Spring Peepers

28 views
Skip to first unread message

John Kearney

unread,
Mar 20, 2021, 8:05:28 PM3/20/21
to birdvox
Hi All. Spring peeper season will soon be upon us. Last spring, BirdVoxDetect would produce thousands upon thousands of detections of spring peepers. Adjusting the threshold did not alleviate the problem since most detections were still peepers. Is there any work around for this problem within BirdVoxDetect? The only solution I can think of is filtering out all calls in the peeper frequency band by my AudioMoth. However, all birds within this range, namely, thrushes will also be filtered out. Thanks for considering this issue.
John

Justin Salamon

unread,
Mar 22, 2021, 5:52:38 PM3/22/21
to John Kearney, birdvox
Revenge of the spring peeper!

In "Towards the Automatic Classification of Avian Flight Calls for Bioacoustic Monitoring" we found that spring peepers can be a serious confounding factor for certain species, e.g. Swainson's Thrush. We'd have to double check, but it is possible that the BVD species classifier wasn't exposed to many samples of spring peepers, thus the false positives.

I believe we should have some ~5k annotated peeper calls, so we could look into re-training the species classifier where we expose it to all of these calls to make it robust against them. But that may take us some time to get to.

I'm afraid I don't have a better interim solution than the one you have mentioned (filtering by frequency), which would alleviate the problem at the cost of losing all calls in the overlapping range.

Best,
Justin

On Sat, Mar 20, 2021 at 5:05 PM John Kearney <j.f.k...@gmail.com> wrote:
Hi All. Spring peeper season will soon be upon us. Last spring, BirdVoxDetect would produce thousands upon thousands of detections of spring peepers. Adjusting the threshold did not alleviate the problem since most detections were still peepers. Is there any work around for this problem within BirdVoxDetect? The only solution I can think of is filtering out all calls in the peeper frequency band by my AudioMoth. However, all birds within this range, namely, thrushes will also be filtered out. Thanks for considering this issue.
John

--
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/06194f3b-9f00-4bce-9a87-c0c8af3ca8fdn%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

John Kearney

unread,
Mar 23, 2021, 11:00:12 AM3/23/21
to birdvox
Thank you for your response and the link to your paper, Justin. Last spring I had 25,000 or more spring peeper calls per night of recording. I sent one file to Andrew Farnsworth last year but I could send you one too if you wish. I think the main problem is that a spring peeper call is almost indistinguishable from a distant Swainson's Thrush flight call. The only way I can with any accuracy tell them apart is by listening to the recording or viewing a segment of the recording so one can hear or see the cadence of the spring peeper calls. Is cadence something you can build into your model?
Thanks!
John

Vincent Lostanlen

unread,
Mar 23, 2021, 11:10:10 AM3/23/21
to birdvox

Hello John, Justin and all,

> The only way I can with any accuracy tell them apart is by listening to the recording or viewing a segment of the recording so one can hear or see the cadence of the spring peeper calls.

I agree. Telling apart one from the other is more a matter of call rate ("cadence") than of spectrotemporal pattern.

We don't have a model for call rate yet in BirdVoxDetect. The species classifier make a prediction based on a small amount of context (150 milliseconds), encompassing a single call.

We have published a version of our flight call detector which is "context-adaptive" in the sense that it takes into account the background noise at a duration of 30 minutes. This is mostly for dealing with variations in insect noise.

But the call rate of spring peepers range between those two time constants (150 ms and 30 minutes), and thus beyond the scope of BirdVoxDetect. For this reason, we expect BirdVoxDetect to perform better during the fall than during the spring.

Distinguishing flight calls from frog calls is an open research question and one that we'd like to tackle in the long term; but it's unclear why. A straightforward approach would be to replace the convolutional neural network (CNN) by a convolutional _recurrent_ neural network (CRNN), but these are much slower than CNN, both at training time and at prediction time. So i'm not convinced that this is an approach worth investigating. For the time being, we're focusing on bird-vs-bird classification which is already a challenging problem when posed at the level of individual flight calls.

I hope this helps and thanks again for your interest in BirdVox


Sincerely,

Vincent.

