Google Perch v2 and BattyBirdNET Bat Classifiers Now Available

[tphakala]

Hello All,

The release I talked about in the previous announcement is now available. Today's nightly build adds support for Google Perch v2 and BattyBirdNET bat classifiers through a new model gallery.

What's new

BirdNET-Go now ships with a model gallery where you can install additional AI models alongside the default BirdNET v2.4. You'll find it under Settings > Analysis > Models. Available models in this release:

• Google Perch v2 (14,795 species)
• BattyBirdNET (11 regional bat classifiers covering Africa, Americas, East Asia, Europe, Middle East, South Asia, Southeast Asia, and USA)

How to set it up

1. Go to Settings > Analysis > Models and install the model you want from the gallery
2. Go to Settings > Audio and select which models each audio source should run

You can assign one or multiple models to a single audio source, or dedicate separate sources to different models. For example, a standard microphone running BirdNET and Perch for bird detection, and a separate ultrasonic USB microphone running BattyBirdNET for bats.

Hardware requirements

These new models need more resources than BirdNET v2.4 alone. You'll need at least 2 GB of RAM and Raspberry Pi 4B class hardware at minimum. Raspberry Pi 5 is recommended, especially if you plan to run multiple models simultaneously.

Known limitations

This is very new and you should expect rough edges. Some things to be aware of:

• Perch v2 species scope is currently limited to BirdNET v2.4 species. Perch outputs scientific names which are mapped to common names using BirdNET labels, so only species that exist in both models are usable for now.
• Bat detection requires an ultrasonic-capable audio capture device like AudioMoth (which can act as a USB microphone with sample rates up to 384 kHz) with a sample rate of 192 kHz or higher.
• Bat detection is Linux only for now. The exclusive sound card mode needed for high sample rates is not yet available on Windows or macOS.
• Bat detection works with sound card sources only, no RTSP stream support for bat audio.
• The settings UI is still work in progress, feedback on layout and usability is welcome.

With this many changes there are bound to be issues I haven't hit. If you run into problems, please file an issue with your config and logs.

Acknowledgments

Thanks to Justin Chuby for converting Google Perch v2 to ONNX format, which made integration into BirdNET-Go possible. Thanks also to R.D. Zinck for the BattyBirdNET bat classifier models.

Full changelog is in the release notes.

[Summertanager]

No way around the ultrasonic capable auido capture? Because I switched over models to using the USB Microphone and is not picking up anything despite hearing a robin in the audio stream

[tphakala]

I am sorry but I don't understand what do you mean? Can you please be more specific, or better yet, file a bug report and submit support dump if you are having some issues.

[Aimache]

Updated, bat model installed, and a (real) "Pipistrellus nathusii" bat tracked less than 2 minutes after. Amazing work! Microphone is a Clippy, by the way.

[Summertanager]

Updated, bat model installed, and a (real) "Pipistrellus nathusii" bat tracked less than 2 minutes after. Amazing work! Microphone is a Clippy, by the way.

What Microphone are you using?

[alexbelgium]

Clippy EM272 is great to capture bats up to 55khz (needing to be used with pcm with 192khz stream) ; no need to use 384khz as the microphone isn't sensitive to such high values and anyway very few common bats have specific sounds above 60khz

[rindlerblabla]

Wow, cool!

Am using a clippy em272 even for bat detections. Not optimal, but... Maybe the 258 is better if you want to build the microphones yourself.

Am a total novice when it comes to these detections.

Get a lot of double detections of bats at the same time as birds but it is difficult to understand whether the bat captions are correct.

Would it be possible to show the spectrograph for higher frequencies than now? Would it be possible to add another low pass filter just for bat detections, so I sort out everything under ex 16khz or something.

[tphakala]

Bat detection audio clips do contain full frequency range up to sample rate you have configured your sound card to, only bird detection clips are downsampled to 48 kHz to conserve disk space. However spectrogram visualisation in BirdNET-Go is capped at 12 kHz at the moment, so you could validate bat detections by downloading them to your PC/mac and opening them on audio editor like Audacity to see if there is activity above 20 kHz.

I'll see if I can improve this later.

[rindlerblabla]

Obviously the turdus pilaris is mistaken for a barbastella barbastellus.

But will examine if I can change something with my detection settings to avoid this.

[tphakala]

At least increase bat detection confidence threshold, it is by default 0.5 which is quite low value.

[rindlerblabla]

Good idea! However there's no save button when changing the values on the Analysis page and the changes are not stored automatically.

[matkaan]

I have Perch v2 model on same rtsp-audio source ast the default model. I have started to get interesting identification of birds and one funny looking frog, which has been seen in Europe only once before on 1980s. Haven't seen those yet on my backyard, so I started to doubt that maybe the Range filter does not apply to Perch v2 model?

