Bat detection with BirdNet-Pi

[VDierks]

Hi,
i'm using an AudioMoth as my USB-Microphone, which ist capable of recording sounds in the ultrasonic range.
The analysis of bat-recordings is quite time consuming, because most recorded files contain no calls or long periods without calls.

My idea is to add an ultrasonic-option to BirdNET-pi, that stores sound-files with time-stamp to a seperate folder, when the microphone detects events above 18kHz . Automated bat identification would be a nice feature.

[mcguirepr89]

Hello! Sorry for the late reply --

The first thing that came to mind for me was the great work going on with @audevuilli's https://github.com/audevuilli/shazam4bb you may consider reaching out to Aude there for ideas on implementation.

Perhaps also you may find something of interest in this conversation 🦇

Please keep us updated on your progress getting bats into the BirdNET-Pi cave!

[tomwhite101]

Hi @VDierks did you get anywhere with progressing this idea?
I would like to do the same thing.
Interestingly I have successfully detected bats with the audiomoth using the BTOpipeline so the microphone is capable of at lease some bat range.

[blogdemoi]

I am also interested, I also use an audiomoth for bat recordings.

[mcguirepr89]

#69 (comment)

[mcguirepr89]

}\^-,-^/{

Original:

2022-09-10_21.37.05.mp4

Pitch-shifted

2022-09-10_21.37.05.pitch-shifted.mp4

[tomwhite101]

@mcguirepr89 this is so cool! Well done you! is the model identifying?

T

[mcguirepr89]

Nope, no classification at this point, just bringing things into the audible frequency range for now :)

[nilspupils]

That is very impressive! What mic did you use?

[mcguirepr89]

I have an audio moth that I use for the bats but still use the electret+ugreen for the birds (I am still physically swapping the mics out in the evening for now)

[VDierks]

Thanks for your efforts. Which sample rate do you use? I use my AudioMoth with 192kHz.

[mcguirepr89]

Same -- I set the USB firmware for the Audio Moth to 192kHz, though I think 96 would also suffice for my neck of the woods.

[croisez]

My post here is a bit off-topic. I just want to share my mind with you all about bats. I have seen a very captivating documentary on the technique bats are using to "see" their environment. Changing frequency of their pulsing signal when moving their head, to be able to discriminate the different reflections coming back. It is a stunning animal. I also learned that they possibly not see surfaces that are too planar, because there are no signal reflection coming back to their ears.

[mcguirepr89]

I wanted to share that I have the beginnings (only testing in my region on the bats outside my window) of a pretty cool new little feature for BirdNET-Pi and folks with access to an ultra-sonic microphone.

The new feature is a live stream of the bats -- for folks able to check out my beta installation, you'll be able to listen to my bats live https://newbnp.birdnetpi.com

[davidmam]

Hi, i'm using an AudioMoth as my USB-Microphone, which ist capable of recording sounds in the ultrasonic range. The analysis of bat-recordings is quite time consuming, because most recorded files contain no calls or long periods without calls.

You can set triggered recording so that it will only save a recording when there is something worth analysing. This has been a godsend for me as I can then put detectors out for weeks instead of days.

[VDierks]

I‘ve tried the triggered recording several times, but did not find a configuration that worked for me. I use the AudioMoth in an urban Environment with a lot of noise in the ultrasonic frequency range. Without amplitude-threshold the Audiomoth is almost constantly triggered and records huge files, with amplitude-threshold only the loudest calls are recorded and most bats will go unnoticed.

[davidmam]

Tell me about it.. I had a semi urban setting with little noise. This was a long term study pre/post infrastructure change. I then discovered it had been triggered with loads of recordings that mapped as Horseshoe Bat (not present in Scotland). One of the neighbours had installed an ultrasonic pest detector that gave a perfect 80kHz harmonic on a square wave, just like horseshoe bats.

[rdz-oss]

If we manage to build in a detector that woks good enough, it can continuously scan and record only when a detection happens. For that. actually all I need are enough bat call samples to train a network.

[davidmam]

As well as looking at batdetect2 (I'd suggest reaching out to Oisin) you could also look at the work done by the audiomoth team for thresholding.

[rdz-oss]

History repeats. Something similar seems to have happened to Svardsten53. Presumably with crickets. Curious this also was mapped to horseshoe bats.

[incandescental]

I've been having similar thoughts - it'd be cool to switch the Pi to 'bat mode' once the birds have gone to bed. I'd be happy to be involved in training up a TF model if any chiropterologists out there know of a good source of labelled bat recordings?

[JohnJoosten]

@incandescental
Some bat processing stuff:
https://www.nabatmonitoring.org/methods-and-sample-design
https://code.usgs.gov/fort/nabat/nabat-ml/-/tree/v1.0.1

The used samples can be found here:
https://www.sciencebase.gov/catalog/item/627ed4b2d34e3bef0c9a2f30. Unfortunately it is only based on USA species

[rdz-oss]

Hi BirdNET-Pi team,

first of all thank you for your great project! My family enjoys the BirdNET-Pi very much.
As I am also interested in bats I have checked the code to see what options there are to extend the functionality to cover bats also. I work as a cloud solution architect and could provide pull requests to your projects if you are open to that in general.

In order to cover bats I currently see the following options

• Extend the BirdNET-Pi system to make the sampling rate a configurable variable. It currently is set to 48000 Hz. Sampling for bats would need twice the frequency that they call on, so for a 60000 Hz call that would be at least 120000Hz. I think I could make the necessary modifications in the code and provide a pull request. The only thing this would do is to make the sampling rate a configurable variable. I have a hardware setup that already works and can record both birds and bats in principal.

• The classifier from BirdNET-Analyzer is trained on data sampled to 48000Hz. This is fine for many animals and birds, but would only allow to detect bats and insects to max 24000Hz. So the Network would have to be adjusted to use say 125000 Hz.
  There are a few options here:
  - If the BirdNET Lite team would agree, one could try to 'upsample' the training data for birds and
  add some for bats. The performance would need to be validated, but this could work.
  - Alternatively, the bat calls could be shifted into the audible range and then resampled
  at 48000Hz. For training and live analysis. This would additionally require a custom modification
  to the BirdNET-Pi system.
  - I could make a version of BirdNET Lite that operates on 125kHz. The training data would not be available to me (for birds),
  but I could gather some for bats and birds from different sources and train a network myself.
  This could then be plugged into the prepared BirdNET-Pi version (adjusting the sampling frequency to 125 kHz).
  BirdNET-Pi would then either operate as is with 48000Hz set or be used as a bat detector with alternative classifier and 125kHz.

Of course I would honor the License terms in all this and preferentially provide pull requests to the respective projects.
What do you think? Which version might work for you? Perhaps you have other ideas?

Let me know.
Best regards,
Richard

[guillaumarchais]

I just have had a look at https://somersetbat.group/advice/which-bat-detector/audiomoth/process/batclassify/ and both tables show 0 identified for R. hipposideros, except for 384 kHz fs setting.
But don't get me wrong, 256 kHz may be the best setting for Audiomoth... just with a limit for that particular species + R. euryale.
And sometimes harmonics (h1) at 50 something kHz can be picked up if close enough to the microphone (see fig A of https://www.pnas.org/doi/full/10.1073/pnas.0703550105)

[BobSpong]

Isn't that a just a commentary on the accuracy of a classifier and the lack of such calls in the sample?

Another commentary on sample rate requirements from https://www.researchgate.net/publication/368683386_Good_practice_guidelines_for_long-term_ecoacoustic_monitoring_in_the_UK

"Sample Rates - the sampling rate is how often the recording device samples
the analogue signal in order to convert it to a digital representation. The sound
signal needs to be sampled at least twice the rate of the maximum frequency
of the sound of interest, e.g. using a sampling rate of 4 kHz will allow recording
of sounds up to just 2 kHz, whilst a sampling rate of 36 kHz will allow the
recording of sounds up to 18 kHz. It is therefore necessary to ensure that
the chosen ARU has an available sampling rate double that of the maximum
frequency of the sounds you wish to record."

