Unfamiliarity: A Field Mark
On the xeno-canto forum, people frequently post unidentified bird sounds from all around the world. I can’t be of much help on the mysteries from Europe or Africa or India, but if I see an unidentified sound posted from North America, I check to see if it’s something I recognize.
Last year Bates Estabrooks uploaded an odd call he’d recorded in Tennessee. It didn’t exactly resemble anything I’d heard before, but I told him it was probably a Tufted Titmouse. Several other people on the forum thread agreed. Here’s the sound:
Bates wrote back with a perfectly legitimate question: what made it a Tufted Titmouse? If it didn’t exactly resemble anything I’d heard before, how could I be sure?
Here’s what I told him:
It’s simultaneously pretty simple, and frustratingly difficult. Here’s the reasoning I used:
- It’s very high-pitched (almost entirely above 6 kHz).
- It’s medium-complex (2-syllabled).
- I’ve never heard anything quite like it (i.e., it doesn’t fit for any of the other birds that regularly give very high, 2-syllabled calls, like Brown Creeper or Golden-crowned Kinglet).
“I’ve never heard anything quite like it” is actually an excellent field mark for Tufted Titmouse (and a couple other species, such as Blue Jay). The frustrating part is that it takes time to get to the level of experience where this field mark is helpful.
When it comes to identifying bird sounds, unfamiliarity can actually be a very useful mark. Of course, it becomes more useful as you learn more bird sounds (and can therefore rule them out). The legendary Ted Parker knew almost all the bird sounds in the Western Hemisphere — so when he heard something unfamiliar in a tape from Bolivia, he postulated that it must be a species new to science (and it appears he was probably correct). Most of us are never going to arrive at this level, of course.
Note that I’m not talking about unpredictability, which is a field mark for mockingbirds, thrashers, catbirds, or Yellow-breasted Chats. Unpredictability is different from unfamiliarity. Unpredictable birds might sing unfamiliar notes or phrases, but then they quickly move on to another type of sound. I’m talking about birds that repeat the same unfamiliar sound over and over.
When I was a young birder in eastern South Dakota, I soon realized that most weird, unfamiliar bird sounds came from Northern Cardinals. When I moved to Massachusetts for college, I learned to bet on Tufted Titmouse as the source of a wacky sound. In graduate school in western Oregon, the vocal tricksters were Hutton’s Vireo and Bewick’s Wren. And of course, two other consistently unfamiliar birds are Blue Jay (especially in the wide variety of so-called “pumphandle calls“) and Red-winged Blackbird (especially the many whistled variations of the alert calls).
I’m curious about other people’s experiences. What birds have you learned to identify by their unfamiliarity?
3 thoughts on “Unfamiliarity: A Field Mark”
I had that experience with a House Wren today – although it’s the first time I have witnessed it with this species. It was singing a song I didn’t recognize (significantly different from normal House Wren songs). I even played back through all the songs on my phone – but nope, it had invented a new one. About 10 minutes after repeating the unfamiliar song every few seconds (i.e. many times) it threw out a typical House Wren song, then flew away. As if to say, “got you”…
I have often thought that we birders who cut our aural teeth in the Northern Hemisphere have a bit of a leg up on those who learn their skills in a place like the Neotropics because we have to learn bird songs independent of specific pattern (a specific pattern being, for example, what one does when searching through XC for an exact match to a mystery recording). Allow me to clarify: nearly all of the passerines in North America, and all in Europe, are oscine, so they have dialects thanks to learning error and many have multiple song types within an individual’s repertoire. In tropical America, most passerines are either suboscines (which overwhelmingly have inherent songs with little or no local variation, although many–such as furnariids–do have multiple song types) or are things like tanagers, which only occasionally have truly distinctive voices one can learn (think of all the “seets” and “tsits” of Tangara tanagers and Dacnis and related things… most of which are bafflingly unidentifiable).
So learning sounds based on “texture” (for lack of a better word, sort of the audio version of GIS) and general pattern, rather than specific pattern, seems to make it that much easier to identify vocalizations that are strongly stereotyped because of genetic programing. It also allows us to keep our mind open when listening to an unfamiliar sound that we can often categorize into a collection of possible suspects almost immediately based on texture and general pattern. When I hear a woodcreeper vocalization that is unfamiliar to me, I still seem able to categorize it as “woodcreeper” due to its texture. How, I can’t exactly explain, but I believe it is because I have learned how to recognize a wide range of distinctive sounds into categories such as “Song Sparrow” or “Yellow Warbler” or “Bewick’s Wren” or “Tufted Titmouse” thanks to my early learning experience in North America.
In this era of increasing efforts to create computer programs to identify bird sounds, I wonder how they will handle such variation and unknowns?
Regarding how modern “machine intelligence” or “deep learning” algorithms would learn to distinguish calls, it is through training. Specifically, supervised training. i.e. the algorithm is fed a collection of recordings of known songs and calls for a particular species. If (big “if”) the data is comprehensive and representative of the full range of vocalizations then the algorithms (once trained) can achieve high accuracy against new calls.
The more good data that can be fed to these algorithms, the better. Time of day, date, habitat call is recorded in, GPS location – all will provide clues. As for the call itself, it is ideal if the call / song is isolated from any longer recording – but these algorithms can do surprisingly well even if it is not (although the size of the training data set would have to be larger). Typically a frequency analysis would be run on the audio. All of these features are then usually put in a 1 dimensional vector that is used to train the neural net.
How this is done deterministically is unclear. Neural nets in particular are still black box in terms of how they work, even to researchers. But work they do. The features selected by these neural nets as the key areas of interest / differentiators sometimes correspond to what we humans might identify, but often do not. It’s intriguing.
All of this to say – yes, it can be done. It is being done to a limited extent today, but you can expect excellent automatic bird call ID sooner rather than later. It will do wonders for ornithology …
Comments are closed.