Biosonic - An interview with Josef Carlson
Can you tell us a little bit about yourself?
I've always been interested in birds and nature. Ever since I was a child I was fascinated by nature and the way birds communicate with each other. It's amazing how complex their language can be.
I'm not an ecologist, I come more from the engineering side. I've been into coding for a long time. About two years ago my brothers and I wanted to use our coding and AI knowledge for biodiversity, which we are passionate about. That's how Biosonic was born. Interestingly, we started with birds, but now our days are filled with bats.
I work on Biosonic together with my brother Jonathan, specialist in AI image recognition. And our third brother Simon is doing his PhD at Cambridge, using very similar image recognition models to detect cancers like leukemia in blood cells. So you could say image recognition runs in the family.
Can you explain what Biosonic is?
BioSonic helps bat consultants with AutoID of bat species with high accuracy. Even separating social and feeding calls.
Identifying bat species is much harder than identifying birds, for one simple reason: With birds, you can use both your eyes and ears. You see the bird, you hear their song, you can know which species it is. With bats you can't see them because they are out at night, and you can't hear them because their calls are ultrasonic. So to identify bats you have to record their sounds and play them back much slower to actually hear anything.
If you record every night for a few months you end up with a huge amount of recordings. Traditionally, experts went through all of those files by hand, listening to the slowed-down audio and looking at spectrograms to work out which species is which. When you have months of recordings, that analysis takes forever.
Many bat consultants now use BioSonic to do that work much faster. Our AI filters out all the non-bat sounds and classifies the bat recordings with high accuracy. In the end, a human expert validates the results. That saves a huge amount of time on the boring parts so the ecologist can focus on the interesting species.
On the Biosonic website, you say you try to make AI "see" the spectrograms. What do you mean with that?
Most traditional bat algorithms rely on simple triggers. They assume bat calls are simple enough that you can define rules like "if the frequency is above 52 kHz it must be a Soprano Pipistrelle."
In reality, bats are much more interesting than that. They throw out social calls and feeding buzzes. They change their frequency in narrow spaces, and when they meet another bat they go a bit crazy. So a rule-based algorithm quickly breaks down in real recordings, especially when there is noise or multiple species in the same file.
What bat workers have known for many years is that looking at the spectrogram as an image is actually faster and more accurate than listening. They see a shape and say "I know that shape, that's a Serotine bat.
So we thought, why not train an AI to do exactly the same thing? My brother Jonathan is a specialist in AI image recognition, and our brother Simon is doing his PhD at Cambridge on image recognition with AI on cancer images. That combination was perfect with this approach.
When we built Biosonic on image recognition of spectrograms, the accuracy jumped significantly. A third party benchmark by Wilder Sensing and Annabel Jeffries showed that a commonly used software had 84.2% accuracy (F1) on their benchmark. BioSonic soared past it to 98.9%.
Are there other companies doing something similar, or is Biosonic unique?
The AI approach we use took a long time to develop, and only a small number of people have the deep knowledge needed to build this. Also we have a huge amount of training data that we have been building for years with bat workers from many countries.
That’s why I'm very confident that no one will come close in terms of accuracy for a while. Of course more companies will try, because this field is growing fast, but we have a big lead and increasing this lead with launching new AI models 3 times per week.
You used 2.5 million audio files to train this AI model. How did you get all these files?
We’re fortunate to be collaborating with t consultancies and researchers in Denmark, the Netherlands, the UK and other countries.
On top of the 2.5 million bat files, we also have around 3.5 million noise files in the training set, which is very important. That's what makes Biosonic good at filtering out rain, wind, grasshoppers and all the other things that can look and sound a bit like bats.
For reference, that is roughly 10 times more training data than the second largest algorithm we know of.
How does Biosonic save time on bat analysis?
Imagine you're working with 100,000 audio files. Around 80,000 of those will probably be noise, like rain or grasshoppers. In traditional software you still have to go through many of them to make sure you didn't miss anything.
Biosonic filters out most of that noise automatically. Then, of the bat files that are left, a huge part will be very common species like Common Pipistrelle that the AI recognizes with very high confidence. Those get marked as "AI approved" and you can trust them without reviewing each one.
What you end up with is a small set of files that the AI is uncertain about, or rare species, or interesting behavior. Those are flagged in red so you focus your time there.
The platform also keeps projects organized. Multiple people can work on the same project at the same time, leave internal comments, and discuss specific recordings inside the system instead of emailing WAV files back and forth.
Why is this kind of analysis so hard on the ecologists who do it manually?