John Kearney

unread,
Mar 23, 2021, 3:31:44 PM3/23/21
to birdvox
Thanks for your contribution to this discussion Vincent. For those who need to do spring bird studies, it seems that using a recorder with a frequency filter is a good option if you can do without the birds in the peeper frequency range. Otherwise, it seems using a detector in other bioacoustics software is the only option. I found that increasing the minimum separation distance in the Raven Pro band limited energy detector would greatly reduce the number of peeper detections but not without reducing some of the bird detections as well. If I remember correctly, the original Old Bird "Thrush" detector had a feature where the software would stop detecting during a long series of repetitive calls. I don't know if the updated "Thrush" detector has this feature. In any case, I am grateful for the work of the BirdVox team. I now use it in all my projects with this possible exception as explained by Vincent. I should also mention that I use BirdVoxDetect for morning recordings as well as nocturnal monitoring during the migration season.

Justin Salamon

unread,
Mar 23, 2021, 4:55:15 PM3/23/21
to John Kearney, birdvox
Thanks John, it's great to hear BirdVox is useful in your work! 

I just thought of one more option: you could run BVD on your data without filtering, and then, if the peeper calls are regular, the cadence should be detectable via e.g. cross-correlation (or fourier analysis) applied to the time series of detection counts. Then you could discard calls that occur on the cadence, and keep calls that occur "off-beat". It would probably require some experimentation to get right.

Cheers!
Justin

p.s. kindly remember to cite [1] in publications that make use of BVD, or [1] and [2] if you're using version 0.4 and up.

[1] Vincent Lostanlen, Justin Salamon, Andrew Farnsworth, Steve Kelling, and Juan Pablo Bello, "Robust Sound Event Detection in Bioacoustic Sensor Networks", PLoS ONE 14(10): e0214168, 2019. DOI: https://doi.org/10.1371/journal.pone.0214168.

[2] Jason Cramer, Vincent Lostanlen, Andrew Farnsworth, Justin Salamon, and Juan Pablo Bello, "Chirping up the Right Tree: Incorporating Biological Taxonomies into Deep Bioacoustic Classifiers", In IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Barcelona, Spain, May 2020.


For more options, visit https://groups.google.com/d/optout.


--
Justin Salamon | Adobe Research | www.justinsalamon.com

Joe Gyekis

unread,
Mar 23, 2021, 5:56:42 PM3/23/21
to Justin Salamon, John Kearney, birdvox
I'm a strong believer that the shortcut to solving these problems is just like normal conference microphones have built-in an extra four or six microphones for less than $100 extra price, sound source localization can filter noises by direction just like our ears do and easily recognize what's coming from the sky above and what's coming from the ground below or at least near the horizon. Directional filtering on the original incoming signal (prior to detector analysis) will severely reduce detections of birds calling from low angles but will also reduce a wide variety of extraneous noises (bugs, mammals, frogs, humans) by a massive margin and will be worth it. Anybody who wants to work on this side project with me over the summer (~May 15 to Aug 15) send me a message! I am a horrible newb on this kind of technical problem solving but I feel this is achievable and simply must be done, at least at proof of concept level.

Wim van Dam

unread,
Mar 23, 2021, 6:16:53 PM3/23/21
to Joe Gyekis, Justin Salamon, John Kearney, birdvox
Joe: That's a good idea. If you take your approach and do it precise enough you should also be able to determine the altitude of the calling bird. Another potential hardware solution would be to use a parabolic microphone to focus on a specific direction. You know, something like this: https://en.wikipedia.org/wiki/Acoustic_mirror

Wim van Dam
Solvang, CA



Joe Gyekis

unread,
Mar 23, 2021, 6:22:54 PM3/23/21
to Wim van Dam, Justin Salamon, John Kearney, birdvox
Altitude can't be estimated with any accuracy without a gigantic microphone array to triangulate with, talking about multiple microphones on booms at least 10 m tall plus additional microphones on the ground beneath those booms. 

But simple up vs down direction of arrival is comparatively easy, there are implementations of that built into any normal priced conference microphone system.

Justin Salamon

unread,
Mar 23, 2021, 6:43:00 PM3/23/21
to Joe Gyekis, Wim van Dam, John Kearney, birdvox
Thanks Joe, that's a great point.

Now that you mention it, we actually discussed exploring direction in the context of BirdVox, but it ended up being out of scope (there's only so much we can cover).

But yes, I fully agree that if the problem can easily fixed at the point of capture, that's going to be easier and more robust compared to trying to tease out the calls post-hoc. A quick online search suggests there you can get an 8-mic USB array for ~$100.

Best,
Justin


For more options, visit https://groups.google.com/d/optout.
Reply all
Reply to author
Forward
0 new messages