[tphakala]

It does not on release you are using, it is fixed in codebase but requireds a new release, will be available in day or two. Also when new release drop you need to reinstall Perch v2 model to download new range filter model.

[matkaan]

Perfect! And thanks a lot for great solution, whole family is loving it. Still I am bit sad that we didn't really have Australian "Crinia signifera" on our garden here in Finland.

[hpk60]

I really doubt that the clippy can detect bats correctly, mine is 'hearing' 4 different bats, also some, that usually live near waters...

[rindlerblabla]

I guess it is true especially when connected to sound cards/adapters which are not able to record 192 kHz. From what I have read the clippy/primo em272 is also worse than em258 on higher frequencies.

I would guess that all my detections with an em272 and an ordinary 3.5mm to usb-connecter are false positives. Will consider another microphone for bats.

[tphakala]

I would not rule out Clippy from detecting at least some bats, I just got few true bat detections last night using Rode NT5 studio microphone, it is connected to Focusrite Scarlett audio interface and I am running it at 192kHz samping rate. By manufacturer specifications Rode NT5 only covers 20hz to 20 kHz range. https://rode.com/en-fi/products/nt5

I applied some noise reduction and increased gain by 20dB for this screenshot.

BirdNET-Go spectrogram does not visualise bat sounds at all for now, I need to think how to fix this.

[Aimache]

EM272 is known to be pretty capable with frequencies up to 35-40 kHz. That's why "field recorders" like it. Do not overtrust manufacturer's specs, it does not brutally stop working at 20 kHz ;-)

Which means that, with a proper audio card, it's possible to catch multiple bat species which are in the 20kHz-30kHz range (Nyctalus,..., up to Pipistrellus nathusii).

[rindlerblabla]

My comment about the em258 is based on this comparison. https://zachpoff.com/resources/cheap-microphones-for-ultrasound/

Haven't compared them myself.

Guess the sound card is the issue with my detections.

[rindlerblabla]

Speaking of sound cards I found this one. https://github.com/Mariano-Esteban/Ensor-384kHz-32bit

Think I will try to order some of those. But will use external mics, the em258 for bats at 192khz and the em271 at 48khz for birds.

[tphakala]

There is a new release which fixes several issues, it reduces bat detection false positives by enabling night time scheduler by default, it makes bat detection to run only between sunset and sunrise. Second there is post detection validation which looks for bat detection clips for audio precense in ultrasonic frequency range, if there is none significant audio in there then detection is flagged as "Unlikely".

Google Perch v2 gains range filtering with new recently released BirdNET Geomodel v3.0, to enable this you need to reinstall Perch model from model gallery, just click Reinstall button on model card and it downloads Geomodel, it gets enabled by default.

[rindlerblabla]

There is a new release which fixes several issues, it reduces bat detection false positives by enabling night time scheduler by default, it makes bat detection to run only between sunset and sunrise. Second there is post detection validation which looks for bat detection clips for audio precense in ultrasonic frequency range, if there is none significant audio in there then detection is flagged as "Unlikely".

Most of my bat detections are tagged as unlikely even if there seems to be genuine calls, when opening the files in the oversample web app or Audacity.

CV: 0.100, but I have also got what I think is true detections with CV: 0.077

How can I help out with tuning this setting? Would you like to have the wav-files?

[tphakala]

This is really good feedback! WAV files would be very useful so I can improve this filter, can you upload them to my personal filedrop site? https://filedrop.pobox.fi/

[rindlerblabla]

Perfect, done!

[maxdraki]

@tphakala
Adding to this — I've been independently verifying BirdNET-Go's bat detections against BatDetect2 for the past 21 days as a second-opinion cross-check, and have systematic evidence on the unlikely flag that may help tune cvthreshold. Site is Winchester, UK (Hampshire); AudioMoth USB mic @ 256 kHz; BattyBirdNET-UK.

Dataset

• 21 days continuous (2026-05-15 → 2026-06-04)
• 5,699 bat detections cross-checked against BatDetect2
• 295,404 individual echolocation calls analysed by BatDetect2 across those clips
• 2,593 currently flagged unlikely in BirdNET-Go's DB; 1,792 of those are in my cross-checked set
  Of 1,792 clips BirdNET-Go flagged unlikely:

BatDetect2 verdict	Count	%
✅ Confirmed (BD2 independently agreed on species)	1,380	77%
⚠️ Species disagreement (BD2 did find echolocation, called a different species)	292	16%
🔇 No echolocation found (BD2 also found nothing)	120	7%

I'm happy to share whatever data might be helpful

I can provide:

• Curated WAV bundle — ~30 clips spanning the failure modes (unlikely-flagged but BD2-confirmed strong pipistrelle / barbastelle / Leisler's passes, plus a few genuinely-empty negative controls), each with a manifest line: BirdNET species + confidence, current unlikely flag, BatDetect2 verdict + call count, BD2's per-call species distribution.
• Full CSV of the contingency table — every clip's basename + BirdNET-Go's claim + unlikely flag + BatDetect2 verdict + match fraction + call count. ~5,700 rows, easy to re-analyse against any CV change.