[rdz-oss]

Now ready to be tested for North America!

Classifier and system are ready for a field test in North America!

Who would like to try and send feedback? Let me know ! Write to: [email protected]
Your bonus: The noise engineering is an important part. Send some of your noise and it will work at your place :)

Two systems are running well in the UK and GER. Time to expand :)

To use the North American version:

• install as described here: https://github.com/rdz-oss/BattyBirdNET-Pi
• go to the tools setting
• log in as user 'birdnet', leave password empty
• Settings -> Advanced Settings -> Bat Classifier (use USA)
• the system will reboot, wait two minutes and refresh browser

On my balcony tonight - including a social call -

[guillaumarchais]

Hello Richard, one thing I have not fully understood yet is whether this BattyBirNetPi is working simultaneously with BirdNetPi using the same SD card ? For the install procedure as detailed here, you mention "You call the install script from this repository (not from BirdNEt-Pi - unless that is what you want to install):"
How do I specify the right repository please ? It is not something I do often and I am missing a step here.
When I open the LXTerminal, I get the hostname and ~/BidNET-Pi$
Is it OK then to copy and paste your install srcript there ?

curl -s https://raw.githubusercontent.com/rdz-oss/BattyBirdNET-Pi/main/newinstaller.sh | bash

[rdz-oss]

Thank you for asking. I will try to clarify that on the readme.
When you open the Webinterface, does it show a bat ? Then all is ok and you are on a system scanning for bats only.
Otherwise (if you seen no bat in web interface)
First, you flash your SD card with Raspberry Pi OS 64 bit lite. Before flashing, you set the Wifi name and password as well as a name for the system. You should now have a blank install of the OS, nothing more.
Then you put the sd card into the Pi 4 and power up the system. Log into the system with ssh and name and password you set while flashing the OS.
Now you call the

curl -s https://raw.githubusercontent.com/rdz-oss/BattyBirdNET-Pi/main/newinstaller.sh | bash

There have been some issues wih installs of BirdNET-Pi of late that also affect BattyBirdNET-Pi. If the install fails with unsupported tensorflow, try the older version of the 64 bit Rasperbby Pi OS lite.
Let me know if it works.
BR
Richard

[croisez]

I have done some tests of frequency shifting for batt sounds with Patrick one year ago. The goal of the test was to shift down sounds recorded from bats. In order to keep the sound quality for the automatic recognition, we had to use a numpy script for the freq shifting process, which was quite resources hungry. I had a RPI3 at that time, and the computation process took about 30s for a wav file chunk. Now, on my RPI4, it takes about 5s for the shift. It is important that we keep compatibility with all RPI versions, because every people does not own a RPI4. I am not sure that it is a good idea to introduce bats into the classifier, because it will consume a lot of resources on every RPI, and it will be only used by fews. What is your opinion? Le mer. 26 juil. 2023 à 18:01, rdz-oss ***@***.***> a écrit :

…

Hi BirdNet-Pi team, first of all thank you for your great project! My family enjoys the BirdNet-Pi very much. As I am also interested in bats I have checked the code to see what options there are to extend the functionality to cover bats also. I work as a cloud solution architect and could provide pull requests to your projects if you are open to that in general. In order to cover bats I currently see the following options - Extend the BirdNet-Pi system to make the sampling rate a configurable variable. It currently is set to 48000 Hz. Sampling for bats would need twice the frequency that the call on, so for a 60000 Hz call that would be at least 120000Hz. I think I could make the necessary modifications in the code and provide a pull request. The only thing this would do is to make the sampling rate a configurable variable. I have a hardware setup that already works and can record both birds and bats in principal. - The classifier from BirdNet-Analyzer is trained on data sampled to 48000Hz. This is fine for many animals and birds, but would only allow to detect bats and insects to max 24000Hz. So the Network would have to be adjusted to use say 125000 Hz. There are a few options here: - If the BirdNet Lite team would agree, one could try to 'upsample' the training data for birds and add some for bats. The performance would need to be validated, but this could work. - Alternatively, the bat calls could be shifted into the audible range and then resampled at 48000Hz. For training and live analysis. This would additionally require a custom modification to the BirdNET-Pi system. - I could make a version of BirdNET Lite that operates on 125kHz. The training data would not be available to me (for birds), but I could gather some for bats and birds from different sources and train a network myself. This could then be plugged into the prepared BirdNET-Pi version (adjusting the sampling frequency to 125 kHz). BirdNET-Pi would then either operate as is with 48000Hz set or be used as a bat detector with alternative classifier and 125kHz.. Of course I would honor the License terms in all this and preferentially provide pull requests to the respective projects. What do you think? Which version might work for you? Perhaps you have other ideas? Let me know. Best regards, Richard — Reply to this email directly, view it on GitHub <#300 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANRFSITHBLJYOEWP6YJBW3XSE5PLANCNFSM5WSADX5Q> . You are receiving this because you commented.Message ID: ***@***.***>

[rdz-oss]

Thank you for the Info!
I can see the point about maintaining compatibility with existing setups. Although it would be desirable/preferable to have one system that can do both, it might be the most pragmatic way to separate ultrasound from non ultrasound based detection.
This would still leave the question if the systems should be separate or if one would make BirdNET-Pi configurable for either birds - as is - or bats/insects etc. For bats, one would need to set the sampling frequency to a higher rate (say 125kHz or more) and use a new/different classifier.

• I could develop the ability to set the frequency as needed into BirdNET-Pi and provide the necessary pull request. This should not be a burden on the BirdNET-Pi system in any way. It would keep the system as the main development branch for both purposes also in the future. Regular BrdNET-Pi users would not even notice.
• Then, if I get enough samples for bats, I can train a network e.g. based on BirdNet to recognize them. This would also need to operate on the higher frequency for training and detection, but that should be possible if I can get my hands on enough data. This can be an OSS effort also.
• The hardware setup is straight forward and the necessary ultrasound microphones are already working with the BirdNet-Pi, at least in my setup which could serve as a template.

If somebody else would like to do any of this, of course welcome as this is in my spare time and not done to generate money :)
But I can set out and try it in about that order. My current feeling is this can work - depending on the quality of the classifier. If a classifier were to be used that does not have the BirdNET interface it would require more changes.
It hinges on the question:

Would the BirdNET-Pi team be open to making the sampling frequency a configurable variable?

[croisez]

If we allow users to configure the sampling frequency, then the recognition model must follow and be compatible. Right? The whole chain must be coherent Le jeu. 27 juil. 2023 à 14:15, rdz-oss ***@***.***> a écrit :

…

Thank you for the Info! I can see the point about maintaining compatibility with existing setups. Although it would be desirable/preferable to have one system that can do both, it might be the most pragmatic way to separate ultrasound from non ultrasound based detection. This would still leave the question if the systems should be separate or if one would make BirdNET-Pi configurable for either birds - as is - or bats/insects etc. For bats, one would need to set the sampling frequency to a higher rate (say 125kHz or more) and use a new/different classifier. - I could develop the ability to set the frequency as needed into BirdNET-Pi and provide the necessary pull request. This should not be a burden on the BirdNET-Pi system in any way. It would keep the system as the main development branch for both purposes also in the future. Regular BrdNET-Pi users would not even notice. - Then, if I get enough samples for bats, I can train a network e.g. based on BirdNet to recognize them. This would also need to operate on the higher frequency for training and detection, but that should be possible if I can get my hands on enough data. This can be an OSS effort also. - The hardware setup is straight forward and the necessary ultrasound microphones are already working with the RasPi, at least in my setup which could serve as a template. If somebody else would like to do any of this, of course welcome as this is in my spare time and not done to generate money :) But I can set out and try it in about that order. My current feeling is this can work - depending on the quality of the classifier. If a classifier were to be used that does not have the BirdNET interface it would require more changes. It hinges on the question: Would the BirdNET-Pi team be open to making the sampling frequency a configurable variable? — Reply to this email directly, view it on GitHub <#300 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANRFSMV47SGVXMUL4FMQV3XSJLVVANCNFSM5WSADX5Q> . You are receiving this because you commented.Message ID: ***@***.***>

[rdz-oss]

The whole chain needs to be coherent -> correct

• Users would not need to set the frequency directly themselves. This would happen behind the scenes when they select an option for ultrasound bat detection. It can be controlled via the user interface or its controller.
• The default set remains as is at 48kHz
• BirdNET-Pi uses system variables a lot. It also uses hard coded variables sometimes. The sampling frequency would become just one more of these system variables. The users would not necessarily need to be shown this in the webinterface at all.
• There are many ways the webinterface could be used to allow the user a choice - birds or bats
• Default is birds, all settings ready for that
• If bats are chosen, different classifiers need to be offered and the variables in the background change also. Further, the microphone setup would be different as it would be ultrasound mics.