Honestly, it's tough. People sit for weeks pressing the same button, file after file. I have heard from many ecologists that they develop neck and shoulder pain because of how they are sitting and clicking for so many hours. And nobody really enjoys spending their day checking 10,000 files of grasshoppers.
That's part of why I'm on a mission to get everyone working with bats to move to BioSonic. It's partly about saving time, but it's also about people's health. Ecologists should spend their time on the interesting parts of their job like fieldwork, not on repetitive clicking that gives them a sore neck.
How much time can Biosonic actually save you?
We had an ecological consultancy called Wharton in the UK do a proper comparison. On 2 terabytes of data they compared Biosonic against their previous software and saved 327 hours, just on the sound analysis. That's around 66% time saved.
Other consultants tell us they save up to 86%, depending on the project. So it's usually somewhere between 66 and 86% on the analysis itself. And that's before you count the time saved on writing reports, because Biosonic also generates all graphs, heatmaps and tables needed for reports automatically.
Does it take GPS into account as well?
Yes. Biosonic works across multiple countries and not every region has the same bat species. GPS data helps make sure the species classification is geographically accurate. A bat that is common in the Netherlands might not exist in the UK, so that context matters.
GPS also unlocks the visual side of the platform. From the locations of the files you can create maps directly in Biosonic, heat maps of bat activity, species distribution, emergence times, and so on. All of that feeds straight into the reports that ecological consultancies write for their clients.
We always think of Biosonic as taking you from the raw sound files to a Word report as fast as possible. That whole chain matters.
What adjustments were needed to make Biosonic fit into the Dutch ecological consultancy world?
In the Netherlands ecologists often do transects on foot or by bicycle rather than leaving a static recorder for weeks. So extracting the GPS of every single file along the route was essential. In other countries, stationary recorders are more common, so we handle both.
For SMPs (Soortenmanagementplan) recordings have to be done within specific time windows, and the reporting has to follow strict rules. We built that workflow into BioSonic with the help of Dutch consultants. For example, you can tag a colleague on a microphone and say "this needs to be reviewed within 5 days for the SMP." If they don't review the species on time, they get an automatic email reminder.
We shipped that deadline last Friday, actually. That's the pace we try to keep. If a consultant tells us on a Monday that they need something, it can appear for them in BioSonic by the end of the week.
Is it difficult to start with Biosonic?
Not at all. We have free trials at https://biosonic.io/ so anyone can explore the platform. You create an account in your web browser, upload a few files, and see the results for yourself. That's how most people get started, even before talking to us.
We've tried to make the software as self-explanatory as possible. When you log in you get a short tutorial video and a few guides, and then you just upload. If you want more support, you just call me on Whatsapp or book a meeting with me and I'll walk you through everything in detail. That's also useful for me because I learn a lot from how different ecologists work and how to change and add features.
We also set up a WhatsApp or Teams group for every customer with their whole team. If anyone has a question, I often respond with a quick recorded video showing exactly where to click. We try to be very fast, because every question is also a chance for us to improve the software so the next person doesn't have the same problem.
For Veldshop customers, you can just sign up directly through the Veldshop link and get started You get a 7% discount as well!
https://bat.biosonic.se/signup/veldshop
One more collaboration question. What if my team is big?
Totally fine. Some of our customers have 100+ ecologists working in the same Biosonic organisation at the same time. Everyone has the same access, you can see who has reviewed which files, tag colleagues for a second opinion, leave comments, and mark files for review. You have a “chat” on each file, to directly communicate (instead of sending files through emails back and forth)
We don't charge per user. You get unlimited people in your organisation, because we think collaboration shouldn't be something you pay extra for. Even people that are not working with BATs can get added and download the graphs, reports, and check-in, such as the project managers
What are the plans for Biosonic?
Right now my focus is simple. I want every bat consultant in the Netherlands, the UK, Denmark, Germany to know that they can work 3 times faster with BioSonic.
Longer term, we are adding more species. We actually started with birds before bats, so the bird AI model exists. We pulled it out temporarily to focus fully on bats, but in a few months we will bring birds back. Frogs are also coming.
The bigger picture is this; Scientific research is really clear that sound is one of the best ways to measure biodiversity. Most animals make sounds, and sound is objective and scalable in a way that visual surveys often aren't. If we can measure biodiversity well, we can actually do something about improving it. That's why we do this.
We hope BioSonic can be a small part of restoring biodiversity together with ecologists.
-
Interested in trying Biosonic? Veldshop customers can sign up directly through our Veldshop partner link: https://bat.biosonic.se/signup/veldshop
or book a demo with Josef here: https://cal.com/biosonic/30min