[tphakala]

Can you share subset of both incorrectly and correctly flagged files through filedrop service linked above?

[rindlerblabla]

What interference time have you got?

Am running a Raspberry Pi 4 Model B Rev 1.4 (using all cores) and have:

BirdNET v2.4 228ms
Google Perch v2 2010ms

With the perch model it seems like 4b soon is not fast enough...

[tphakala]

RPi 4

• BirdNET v2.4 199ms
• Google Perch v2 2036ms

RPi 5

• BirdNET v2.4 81ms
• Google Perch v2 678ms

With both of these models running RPi 5 interface is significantly snappier. I think Perch model could be optimized somewhat, also batching could help. Currently implementation is still very fresh and not fully optimized.

[hpk60]

Is there a way to find out, when Birdnet 2.4 and Google Perch come to a different result at a specific sound sample?

[tphakala]

@hpk60 no at this time, you can monitor detection events from logs/action.log file, but for now there is no easy way to correlate when models disagree. Here is some documentation on how detection pipeline works at the moment https://github.com/tphakala/birdnet-go/wiki/detection-pipeline.md#cross-model-detection-consensus

[tphakala]

@6zXo8aMn2TRXOs I just merged a fix which optimizes inference code, instead of running all models in parallel they are now ran in serial so each model gets full access to system resources like CPU and memory bandwidth. This cut Perch inference time to half.

RPi 4

• BirdNET v2.4 175ms
• Google Perch v2 933ms

RPi 5

• BirdNET v2.4 80ms
• Google Perch v2 353ms

[oturn]

What is the difference between BattyBirdNET USA East (High) and BattyBirdNET USA East? Should I just use both?

[tphakala]

I had to consult Claude since these models are from BatttyBirdNET project and I am not so familiar with bat models.

USA East (standard) has 22 output classes: 16 bat species + 6 non-bat/noise classes (background, collected, noise, noise-us, other, silence). Model size: 93 KB.

USA East (High) has 17 output classes: the same 16 bat species + only Cricket as a non-bat class. No noise/background categories. Model size: 69 KB.

The "High" variant is a higher-precision classifier that strips out the ambient noise rejection classes. Instead of the model itself deciding "this is background noise" or "this is silence," it focuses purely on distinguishing bat species (plus cricket, which overlaps in frequency range). The trade-off:

• Standard: includes built-in noise rejection via dedicated output classes, so the model can explicitly classify non-bat sounds. This is better for noisy environments where you want the model itself to filter out false positives.
• High: omits noise classes, which means all classification probability mass is distributed across actual bat species. This can yield higher confidence scores for real bat calls, but it's more prone to false positives in noisy conditions since it has no "this is just noise" output to absorb non-bat sounds.

The same pattern holds for the Bavaria models: Bavaria-high has 24 classes (species + 2 non-bat categories), while standard Bavaria has 32 classes (species + 10 non-bat/noise categories like audiomoth, bird, electrical, rain, etc.).

In short: "High" means higher sensitivity/confidence for bat detections at the cost of less built-in noise rejection.

[rindlerblabla]

I got a notification of the first detection of a Thinornis dubius. Guess it is a perch detection. Only the Latin name is shown in the UI.

However I already have some 50 detections before of Charadrius dubius , which obviously (?) is the same species but wrong/old label.

Which is best way to avoid these duplicates? And also show the correct common name in the UI? Should I rename the old one in the label file or is this a "bug"?

[tphakala]

Please file a bug of this, and submit support dump, it helps me to troubleshoot this.

[PeteLawrence]

Thanks for your efforts on this!

I’ve successfully got both bird and bat detection working on a RPI5 using a pippyg flashed with their ultrasonic microphone firmware, listening at 196khz.

It would be nice when listening to the live stream to be able to switch bitrates - at present when listening during the day to birds you just hear the (presumably) time-expanded versions of their calls.

[hpk60]

Updted a system today. From older version (6 weeks ago).
After the update all historical data of several month are gone. Are they lost now?

[tphakala]

There is nothing in the update which should cause this, are you sure you installed it using same user it was originally installed as?

Have you looked at file system are birdnet.db and clips in some other users home directory.

[hpk60]

I used
bash. /install.sh Option 1
nothing else
The directories 2026 02-03-04-05 are there
Bird net db is there as well - 110GB in size

I'll send a support package #5555

[alexbelgium]

This is a great feature. Sorry if you answered it elsewhere but will you add rtsp stream capability for bats in the future ?

Edit : just saw it is already available in the rtsp streams options. Thanks !!