An intermediate step would be to make the sampling frequency a system variable set to 48kHz (default) and make no changes to the UI at all until a classifier can be offered that can actually work with bats. Then, integration options could be weighted.

[davidmam]

I would be interested in this. For automated classification I have used Batdetect2from Oisin MacAodha https://github.com/macaodha/batdetect2 which seems to work very well for UK bats.

[rdz-oss]

Thank you for the link! I checked the repo and it appears they are using pyTorch rather than tensorflow for the ANN part. I will look into it when I have some more time. Such a classifier can be hidden behind an API and then used in a common way. For BirdNET-Pi that would probably also require a custom adapter for integration.
I could make a few mods to the BirdNET-Analyzer to allow it to train an ANN for bats which would be compatible to BirdNET-PIs ANN interface. If Oisin MacAodha makes the bat calls available he uses for training I can train such a network. If not, perhaps Oisin would be willing to run the modified BirdNET-Analyrzer scripts on his data.
In general, I forked the repos and start to look for bat call training data. If there is enough out there in the right quality (original recordings in ultrasound), then I can make the mods, train a compatible ANN and integrate that into a version of BirdNET-Pi that can then detect bats.

Can you help out with some data sources?

[davidmam]

I have many terabytes of ultrasonic recordings, most of which are pipistrelles. Unfortuately I do not have them well organised for test/training of ML algorithms.

[davidmam]

There is also the issue of regiospecificity. My interest is almost exclusively UK bats which are limited in number and generally easy to distinguish.

[davidmam]

I may well see if I can get a Computing Honours studetn to take on wrapping Batdetect2 on a Pi 4 under BirdNET Pi

[rdz-oss]

Hi David,

I have forked the BirdNET-Analyzer repo and found out some more details.

• I am using the NABAT (North American) data for testing, about 90GB sound data for 30 or so species.
• What BirdNet-Analyzer offers is something like cross-learning: you use the network trained on birds and add another layer at the end, an extra classifier you train with your specific data.
• The limitation is that the BirdNet ANN expects 48000 Hz at 3 Seconds = 144000 inputs in its entry layer. One can however use e.g. 144000Hz at 1 second to generate such input. Or 360000Hz sampled at 0.4 seconds, 240000 at 0.6 seconds etc.
• This works well with the NABAT data. The classifier reaches .95 precision (I have not tried to optimize yet) on bat calls that are routed through BirdNet and the extra classifier at 144000Hz.

What does this mean?

• I can train the extra classifiers for BirdNet that are able to detect bat calls (already works)
• If I also get enough European and British bat data I can add that capability, too. Requires about 100 or more samples per species.

So BirdNET-Analyzer as a program would work for analyzing bat calls pretty much immediately. I can upload the trained classifier for NABAT right now.
For BirdNET-PI, some integration work would be necessary.
I will push the current state and .tflite classifier for North American bats here soon: https://github.com/rdz-oss/BattyBirdNET-Analyzer

Not sure yet what I will do next.
BR ,
Richard

[rdz-oss]

A good evening to everyone! I have good results from cross training BirdNET-Analyzer for bats!
I am rather pleased it appers to work well. It is currrently trained on the NABAT data set so it works for North American bats.
Since I live in Europe, I will try to get a similar collection for European bats and train that, too.
This is now my summer project. If that works, I want to test the system live with the bats here and make the necessary modifications to BirdNET-Pi to run it 24/7. How cool would it be to have a Bat-Pi based on BirdNET?
My goal is to get it done in this bat season.
So if you want to help let me know!
I will gladly train the classifier for your local bats if we can arrange the data !

Test the classifier here, if you like: https://github.com/rdz-oss/BattyBirdNET-Analyzer
Should all be there.

[davidmam]

That sounds very interesting. I'd be keen to see that on Scottish bats (a subset of UK ones).

[rdz-oss]

I am working on it. Getting a suitable data set is a lot harder than for the US species. How about fine-tuning a classifier to your local bats? Can you get samples with say 20 samples of 2-5s length for each species? I will use them to fine tune the classifier on top of what I can find to excel at Scottish bat recognition ;) . Ideally .wav files in real time.

[spambake]

Thanks so much for this work. I just started trying this out and have much more to absorb. Occasionally I get a Townsend's big-eared bat and one came by last night. Here is a photo and the associated results from a six second audio clip taken at the same time. Many other bats last night. Marin County, California. I have a raspberry pi standing by for the project.
Not sure how to test this. Let me know if you have suggestions.
(these particular Myotis are eastern bats; nowhere near here.)
1 Spectrogram 1 1 1.0 2.0 150 15000 Corynorhinus townsendii_Townsend's big-eared bat Townsend's big-eared bat 0.9907
2 Spectrogram 1 1 1.0 2.0 150 15000 Myotis austroriparius_Southeastern myotis Southeastern myotis 0.9531
3 Spectrogram 1 1 3.0 4.0 150 15000 Myotis grisescens_Gray bat Gray bat 0.8633
4 Spectrogram 1 1 3.0 4.0 150 15000 Myotis austroriparius_Southeastern myotis Southeastern myotis 0.1957
5 Spectrogram 1 1 4.0 5.0 150 15000 Myotis grisescens_Gray bat Gray bat 0.9819
6 Spectrogram 1 1 4.0 5.0 150 15000 Myotis austroriparius_Southeastern myotis Southeastern myotis 0.4359
7 Spectrogram 1 1 5.0 6.0 150 15000 Myotis grisescens_Gray bat Gray bat 0.7976
8 Spectrogram 1 1 5.0 6.0 150 15000 Myotis austroriparius_Southeastern myotis Southeastern myotis 0.1863

[rdz-oss]

I love it. Nice picture. A fast answer:

• It appears you are using the classifier at 144kHz. It sections your sample in one second intervals and puts out the two most likely species for that section. You could also try the classifier trained at 240kHz and compare results.
• While Townsend's big-eared bat is the clear winner in the first second, which is great, other species crop in in the following segments, although with lower overall likelihoods. So not bad at all at this point where things are still a Proof of Concept.
• You can change the minimum likelihood to a higher value by using the '--min_conf 0.5' option on the command line. That would get rid of some of the far fetched candidates.
• The default analysis does not overlap the sections (of one second each). You could try what overlapping does by setting the parameter to '--overlap 0.2' (any value between 0 and 0.3 could work I think. If you use the 240kHz classifier, this would have to be lower, say between 0 and 0.1.
• You can run the analysis on a directory in which you have several .wav files. It finds and works on them all. In your path, specify with /path/to/directoryfullofsamples/ with the slash at the end.

Let me know what happens. I am not sure if this becomes to specific for this discussion thread. If you like, you can use the discussions thread here: rdz-oss/BattyBirdNET-Analyzer#1

BR
~Richard

[spambake]

I'll start posting results over on rdz-oss/BattyBirdNET-Analyzer#1
We should probably post at least milestones here since it has a wider audience.

