birdNET Analyzer model update (1 Viewer)

[S_Man]

Ok did a bit of testing this afternoon. On the whole, it has improved some....

Thanks, I may try to update, although Costa Rica is certainly better served than Asia... (both Merlin and BirdNet were almost useless on a recent trip, even in excellent conditions, it was really shocking compared to Europe - but I still have some files so I can try again...)

 

[Buckster]

updated my BirdNET-Pi today although made a real meal of it

I used the in-built File Manager to delete the old file, and then went to upload to replace but that wouldn't work (maybe upload in the filemanager is restricted)

so had to copy into the boot directory (only directory I could access in Windows when the Micro-SD put into it as only one in FAT)

then copy from there via terminal in BirdNET from the Boot directory/partition to the Model directory on the BirdNET partition

works fine now though

 

[Buckster]

2 questions please

(1) what do people find the best combination of recording length, overlap, and "extraction length". I've changed mine to 18 seconds, 0.2 overlap, and 9 seconds extraction - and have changed extraction file format from mp3 (as was very lowbitrate mp3 (64kbps I think) to FLAC which is lossless

(2) how does the model work please ref the initial identification then comparing against the Mdata of whether the bird is likely or not ?

I think it looks at a 3 second window of audio, and identifies the most likely % bird, then compares that against the likelyhood of that bird being possible for that location

So in the scenario say the highest possibility bird in that 3 second window is identified at 78% possibility, but at the next stage it is rejected as "not likely" for that location

what if there was a bird at 76% in that 3 second window possibility that was likely for that location ?

could the model do multiple passes (or does it do this already)

so extracted audio

(1) Bird 1 - 78% - not likely - re-analyse excluding that Bird
(2) Bird 2 - 76% - likely - record

This maybe is how it works already - or does it just look at every possible bird within a 3 second window ?

 

[LittleBitOfBreadNoCheese]

It builds the localized species list first using the M_ model then runs through each file using the main model. All open source so not too difficult to track if you look at gui.py, analyze.py and species.py. The M_ model call is inside model.py

 

[TicoTyler]

Thanks, I may try to update, although Costa Rica is certainly better served than Asia... (both Merlin and BirdNet were almost useless on a recent trip, even in excellent conditions, it was really shocking compared to Europe - but I still have some files so I can try again...)

For sure. I know Merlin's sound ID requires 100 - 150 recordings to train their model. I'm not sure if BirdNET has a similar threshold. Costa Rica has around 215 species under 100 and 300 some under 150. So around 30-40% of the species are not even at the minimum threshold.

 

[S_Man]

For sure. I know Merlin's sound ID requires 100 - 150 recordings to train their model. I'm not sure if BirdNET has a similar threshold. Costa Rica has around 215 species under 100 and 300 some under 150. So around 30-40% of the species are not even at the minimum threshold.

It must also be in this range, these models are easily spoiled with too many samples from what I understand.
And it's apparently quite easy to train them, using xeno-canto it should be possible to have many more species and countries, but it takes some time and organization, obviously... I suppose these sounds are free to use in this case.
That said I had some other issues on Windows in addition to the lack of geographical coverage, ffmpeg module, also sound files that are not 48kHz wav, so I haven't been using BirdNET recently, and I'm not going to dive into this "AI" thing!