[alexbelgium]

EDIT : rtsp stream for bats doesn't work it seems ; Claude's opinion on it is due to fixed sample rate of rtsp with "SampleRate: conf.SampleRate, // = 48000, hard-coded constant" in internal/analysis/audio_pipeline_service.go

[tphakala]

Yeah it is not hooked up properly, I will fix this for next nightly. Problem just is, 99% of streams are unable to support bat detection due to compression etc.

[PE1PWF]

Just got the BirdNET-Go working and the bats are out.
I have been using a raspberry 2040 based 384kHz USB microphone "vleermic" to capture the birds during the day.

I did add batty birdnet EU and selected both to work with the Vleermic in the audio settings but only captured bats when I turned off the birds in the audio settings.
Is this as expected? It would be nice if I did not have to switch manually, maybe I did not quite understand what settings should be made to do this.

Also the spectrograms do not show the bat sounds the lower part of the spectrum is shown, only the live audio seems to show the full spectrum. The bats however seem to be recognized quite fine.
Thanks for this nice piece of work.

Edit:
The audio settings show "bat" but if I try to see what model is activated I only see the birdnet model (not selected)
Edit
After enabeling the birdnet again the bats kept working bit in setting there is no bat model (with a selection box)

[oester]

Could you share some details on this setup? I would be interested in putting together something for bats as well. Thanks.

[PE1PWF]

Runnig a raspberri pi 5 with 8gb of memory, I first tried the battybirdnet for raspberry. My microphone is rp2040 based sold in the Netherlands you could build yourself a Griff, or use Pippyg, Pipistrelle mini or audiomoth mic or regular audiomoth with microphone firmware that should also work. The micrope is simply uses on a USB extension cable.
I built me the BattyBirdnet raspberry project first with the same materials but the BirdNET-Go looks much nicer and it seems like a very active group of people here. Also switching between classifier models seems more easy in BirdNet-Go

[tphakala]

I did add batty birdnet EU and selected both to work with the Vleermic in the audio settings but only captured bats when I turned off the birds in the audio settings.
Is this as expected?

No this is not expected, you don't need to turn classifiers on and off. Are you sure you are on latest build?

Spectrograms do have a known limitation, it requires deeper changes to get spectrogram frequency scale to adapt to species and I have gotten to that yet.

[DD4WH]

there is a small "bug": you can only select bat classifiers for ultrasound mics as soon as you restart services in the System -> overview menu
After doing that, everything is fine and you can select in Settings -> audio which classifier should be used with each soundcard

[PE1PWF]

Selecting was not a big deal, but once I added it and had it selected the option was gone after I clicked the gear, I attached a link in an other reply. (sorry I messed that up by simply using the mail.) I think we also are missing the option to disable some bat species, it is however quite fun to play around with.

[PE1PWF]

It seems to have worked afterall after activating the birds again last evening the bat kept being recognized and when the birds came out these also were recognized properly. In the soundard settings one can see bat, BirdNET v2.4 (TFLite) Clicking on the gear however does not show the bat model bit I am sure it is active. The double button to save changes is a little confusing. In BattyBirdnet I can apply a noise profile for my microphone, is that also available for BirdNET-Go? I failed to find it. <img width="1394" height="1003" alt="bat settings" src="https://github.com/user-attachments/assets/50336f25-797b-4478-80d4-0a5222ebb041" />

[tphakala]

The double button to save changes is a little confusing.

Yes I agree, I have been meaning to fix that

[DD4WH]

This (BattyBirdNET-Pi) is a really awesome extension to BirdNETGo !
It works very well and the new UI is extremely good! However, my RPi4 with 4GB RAM gets really hot when using a Bat model and the BirdNET model simultaneously. I have an audio USB sound card with an AOM5024 mic and additionally a selfbuilt Pippyg as a USB ultrasound mic attached via USB to the RPi

It would be a major step forward to have the ability to view bat spectrograms from 10kHz to 130kHz.

Maybe you could implement some kind of switch in the software that shifts automatically the frequency scale (0-12kHz for birds, 10-130khz for bats) ?

Keep up the good work! Frank DD4WH

[PE1PWF]

I am running just one 384kHz microphone on a raspberry pi5 8GB for both the bats and the birds. I don't quite see why one wants two audio streams to process. CPU load is around 10% during daytime the temperature about 40C when the batas are out it goes up to about 27% system load and 60C on a passively cooled (all aluminium compact housing) pi.

[tphakala]

However, my RPi4 with 4GB RAM gets really hot when using a Bat model and the BirdNET model simultaneously.

You could try to reduce CPU threads to 3 or even 2, that might help with temps while still being fast enough to avoid audio buffer drops. CPU thread setting is in bottom of Settings > Analysis page, it currently lacks proper hot reload so you need to restart BirdNET-Go process to make new value active.

There is still gaps in features like spectrogram view, I try to close them soon.