[rdz-oss]

@spambake The system is ready for testing in the US with audiomoth!
It runs in the UK and GER now stable for a while classifying bats and not too much noise. There will have to be some noise engineering to be done for the US classifier, but then it should work similarily well (or better)
Would you like to try? Shot me an email:
[email protected]
and I will tell you how to set the US classifier for the BattyBirdNET-Pi!
Just in time for the end of the bat season (at least here).
It is fully integrated and works similarily to BirdNET-Pi, now just for bats!

Also: any othe North Americans interested in testing and sending feedback?

https://github.com/rdz-oss/BattyBirdNET-Pi

[rdz-oss]

The classifiers for Europe are uploaded!
I hope I did not forget to push anything relevant to the repo. Let me know if that should have happened!

Otherwise, have fun using the new bat sound classifiers!

[Svardsten53]

I have done an installation in Raspberry Pi OS without any problems, but when running the analysis command I'm getting the following:

svardsten@birdnetpitest:~/BattyBirdNET-Analyzer $ python3 bat_ident.py
File "/home/svardsten/BattyBirdNET-Analyzer/bat_ident.py", line 361
match args.area:
^
SyntaxError: invalid syntax

I hope you can find out whats wrong. Thanks!

[rdz-oss]

Glad to hear the modification works! I will integrate your translations so that it can be used with --locale se in the future with the EU classifier. The Bavarian one automatically outputs to Latin-German names.

I run the analyzer on Linux and Mac machines and it is rather fast. But I should also try on a raspi. I have a raspi 4 with me on my vacation. It might be

• resource limitations of the Raspi model - which one are you using?
• length of the recordings - a 10 min recording will take some time to process
• the sampling rate of your recording. It is down sampled to 144kHz before analysis. That can take a while if you record say at 500kHz or so.
  I will try on a Raspi when I get around to it. Currently I need a little 'bat' break as it was a lot :)
  Thanks for the feedback on Raspi performance.
  Let me know how well it works on the prediction part once you have tried a few times.

[rdz-oss]

The swedish labels are now in the repo. Use --locale se . Works for EU classifier.

[rdz-oss]

Btw. the sample ( in the example folder) is form a US species .... I included two European bat samples now in the example folder.

[rdz-oss]

I will go out now and record a few bats :)

[Svardsten53]

I have used the Analyzer tool for bird sounds before and then each recording has been an hour and it has worked perfectly fine to analyze in Windows environment. What I didn't think about was that the files with recordings of bats contain a larger sound spectrum and then of course the files also become larger and that was the problem with "long" analysis times - it simply timed out in the analysis. I have now changed so that each file is five minutes long and now it works much better. Problem solved!

[rdz-oss]

It is fun to compare the results with what the app from wildlife acoustics said when recording. The results tend to be more detailed - myotis myotis and pipistrellus pipistrellus fly at the same location and are segmented in a plausible way by the BattyBirdNET. It would be good to make a list of interesting features and prioritize soon.
Such as

• getting the segment.py to work for bats - cutting the recordings into pieces according to detections. This could also be a useful feature for generating ( a lot) more training data
• a graphical user interface adaptation
• Including more areas on the planet - ChiroVox would allow India/Malysia/China and more. Not enough data for a great classifier but maybe enough for a useful one at this point
• Integration into BirdNET-Pi
• I also want to write up a paper on training data set and method

Maybe there are people who would assemble the data in the right way for ChiroVox East Asia or work on the GUI?

[rdz-oss]

SEGMENTATION now available

Once you ran bat_ident.py and have a result file, you can obtain the segmented audio files sorted by detected species using

python3 segments.py --audio path/to/input/audio --results path/to/input/audio/results --o path/to/input/audio/segments

If you have put your files in the folder 'put-your-files-here', the you can use

python3 segments.py

If you use the above folder and know that you want to extract the segments, you can use the new flag alredy when starting the analysis

python3 bat_ident.py --segment on

and it will generate the segemnts to 'put-your-files-here/segments'.

[rdz-oss]

Graphical user interface now available

after installing pywebview and gradio (with pip install)

python3 bat_gui.py

Sould produce a working GUI for single file, batch file analysis and segmentation.
Tested on Mac OS.

[Svardsten53]

I made a quick install on my RPi 4B of pywebview and gradio, ran the command and got the following:

Traceback (most recent call last):
File "/home/svardsten/BattyBirdNET-Analyzer/bat_gui.py", line 5, in
import gradio as gr
File "/usr/local/lib/python3.9/dist-packages/gradio/init.py", line 3, in
import gradio.components as components
File "/usr/local/lib/python3.9/dist-packages/gradio/components/init.py", line 1, in
from gradio.components.annotated_image import AnnotatedImage
File "/usr/local/lib/python3.9/dist-packages/gradio/components/annotated_image.py", line 13, in
from gradio.components.base import IOComponent, _Keywords
File "/usr/local/lib/python3.9/dist-packages/gradio/components/base.py", line 29, in
from gradio.blocks import Block, BlockContext
File "/usr/local/lib/python3.9/dist-packages/gradio/blocks.py", line 28, in
from gradio import (
File "/usr/local/lib/python3.9/dist-packages/gradio/networking.py", line 18, in
from gradio.routes import App
File "/usr/local/lib/python3.9/dist-packages/gradio/routes.py", line 42, in
from fastapi.templating import Jinja2Templates
File "/usr/local/lib/python3.9/dist-packages/fastapi/templating.py", line 1, in
from starlette.templating import Jinja2Templates as Jinja2Templates # noqa
File "/usr/local/lib/python3.9/dist-packages/starlette/templating.py", line 11, in
import jinja2
File "/usr/lib/python3/dist-packages/jinja2/init.py", line 12, in
from .environment import Environment
File "/usr/lib/python3/dist-packages/jinja2/environment.py", line 25, in
from .defaults import BLOCK_END_STRING
File "/usr/lib/python3/dist-packages/jinja2/defaults.py", line 3, in
from .filters import FILTERS as DEFAULT_FILTERS # noqa: F401
File "/usr/lib/python3/dist-packages/jinja2/filters.py", line 13, in
from markupsafe import soft_unicode
ImportError: cannot import name 'soft_unicode' from 'markupsafe' (/usr/local/lib/python3.9/dist-packages/markupsafe/init.py)

[rdz-oss]

Can you try after

pip install markupsafe==2.0.1

Might need -- force
Might be this
https://bobbyhadz.com/blog/python-importerror-cannot-import-name-soft-unicode-from-markupsafe

[Svardsten53]

Installation seemed to work, but got the following when i run the command again:

Traceback (most recent call last):
File "/home/svardsten/BattyBirdNET-Analyzer/bat_gui.py", line 108, in
output_path: str | None,
TypeError: unsupported operand type(s) for |: 'type' and 'NoneType'

[Svardsten53]

Something about my analysis seems very strange. The recordings I made at night show six species of which only two should be in the area. I have compared it to a professional inventory that was done just a couple of years ago in my area, just a few hundred meters away. Confidence value is set to 0.7. I got lots of finds of Rhinolophus ferrumequinum, a species that used to be found in southern Europe and Asia, not in Sweden. Confidence value was in most cases 1.0 or just below.

I then tried to make recordings around lunch when no bats are supposed to be outside and even then got lots of finds, plus two more species.

Has anyone else here made the same observations and what can be done about the problem? I hope it can be solved in a simple way.

[rdz-oss]

A combined results file is now added by default.

[Svardsten53]

The summary file is excellent!

I made new recordings the other night and ran the command python3 bat_ident.py --area Sweden on some of them. Unfortunately, the result was very similar to what I had before the cricket exclusion filter was activated. There were far too many hits and of such species that had not previously been reported at that location during an inventory that was made a few years ago.

