macaodha
Classify completely ignores feeding buzzes
I only saw the feeding buzz because I looked in the space between classified calls, and the BTO pipeline said there was one, but not where. Elsewhere, I have only spotted the buzz because of random variation in the visible call interval.
Hi @markgloverswaybtinternetcom,
Yeah, this is a known limitation and a bit of a design choice. At the moment, we don't have enough labelled data for feeding buzzes to reliably detect them, so we've chosen not to include them in the current model.
We'd definitely welcome any labelled examples of feeding buzzes, or even social calls (though that’s an even trickier problem). With more data, future versions of the model could potentially include detection for these.
Thanks for the reply with your existing recordings you probably have a lot of feeding buzz samples. Soprano buzz are slightly beyond your stft resolution, mine is very slightly better (above). The interval is the give-away, the echo-location gives the species. Tricky as the interval is closer that a lot of the echo's I see, many of which confuse batDetect2 Regards Mark Glover
------ Original Message ------ From "Santiago Martinez Balvanera" @.> To "macaodha/batdetect2" @.> Cc "markgloverswaybtinternetcom" @.>; "Mention" @.> Date 02/06/2025 19:10:58 Subject Re: [macaodha/batdetect2] Classify completely ignores feeding buzzes (Issue #49)
mbsantiago left a comment (macaodha/batdetect2#49) https://github.com/macaodha/batdetect2/issues/49#issuecomment-2931875591 Hi @markgloverswaybtinternetcom https://github.com/markgloverswaybtinternetcom,
Yeah, this is a known limitation and a bit of a design choice. At the moment, we don't have enough labelled data for feeding buzzes to reliably detect them, so we've chosen not to include them in the current model.
We'd definitely welcome any labelled examples of feeding buzzes, or even social calls (though that’s an even trickier problem). With more data, future versions of the model could potentially include detection for these.
— Reply to this email directly, view it on GitHub https://github.com/macaodha/batdetect2/issues/49#issuecomment-2931875591, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZP7OYFPOQXMDCYVAYHFKV33BSHTFAVCNFSM6AAAAAB464QEVKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSMZRHA3TKNJZGE. You are receiving this because you were mentioned.Message ID: @.***>
-- This email has been checked for viruses by Avast antivirus software. www.avast.com
Hi I think that training with single calls for feeding buzz would not be recommended as you would probably triple the number of false calls in noise. However recognising a small group of calls might just work. This only applies to species that alter their frequencies for feeding. The feeding buzzes will be of lower amplitude, so most will be lost in the noise. I tend to colour code the interval between calls (from the csv files): white closer then normal echo-location, dark grey for longer than normal.
The difference between batDetect2 and the BTO Acoustic Pipeline might just be due to boxing/labelling policy. For human labellers the choice of spectrogram colour scheme and spectrogram parameters greatly affects the boxes. The labellers see the surrounding context, so might tightly label calls that they could not hope to recognise in isolation.
I have noticed that I am seeing higher frequencies in the book samples (that use 384 kHz sampling) than the maximums documented in the book. Regards Mark Glover
------ Original Message ------ From "Mark Glover" @.> To "macaodha/batdetect2" @.> Date 02/06/2025 23:07:17 Subject Re[2]: [macaodha/batdetect2] Classify completely ignores feeding buzzes (Issue #49)
Thanks for the reply with your existing recordings you probably have a lot of feeding buzz samples. Soprano buzz are slightly beyond your stft resolution, mine is very slightly better (above). The interval is the give-away, the echo-location gives the species. Tricky as the interval is closer that a lot of the echo's I see, many of which confuse batDetect2 Regards Mark Glover
------ Original Message ------ From "Santiago Martinez Balvanera" @.> To "macaodha/batdetect2" @.> Cc "markgloverswaybtinternetcom" @.>; "Mention" @.> Date 02/06/2025 19:10:58 Subject Re: [macaodha/batdetect2] Classify completely ignores feeding buzzes (Issue #49)
mbsantiago left a comment (macaodha/batdetect2#49) https://github.com/macaodha/batdetect2/issues/49#issuecomment-2931875591 Hi @markgloverswaybtinternetcom https://github.com/markgloverswaybtinternetcom,
Yeah, this is a known limitation and a bit of a design choice. At the moment, we don't have enough labelled data for feeding buzzes to reliably detect them, so we've chosen not to include them in the current model.
We'd definitely welcome any labelled examples of feeding buzzes, or even social calls (though that’s an even trickier problem). With more data, future versions of the model could potentially include detection for these.
— Reply to this email directly, view it on GitHub https://github.com/macaodha/batdetect2/issues/49#issuecomment-2931875591, or unsubscribe https://github.com/notifications/unsubscribe-auth/AZP7OYFPOQXMDCYVAYHFKV33BSHTFAVCNFSM6AAAAAB464QEVKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDSMZRHA3TKNJZGE. You are receiving this because you were mentioned.Message ID: @.***>
-- This email has been checked for viruses by Avast antivirus software. www.avast.com