback I've gotten on other clips with what look to be similar calls (https://ebird.org/checklist/S73794113 ).

--
You received this message because you are subscribed to the Google Groups "birdvox" group.
To unsubscribe from this group and stop receiving emails from it, send an email to birdvox+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/birdvox/467027eb-fc77-4901-bcd1-ef4ac41de914n%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Peter Keyel

unread,
Oct 5, 2020, 11:43:23 PM10/5/20
to Benjamin Van Doren, birdvox
Hi Benjamin,

Thanks for the feedback! I did not realize Clay-colored Sparrow was not in the program. Is it safe to assume that it's primarily Eastern species currently in the classifier?

Especially since I'm new to NFCs, I'm happy to correct the calls if they are wrong, or if there is something that fits better. Bill Evans and John Kearney were the people who suggested/agreed with CCSP on some of the calls that I posted. Based on those IDs, I tried to match ones that looked/sounded similar to them. Since there is no 'Field Guide to Nocturnal Flight Calls' out yet, I can't find an easy key for narrowing NFCs, Macaulay doesn't yet allow search by ebird protocol type or time range to easily browse others' NFCs, and the Oldbird/other online information is spotty, I had planned to use BirdVox to help learn the NFCs that I heard.

Cheers,
Peter

--
"This is the world we live in
And these are the hands we're given
Use them and let's start trying
To make it a place worth living in."
-Genesis, "Land of Confusion"

Benjamin Van Doren

unread,
Oct 15, 2020, 10:43:25 AM10/15/20
to Peter Keyel, birdvox
Hi Peter,

Yes, we currently only have northeastern species in the classifier, but it is a priority to expand to more species. Therefore I wouldn’t recommend using BirdVox primarily as a learning tool for now. BirdVox Detect should still be useful for extracting calls from recordings regardless of species ID, as the Detect and Classify components run separately. 

Personally I still refer to the OldBird guide quite a bit (see http://oldbird.org/pubs/fcmb/species/sparrows/ccsp/ccsp.htm), but I also share your desire for something more complete.

Thanks again for your feedback, and we’ll keep you posted on further developments!
Benjamin

Vincent Lostanlen

unread,
Oct 16, 2020, 9:10:05 AM10/16/20
to bir...@googlegroups.com

Dear Peter and all,

Thank you for your feedback. I concur with Benjamin in that CCSP is not yet among within our taxonomy, so, at best, it will appear as "OTHE" (other) in the output.

Our short-term goal is to make the output format of BirdVoxClassify richer in terms of taxonomical depth: order, family, and species. So the output label will be "Other sparrow", meaning "Definitely an American sparrow, but none of the species of sparrows in the currently supported checklist".

See this paper (Cramer et al. ICASSP 2020) for our currently supported checklist and the taxonomical neural network classifier (TaxoNet): http://www.lostanlen.com/pubs/cramer2020icassp/

Our mid-term goal is to expand the taxonomy beyond the current checklist of 14 species. Benjamin, Jason, and myself have concrete plans to enact this: we need to retrain the classifier on more diverse calls, which we're collecting and aggregating at the moment.

The long-term goal is called the "open-set scenario", i.e., in which we allow the user to add new species at runtime. This is an open research problem at the moment, so i'm not able to give you a precise timeline for it.

I hope this helps!

Sincerely,

Vincent.

Reply all
Reply to author
Forward
0 new messages