[rdz-oss]

Currently only the EU classifier filters out cricket noise!
I collected some more call samples and will make classifiers at 288kHz next. That would make better use of the higher frequencies.
More data can help optimize the classifier.
I plan to make a proper evaluation then.

[davidmam]

Sorry to jump into this so late. I had exactly the same issue a couple of years ago. I was monitoring bats for a path upgrade in an urban environment and got a lot of greater horseshoe bat ID. Hundreds. So I took a look and it was one of these ultrasonic pest deterrent devices that produces a pulsed 40 kHz signal with 80kHx overtones perfectly mimicking the rise - plateau - drop of a Rf call.
Rather annoying noise pollution.

[rdz-oss]

The system runs on the Pi 4B now fairly stable on two test installations (GER and UK).
Give it a shot and let me know if you have issues!
The info on install etc is here: https://github.com/rdz-oss/BattyBirdNET-Pi

[rdz-oss]

NEW: REST API server.py of BattyBirdNET-Analyzer (not Pi yet) now works for bats!

https://github.com/rdz-oss/BattyBirdNET-Analyzer

You can start an API service to analyze your bat calls locally or remotely. You can use it start a web based analysis service for bats or for integrating into other systems such as a BirdNET-Pi running on 144kHz (I intend to provide an out of the box solution for that in the next weeks).
You set the area and other parameters once when you start the API service

python3 server.py --area EU

e.g. using the client.py or from your remote/local application

python3 client.py --i path/to/audio.wav

which yields a reply such as

Requesting analysis for test_data/EU/Kuhls-pipistrelle-Pipistrellus_kuhlii_Ski0145_S2_From1152515ms_To1170806ms.wav
Response: {"msg": "success", "results": [["Pipistrellus kuhlii_Kuhls pipistrelle", 0.9276], ["Myotis capaccinii_Long-fingered bat", 0.00125], ["Myotis dasycneme_Pond bat", 0.0004], ["Nyctalus leisleri_Lesser noctule", 0.0002], ["Pipistrellus nathusii_Nathusius pipistrelle", 0.0001]]}, Time: 0.1159s

[croisez]

just to understand, you are recording off-line, and you do the classification later ? What tool do you use for classifying ? It is not yet Birdnet-pi ? Le jeu. 17 août 2023 à 20:14, Svardsten53 ***@***.***> a écrit :

…

This morning I went to the forest to retrieve my recorder and have analyzed some files. Everything seems to be working very well and two species have been found by the program - Myotis nattereri and Pipistrellus pygmaeus. I also managed to segment and create spectrograms. I show one which according to the program is a Pipistrellus pygmaeus. [image: 0 837_101_test2] <https://user-images.githubusercontent.com/62435779/261398250-4d9887a5-3dc1-4956-b1d3-7681989f7563.png> In a few days I will try to make recordings in another location at an old castle with large open fields near water. Is it possible to add an exclusion list similar to the one found in BirdNET-Pi? It would help to be sure that unlikely species are not included in the results. — Reply to this email directly, view it on GitHub <#300 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANRFSMVXANOFCLTNBV7WVLXVZNRDANCNFSM5WSADX5Q> . You are receiving this because you commented.Message ID: ***@***.***>

[rdz-oss]

Hi srd424,
the BattyBirdNET-Analyzer has a server.py implementation that I did a few touches on to work with the new bat classifiers. You should be able to use it already (if I understand protocol correctly). You would need to configure your system for 144kHz and 1 second analysis intervalls. Then you should be able to send the requests to the BattyBirdNET-Analyzer server.py from e.g. the drop in version for BirdNET-Pi 's own server.py. In principle it should work.

I intend do such an integration for my current fork of BirdNET-Pi , i.e. use the drop in server.py replacement you pointed me to and add a BattyBirdNET-Analyzer on the same device to run the analysis API.

However, I do like the idea of a modularized and dockerized version. Depends on where the BirdNET-Pi project wants to go.

[rdz-oss]

Hi svardsten53, I can follow your points. One thing that has been on my mind, however is:
Come for the birds and stay for the bats, i.e. many people might start with a bird monitor and find it cool if it also tracks bats. Which might lead to a larger number of stations that could feed bat info into something such as birdweather. What do you think?

[dickestonia]

@rdz-oss

This is exactly how I got this far down a rabbit hole regarding diy bird call monitoring 24/7. I would absolutely love to be able to make a diy unit that can do both birds and bats even if it needs 2 microphones.

This would definitely be cheaper than building 2 separate boxes or needing 2 different raspberry pi’s especially for the casual enthusiast like me and my wife.

I definitely came for the birds but love the idea of the bats too and now am looking forward to seeing if this can be done.

[rdz-oss]

Hi @dickestonia
that can be done - birds and bats on one inexpensive device. I know how to do this.
Before I publish the details I am considering to add a few accompanying features to get the best effect for conservation as well as enjoyment.
BR
Richard

[dickestonia]

Vielen Dank @rdz-oss!

Yes one cheap device to do both would be amazing and exactly what I had hoped to find.

I will wait patiently for those results but really wanted to thank you for all of this as a backyard enthusiast.

The amount of time, work and effort that has gone into this has not gone unnoticed as I know it comes at a cost to family/personal time.

Again I say thank you and follow patiently.

Rick

[srd424]

I'm fairly close to having a "proxy" ready for my dockerized version of BirdNET-Pi that would allow it to talk to external analysis servers that talk the same protocol as the upstream BirdNET-Analyzer project, if that helps anyone?

Done, in a very rough and ready way:

podman run -d \
    -v /state/birdnet:/state -v /vol/birdrec:/home/pi/BirdSongs \
    -e BNPI_REAL_ANALYSIS_HOST=myserver.lan -e BNPI_REAL_ANALYSIS_PORT=5055 \
    ghcr.io/srd424/bnpi-server-proxy:test

This plugs into my docker stuff, you run it instead of bnpi-server, and point it at the real BirdNET-Analyzer-type analysis server with those environment variables passed on the command line.
There are probably much more tidy / efficient ways of doing this in the longer term, but this was the quick way to get it going :)

[DD4WH]

Hi! I just discovered this discussion and I am very impressed by this bat detection stuff!

However, there are a few things that appear to me as important suggestions for you in order to make a bat classifier work properly and make it reliable. (My background is in biology and I have worked many years in bat classification for nature conservation work in Central Europe).

• you really need a classifier that is trained with recordings made with appropriate sample rates. The sample rate has to be 250ksps or higher! --> many bats emitt sound up to 150kHz or even higher. In order to record 150kHz sound, you need at least a sample rate double this frequency (Nyquist dictates that). Many bats (especially of the genus Myotis !) can only be identified with confidence, if the higher parts of the calls are present in the recording. Most Horseshoe bats have their loudest frequency at > 80kHz (Greater Horseshoe bat at 86kHz, Lesser Horseshoe bat at 103kHz), so these two species are impossible to identify with sample rates less than 250sps (which allows recordings of frequencies up to 125khz).
  --> Thus I recommend using a sample rate of at least 250ksps (better 384ksps or even 500ksps) for all training of your classifiers !

• Spectrograms for bats only make sense, if the time axis is presented in much higher resolution, Bat calls are only a few milliseconds in length and the detail relevant for classification can only be seen in much higher time resolution

• you really need a threshold detector, which constantly checks the acoustic energy inside the ultrasonic bands (say from 16kHz to 120kHz, using real time short term FFTs for this purpose is very straightforward). Every time one of the FFT bins contains energy higher than a predefined/user-settable threshold, a recording is made for a few seconds. Only then this is recorded and sent to the classification chain with the bat classifier. --> This works really nicely and is also written in python and contains everything it needs for such a "threshold detector": https://github.com/cloudedbats/cloudedbats_wurb_2020
  and

• @Svardsten53: I am sceptical about the bat species that your classifier identifies at your place: Myotis nattereri cannot be identified with a classifier at a sample rate of 144kHz, because this species is only reliably identified by its VERY high frequency start of the call, which is above 120kHz or even 130kHz. A sample rate of 144kHz can only sample frequencies below 72kHz, so it is impossible to identify that species with such a classifier. Also your spectrogram of Pipistrellus pygmaeus does not look like that species at all. Ppyg has a lowest frequency of 55kHz in its calls.

