Merlin is an extremely radical development -- the biggest since I started birding 35 years ago! It is a huge change to be able to get quick IDs on calls and songs, unless you are already very expert on the bird vocalisations in an area. From what I've seen so far, it works very well in North America (tested on Eastern deciduous forest birds), and well but not quite as well in Portugal. It will be a good excuse/motivation to do some birding in other countries (e.g. South America) when the models are trained well on those species: the fact that individual locations in the tropics have larger species lists (making ID harder in general) and that vocalisations are a large part of tropical forest birding make it pretty exciting to have it working there. And I'm also looking forward to the app evolving a bit, e.g. confidence scores, labeled sonograms, etc.
I’ve been using Cornell’s BirdNET in Portugal and it’s proven very accurate, even when it’s a euphonic riot of overlapping song - I use inaturalist for photo identification of everything and anything. I’d love it if inaturalist brought in BirdNET or similar - afaik their sound identification is currently driven by humans.
Hi! OK that's good to hear. In Portugal it came up with false positive detections of Hawfinch and Little Owl. Both those species were present in the area (saw them!), but if you think about Hawfinch in particular, it has such a nondescript vocalisation (a quiet, short monosyllable) that it is really easy to imagine that it could be spuriously identified by the neural network (after being allowed as an ID by a prior/filter based on the species plausible in that geographic area at that time of year). I don't mean to sound like I know about the details of their neural net and classification algorithms, or like I know how to do better, but I do suspect that it will improve with some tuning to essentially make it demand stronger evidence for a call like Hawfinch. Little Owl is a bit more distinctive, but I don't really think one was calling when it said it was. I think it also spuriously identified Blue Rock Thrush, but maybe it was right and we just failed to hear/see one! The habitat seemed right. I haven't seen as many misidentifications in Eastern US (a couple of days ago I did think it said Kentucky Warbler for an Ovenbird).
Haven't played with this one yet, but I have messed around with the BirdNET app from Cornell. It's very accurate as long as the neighbors don't have their landscapers mowing the lawn and using leaf blowers right next door.
I just stood up a BirdNet-Pi[1] instance underneath my deck[2] and it's a super fun project. I designed a 3d printable enclosure[3] to mount an rpi in a bird-hut looking thing to try and keep the water off the device. I'm currently using a 4-mic array on a Seeed Respeaker hat which is notionally designed for far-field 360 pickup to be used on voice response devices (alexa et al). It works but it's not great, so next stunt is to try and build some more suitable microphone solutions[4].
It really is an interesting use of TensorFlow at the edge and the somewhat accidental side effect is that I'm learning the birdsongs of all the neighborhood birds which is an extra delight when I'm out and hearing their calls in the broader world.
Nice setup and thanks for the info. I’ve been looking at doing a similar setup myself. I have an outdoor network switch in the same place under my deck. I bet I could stick a Pi close by and power it with PoE.
I've recently come across the haikubox[0] which uses the BirdNET dataset and optionally contributes data back to Cornell.
This is a neat little device perfect for gifts for folks who enjoy birds, but are not technical enough to run a BirdNet Pi or similar. My parents have theirs setup to ping them when certain bird species show up in the yard.
Fun project ran by passionate folks I can't recommend enough.
FWIW, I've been exploring ideas of what bird songs really are. One idea I came up with is "birdglyph" - a distinctive hieroglyph-like picture produced by a birdsong. A few examples:
I recently heard of BirdNet Pi, which lets you record and track what birds are in your local area by the chirps they make ( which are recorded by the RPi): https://birdnetpi.com/
I was about to set one of these up until the mocking birds showed up. Does it compensate for the rotation of those sounds? I know seagulls arent hanging around my yard
I got into BirdNET recently and really enjoyed it. I’m now able to identify the species in our neighborhood and learned some facts about them. It really is pretty amazing app. Building a BirdPi is on my list- I’d like to ID the owls we hear some nights.
My sister had to do this as part of a test for her Master's degree. Now she's working for the government, setting up nets in forests and catching birds to track their migratory habits.
I just installed this the other day on my phone! I was hearing a bird I had never heard before and just had to know. I've been meaning to identify all the local birds, and I had never heard this one and also couldn't see it. It turned out to be a gray catbird. It was wonderfully chatty. I was never a birder but am quickly becoming one after moving out of the city.
This is good software. I've been using it for about a month and I've IDed several local birds I've never been able to see. Merlin is from Cornell as well.