So, to sum up:

• I really like your approach and it would be superb to add bats into BirdNET-Pi
• But I really would like to see it properly implemented, so that the sample rate is high enough to be even theoretically capable of recording the entire bat call (up to at least 125kHz --> which means a minimum sample rate of 250ksps)

Best regards, Frank Dziock DD4WH

[rdz-oss]

Thank you Frank for the Info!
The system should also be able to work on e.g. 240kHz , 288kHz or 360kHz (Nyquist would allow for about 180kHz sounds to be registered). So there is hope :)

I can build it into the classifier (in the future) so that it can be set when it is called:
python3 bat_ident.py --freq 240000
with values: 144000, 240000, 288000, 360000 depending also on the microphone used in the recording and the need for precision in the classification. Perhaps 288000kHz is a good default setting - 144kHz bat sounds included (caveat: does it work in real time on a Raspi 4?).

The US data is well suited, that for Europe:
Some of the training data goes to 500kHz - well in the range up to 250kHz calls of bats included.
Other data is on 390kHz (185kHz calls included), yet some is only up to 50kHz (recorded at 100kHz) and recorded 20 years ago :). So a mix.
More data would help ! Do you perhaps know of a source?

For the bats in my area in Germany, I compare results to the wildlife acoustic echo touch 2 and it is very similar. I have not recorded calls higher than 100kHz as yet - the hardware can record calls up to 128kHz.

That said I can train the classifiers for 240kHz and 360kHz and we can see what happens. I can do this in about two weeks as I am traveling and have the original data at home.

In the present case of @Svardsten53, there might still also be a hardware issue from the AudioMoth that I would like to exclude. If I get an original sample, I should be able to test for microphone artifacts. There appear to have been some in a setup in California, perhaps also interesting for you to peruse: rdz-oss/BattyBirdNET-Analyzer#1

The question remains: what is good enough? Depends on the purpose of use.

I want to integrate the classifier into the BirdNET-Pi system - as proof of concept. Then get back to the classifier and make an evaluation aimed for peer review quality. In time :)

Of course, once the technical heavy lifting is done, anyone who has good data can offer another classifier :)
Would be a cool Kaggle competition, too.

[DD4WH]

Very good !
I am not quite sure how you train the classifier.

Do you resample the files to the desired sample rate and do you use an appropriate anti-alias filter in advance of the resampling?
I have to think a little more about the time resolution you would achieve with your approach. Did I understand that correctly:

• original bat call file with 500ksps sample rate
• lowpass-filter @180kHz
• resample to 360ksps sample rate
• input to BirdNET-Analyzer: 0.4sec of original sound "stretched" to 3 seconds input at a pseudo-sample rate of 48ksps == 144.000 samples ?
• What is the resulting time resolution then ? (and would it be sufficient to discriminate very short highly frequency-modulated bat calls which can be as short as 1 or 2 milliseconds) ?
  I will think more about that . . . :-)

[rdz-oss]

I can tell you my current procedure:

• All automated by scripts - can be rerun for different settings easily
• US- NABAT data as source
• Europe: animal sound archive berlin and chirovox combined
• The data is from different sources, yet every single call is accounted for by meta-data (who, when, where, mic, ...)
• Resampled to the same rate, e.g. 144 kHZ or also I have done it for 240000. High pass filter can be set. Currently 20 kHz to cut off people talking etc. Can be 0 or say 17 alternatively
• For the data that is recorded at 96kHZ, I upsample to 144kHz and cut away artifacts over 48kHz
• The other data is effectively down sampled
• The BirdNET is presented with vectors of 144000 length (3 seconds at 48000kHz). So at 144000 it is 1 one second recordings, for 288000 it would be 0.5 second segments, 36000 would be 0.4 second sections.
• My intuition is that bird net believes a bird made these sounds in 3 seconds. Even if it was a bat in e.g. 0.4 seconds. That should be a time factor of 3/0.4= 7.5
• Embeddings of BirdNET are then classified to bat origin

Now BirdNET uses image classification for analysis. My hunch is it might even be better not to focus on a single call line as humans do - but clusters of them that are correlated. The convolutional neural network can take care of what actually matters for identification better than human experts. If the data is good and sufficient in amount.

Side note: It appears we both worked at the Center for Environmental Research Leipzig - UFZ at about the same time :)
My PhD Advisor was Volker Grimm

I much appreciate all the feedback - including from Sweden - I have a civilingenjor degree from Chalmers TH in Komplexa adaptiva system and many good memories of living in Gothenburg

[rdz-oss]

Hello everyone! Thank you very much for your comments and feedback. The classifiers https://github.com/rdz-oss/BattyBirdNET-Analyzer are now

• based on 256kHz (up to 128kHz sound is registered)
• trained on cricket and 1GB of environmental noise to reduce false alarms
• based on more data: animal sound archive berlin, chirovox and xeno-canto
• can be accessed by command line, GUI, REST service
• and can be tested in an online demo for small (3-5mb) files!

BR
Richard

Try the online demo if you like:

-> https://huggingface.co/spaces/Amazetl/BattyBirdNET-Analyze-Demo

Check it out and maybe also compare results to batdetect2 https://huggingface.co/spaces/macaodha/batdetect2

[davidmam]

Maybe Scottish P.pyg have a different accent as we see them typically down to about 50kHz.
I am not sure that upsampling recordings would work at all reliably for this.

[croisez]

My own coding contributions to Birdnet-Pi were mainly to introduce a new shifting frequency feature, to allow the user to hear birds songs initially too high in frequency. As I don't hear myself upper than 7KHz, I was very motivated. A lot of birds produce sounds up to 12KHz and more. This feature is now accessible for off-line MP3 or WAV files, there is a button to shift freq, but also now for real-time audio listening. I wonder if the two ways I proposed the freq shifting (sox and ffmpeg) could be compatible with such very high frequencies produced by bats. I had a discussion with Patrick McGuire on this list about frequency shiftings of bats WAV files. I think it would be of interest that a user could hear bats songs. If you are interested in this new feature for bats as well, I am available to adapt my code for it. If yes, could you send me some WAV files (2s chunks, for limiting size) by mail, so I could try some python code on it ? Le mar. 22 août 2023 à 21:18, Dr. Richard Zinck ***@***.***> a écrit :

…

Segmentation and mel spectrograms available You can get segmented .wav files by species with a spectrogram plot (if you install sox) by using e.g. python3 bat_ident.py --area USA --spectrum on Substitute USA for your regional choice . The spectrograms can help you to better asses if the classifier is likely right - or perhaps off. I will provide classifiers for 288kHz also in maybe two weeks. Sample spectrogram that is generated and classified as northern bat (what do you think?) [image: 0 987_2_EPTNIL_20230526_001811] <https://user-images.githubusercontent.com/63958211/262459994-85134976-a467-41ca-8642-41a7144beae0.png> — Reply to this email directly, view it on GitHub <#300 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANRFSMZZCVPAADPGGFEWETXWUAY3ANCNFSM5WSADX5Q> . You are receiving this because you commented.Message ID: ***@***.***>

[rdz-oss]

That is a good idea. Can you also use a download link ? A few samples are here: https://cloud.h2887844.stratoserver.net/s/J3FSfQiLaCnkFQs

[croisez]

Hello Richard,
you can have a look at my first results.
I am using a python script that Patrick McGuire sent to me a year ago.
The procedure is quite simple:

1. use a Butterworth high-pass filter to only keep energy from 30KHz.
2. translate signal to baseband by multiplying with a sine signal

For producing the *_shifted.WAV resulting files you can listen to in this link (https://we.tl/t-mxmFbht16o), I have translated the signal 29KHz down.
Concerning *_shifted.MP3 files, it has been produced from *_shifted.WAV, by shifting it 12KHz down and convert it it MP3 (sr=48KHz). I made this to enhance the audio signal a little more into low frequencies, to help me to hear it.

Can you tell me your advice ? Could this procedure be of interest ?
Louis

[rdz-oss]

I am currently traveling so sorry for late reply. I like the idea. Would you mind sharing your scripts?
I would have a look at them and see if an option can be added to the bat_ident.py function that generates audio files with your settings as presets.
I found some more bat call data and will upgrade the EU classifiers soon. Then it should be good enough for the next step - integration to a pi system.

[croisez]

Here is the Python script.

## Script written by Patrick McGuire, modified by LM Croisez ##
import sys
import numpy as np
import librosa
from scipy.io import wavfile
from scipy import signal

def butter_highpass(cutoff, fs, order=5):
    nyq = 0.5 * fs
    normal_cutoff = cutoff / nyq
    b, a = signal.butter(order, normal_cutoff, btype = "highpass", analog = False)
    return b, a

def butter_highpass_filter(data, cutoff, fs, order=5):
    b, a = butter_highpass(cutoff, fs, order=order)
    y = signal.filtfilt(b, a, data)
    return y

def bat_filter_shift(ifile, ofile, filterfreq=30000, freqshift=30000):
    x, sr = librosa.load(ifile, sr=None)
    y = butter_highpass_filter(x, filterfreq, sr, 5)
    T = np.arange(len(y)) / sr
    heterodyne = np.sin(2 * np.pi * freqshift * T)
    heterodyne *= 1.5
    heterodyne = np.int16(heterodyne)
    heterodyne = np.multiply(heterodyne, y)
    wavfile.write(ofile, sr, heterodyne)

if __name__ == '__main__':
    # Here get parameters when testing the functionality from command line.
    # NB: Highpass filtering from 30000Hz seems to give good results.
    #        Anyway this should be a parameter to add to Birdnet-Pi configuration page.
    #        Generally, we can safely shift-down the spectrum the same value we HiPass-filtered.
    ifile = sys.argv[1]
    ofile = sys.argv[2]
    FREQSHIFT = 30000 #Good working parameter value
    bat_filter_shift(ifile, ofile, filterfreq=FREQSHIFT, freqshift=FREQSHIFT) #Function to call from 'bat_ident.py'

Cheers,
Louis.

[rdz-oss]

Hi @croisez I integrated the classifier into BattyBitdNET-Pi. The frequency shift mechanism viavffmpeg works like in BirdNETz- Pi you can eg set a mapping from 40000 hertz to 4000.

[rdz-oss]

Check out the online demo for small files ( a few seconds of bat calls) now on huggingface.co:
https://huggingface.co/spaces/Amazetl/BattyBirdNET-Analyze-Demo
If you are from the UK, why not cross check to batdetect2
https://huggingface.co/spaces/macaodha/batdetect2
Enjoy and let me know how it works for you!

[Inachis]

Hi,
Just been throwing a few of some old recordings - made in North East Hampshire, UK in 2018/19 with a Wildlife Acoustics SM4 - at https://huggingface.co/spaces/macaodha/batdetect2 which is producing identification consistent with Kaleidoscope Pro and BatClassify analysis made subsequent to the recordings, whereas https://huggingface.co/spaces/Amazetl/BattyBirdNET-Analyze-Demo is producing unusual/wrong identification with area set to EU.

Let me know if you need more details or specific recordings...there's about 8000 in total - I did manage to 'break' batdetect2 with one recording.

[rdz-oss]

Hi Kelvin,

first of all let me thank you for the feedback! This is very helpful in getting the classifier towards a reliable state.

• I also noticed mixed results. I have made a few changes at once (add data, change frequency, changed preprocessing steps) and will need to exclude the possible factors one by one.
• Your samples would be very useful as a filter for that. So it would be awesome if you could get as many samples to me as possible - as an independent verification set. But if only a few samples make it that is also already good.
• Also, if you have that many (classified) recordings, would you be open to the use of some of that data for training the classifiers? That would make a great difference. The data I collected so far is all under the CC share alike no commercial use license. Would that also work for you?
• http://creativecommons.org/licenses/by-nc-sa/4.0/

I made a temporary email address where you can reach me so that we can talk about details of file transfer etc: [email protected] should be forwarded to my actual email account.
There is also a file drop where you can upload some samples quickly:
https://cloud.h2887844.stratoserver.net/s/Cq3e9C7x5KMsLBf

BR
Richard

[Inachis]

Hi Richard,
Did you get the emails I sent on 28th and uploads too? Just wanting to confirm that they got through.

Cheers,
Kelvin

[rdz-oss]

Yes, thank you. I replied via our new channel !

[rdz-oss]

I made a mistake when changing the demo on huggingface to 1444kHz version. In effect, it was running a hybrid between 144 and 256 kHz. Should be fixed now. My appologies. Some odd results should have been due to this.

[rdz-oss]

Help appreciated to fill up the data gaps!

Hi - I would like to tell you a bit about what is happening. I am back at data engineering. Thank you @Inachis for the data from the UK you sent me! However, the training data is currently heavily imbalanced among species. Please consider helping if you have data for the species with less than 100 samples (see plot. The common pipistrelle is not in the plot as there are already 1600 samples). I would then point you towards an SFTP server where you could upload your data.

The classifiers will only be as good as the data ! If you can help, please consider doing so.
Ideally: 256kHz or more

If a data donor agrees, I would add some (depending on quality etc.) of your donated bat call data to the training set under the non commercial license that e.g. also BirdNEt-Pi is under: http://creativecommons.org/licenses/by-nc-sa/4.0/

I post that also in the repository of BattyBirdNET-Analyer where I cross train the BirdNET to identify bats. This license prohibits commercial use. I personally have no commercial interest in this endeavor - for me, I am curious and know how to do these things (ML and system development) so I decided to go ahead and try.

This license is also what the other data I assemble is under. This way, there should be legal clarity on the matter. So, even if some third party would use it, they would be bound by the terms of this license. For example, if some researcher or other open source initiative were to use the data in training, they would have to adhere to the above license. Reuse of the data by others would actually be really cool because it would mean that more open source, non commercial work would be done on the topic. Only with good data will there also be good non commercial, open source classifiers for all.

[BobSpong]

I can help with barbastelle and horseshoe calls from a song meter mini bat @ 256khz

[rdz-oss]

Data for barbastelle and horsehoe bat would be great! Would you kindly send me an email at a temporary address [email protected] ? Once we are connected I would forward you the details of an SFTP server. You could use e.g. FileZilla or WinSCP or the like to upload the data.

Let me know if the email address works - I can only get a few email a month via this so we might have to go a different way.
BR
Richard

[guillaumarchais]

Bonjour, I have a number of bat ranging recordings with EMT2 from WA collected from France, northern Spain; also specifically of P. kuhlii from Croatia (Split).
I would strongly suggest to add a country or region tag to each species dataset (Mediterrean / west EU / East EU / Scand EU) as they can shift of few kHz for their peak frequency.
PS : I do have FileZilla

[rdz-oss]

Hi @guillaumarchais ,

I try to maintain the meta-data for the samples. At this point, however, the total number is still too small to partition the training sets by region. Thank you for offering data - would you kindly send me an email to [email protected] and I will forward you the specifics via your email address.

BR
Richard

[guillaumarchais]

Fair enough, but this is something to bear in mind.
Basically, an imaginary line from Bordeaux to Lyon would help in distinguishing various typical calls from the same species.

• Example 1 : Pnathusii_NorthEU has Fmax = 35-38 kHz for transit calls ; Pnathusii-SouthEU has Fmax = 39-41 kHz (as Pkuhlli is very frequent there).
• Example 2 : Ppipistrellus_NorthEU 46-51 VS Ppipistrellus_SouthEU 42-47 (as Ppygmaeus is very frequent 47-60, and Pkuhlii too 34-41)

see this article for Ppygmaeus plasticity in Northern Spain (can go down as 44 kHz !) https://link.springer.com/article/10.1007/s00265-021-03002-7
this one in Germany for Ppipistrellus (Urban VS Rural) : https://www.frontiersin.org/articles/10.3389/fevo.2022.939408/full

(and I am not talking about differences between adults and youngs ; also near and far from roosts).

To sum up... as a minimum North/South tags added to species names would be helpful to limit major errors of Auto ID.
Alternatively, it would only auto label as "pipistrelle" genus level.

[rdz-oss]

INFO on integration of BattyBirdNET-Analyzer to a PI

To everyone interested in getting a BirdNET-PI for bats: I am making progress on the integration of the BattyBirdNET into a test system.
The approach is:

• Use wildlife acoustics echo meter touch 2 as mic
• Install BattyBirdNET-Analyzer on the PI (working)
• Set it up as via its server.py as a systemd service (working)
• Install a modified version of BirdNET-Pi at 256kHz (working, buggy at this point)
• Analyze the audio stream for bats via a BirdNET-PI-server.py to BattyBirdNET-Analyzer-server.py bridge
  Status: they find each other, can call each other, but the interface is not as expected and needs tailoring (TODO)

Then, in principle, it should work. As I am working on this on the weekends while the kids jump around process is slow but steady.
If you want to check it out or help get the system up you can look/contribute here:

https://github.com/rdz-oss/BattyBirdNET-Pi

BR
Richard

[rdz-oss]

Hi @BobSpong ,
I sent you the details for the sftp server. Hope that works for you, let me know if you need some help getting started.
Also: I ordered an audiomoth v1.2.0. I continue with the echo meter 2 development, but am open to try the audiomoth once the time is right. I like the open science/product approach behind it.

BR
Richard

[guillaumarchais]

Bonjour, I use the BirdnetPi since last year (https://app.birdweather.com/stations/541).
I also have an Audiomoth 1.2.0 and an EMT2.
I am not good with programming though as I am more an ecologist slowly getting into geek apps stuff.
Guillaume

[rdz-oss]

Update: #300 (reply in thread)

[kjarkema]

Hi @rdz-oss, any idea if this would work with a Echo Meter Touch Pro 2 with the lightning port? It wont work with an Android phone combined with the Lightning/USB-C adapter..

Also can it run on the same Raspberry?

Sorry, total noob. but willing to experiment and maybe add data.

[rdz-oss]

I am not sure about the lightning port, yet there is a test installation now in the UK using the Touch Pro 2. So I expect we will know soon how that works. It might overwhem the system if it cannot be set to work at 256kHz. But that remains to be seen.

[Mattmanandeddie]

This is fascinating and great work.
Can I make an on topic but off class type of animal idea change?
I've been reading this thread and would love to try out bat detection but my cheap mic does not pic up the correct frequencies. I've been looking at my logs at night and noticed something cool. BirdNet Pi detects frogs, toads and some noisy night insects. Can this info easily be added to the webUI to analyze? Maybe on a separate tab.

[rdz-oss]

Hi @Mattmanandeddie the detections fo other species (e.g. Chipmuncks) are also shown in https://github.com/rdz-oss/BattyBirdNET-Pi if set to detect birds during the day. Shifts to detection of bats from sundown to sunup.

[croisez]

hello Richard, you did verify that the frequency shifting as it was implemented with ffmpeg is working well ? The script that i sent to you (and this thread) was implementing a more reliable way by first using a hipass filter and then shifting the high freq content to audible frequencies. I am not convinced that ffmpeg shifting process is compatible with batts very high freq contents Le sam. 14 oct. 2023 à 20:55, Dr. Richard Zinck ***@***.***> a écrit :

…

Hi @croisez <https://github.com/croisez> I integrated the classifier into BattyBitdNET-Pi. The frequency shift mechanism viavffmpeg works like in BirdNETz- Pi you can eg set a mapping from 40000 hertz to 4000. — Reply to this email directly, view it on GitHub <#300 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANRFSN3UH72ST5NT7ZAYLLX7LNZDAVCNFSM5WSADX52U5DIOJSWCZC7NNSXTOKENFZWG5LTONUW63SDN5WW2ZLOOQ5TOMRYGIYTEMY> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[rdz-oss]

It did not - so i changed it to use sox speed - now it sacles by factor 10. 10 times slower and deeper. This is working well for the recordings. It does not work for the live recordings at this point as that is implemented with in browser technology. Apparently, every browser implements their own solution and I have not got them to go above 48000 Hertz as yet. That may be a limitation of browsers at this point.
However, scaling by factor 10 for recordings with sox is working well now. I will make the factor configurable next. Currently it is fixed to factor 10.

[EZStreamer]

I was curious if anyone has tried BattyBirdNET-Pi with the new "AudioMoth USB Microphone"? Link: https://www.openacousticdevices.info/audiomoth. I provisioned a BattyBirdNET-Pi system on a Raspberry Pi 4 4GB using this AudioMoth mic and it's didn't recognize it. I assumed it would work after reading this: "AudioMoth USB Microphone is a variant of the standard AudioMoth 1.2.0 design which removes many of the components which aren't needed when an AudioMoth is being used as a USB microphone. It's designed to specifically run the AudioMoth USB Firmware and uses an identical audio front-end to the AudioMoth 1.2.0 and AudioMoth Dev variants."

I have placed an order for the original AudioMoth 1.2 to see if I can get a BattyBirdNET-Pi system operational in the meantime.

[BobSpong]

Just hooked mine up (had a fully fledged audiomoth connected before the winter) and looks good. Now for the bats to wake up!

[rdz-oss]

A first reply from me here: rdz-oss#1 (reply in thread)

[rdz-oss]

UPDATE: There have been significant improvements to the https://github.com/rdz-oss/BattyBirdNET-Pi
system including:

• start stop automation at dusk/dawn
• noise reduction for mics audiomoth 1.2, WA echo meter 2 and echo meter 2 pro
• improvement of robustness against noise in Bavaria and US classifiers (rain, car, birds, ..)
• Classifiers for US-EAST and US-West

Check it out if you like and let me know your thoughts!

BR
Richard

[rdz-oss]

Works for bats AND birds now: https://github.com/rdz-oss/BattyBirdNET-Pi
Bats at 384kHz
Add your own Classifier (bat or bird) possible

[FreshNiagaraMushrooms]

I'm just beginning my deep my deep dive into bioacoustics and AI.

I live in the Niagara Region, Ontario, Canada.

I'm teamed up with a local biologist familiar with Wildlife Acoustics systems, and exploring a bit deeper.

Our goal is to use the data to get a baseline of different areas of Niagara, as we have 60% if Canada's most fertile agricultural soil within a one hour drive, so lots of birds, bats and bugs.

Very interested in how these bioacoustic baselines will compare and contrast between different land management practices, like regenerative agriculture/permaculture vs. traditional farming.

This seems useful for general public education as well as a qualitative analysis informing local farmers land management practices and not just their effects on the biosphere, but ultimately their farming costs.

This may be a way to prove that a healthy biosphere increases yield and lowers more costly inputs.

I'm very excited to see your work!

Looking for something that can output to pretty website dashboards for the general public, but also be used as a comprehensive bioacoustic database to monitor different areas in Niagara.

Any help or support is definitely welcome.

I'm hoping we can use this information to help Niagara's Wineries resist the invasive Lantern Fly by supporting insectivores.

MushLove,

Craig the Mushroom Farmer