---
canonical: https://biosonic.io/compare/bsg-bats
---

# BioSonic vs BSG-BATS: an open-source classifier and a ready-to-use service

Last updated: 2026-11-19. Written by the BioSonic team.

BSG-BATS is a free, open-source neural network for 21 European bat species, built in the LIFEPLAN programme at the University of Helsinki and published in Methods in Ecology and Evolution in 2026. You download it from Zenodo and run it in Python. BioSonic is a Swedish web service that identifies European bat species in the browser and gives a team one place to review, approve and report. In our head-to-head test on 3,159 manually verified recordings from Denmark, BioSonic named the correct species for 94% of bat recordings against 66% for BSG-BATS (87% on the species BSG-BATS covers), and reported 1% of noise as bats against 95%.

They answer different needs. BSG-BATS is a research tool you can inspect, retrain and run on your own machine. BioSonic is a finished service for survey teams who want the sorting, checking and reporting done in one place. Below are the facts, with sources.

## At a glance

Facts about BSG-BATS come from the peer-reviewed paper, the Zenodo record and the authors' own posts, listed under [References](#references).

| Feature | BioSonic | BSG-BATS |
|---|---|---|
| What it is | Web service for automatic bat species identification, team review and reports | Open-source CNN classifier plus a web portal for annotating training data [1][2] |
| Made by | Biosonic AB, Stockholm, Sweden (founded 2024) | Researchers in the LIFEPLAN programme, coordinated at the University of Helsinki; paper by Meramo, Somervuo and colleagues [1][4][5] |
| How you run it | In the web browser on any operating system; nothing to install | Download a 469 MB archive from Zenodo and run it in Python (PyTorch) on your own computer [2] |
| Works offline | No, analysis runs on BioSonic's servers | Yes, once installed [1][2] |
| Species | European bat species, including all 17 breeding UK species | 21 European species in version 0.21; five phonic groups annotated but not yet used in the test [1][2] |
| Training data | More than 2.5 million verified bat calls and 3.5 million noise files from across Europe | 4,042 annotated 10-second segments, much of it from Finland, Spain and Italy [1] |
| Model updates | New model every weekend, used automatically | Retrained as the community adds annotations; you download new versions [1][3] |
| Confidence and overlaps | Confidence score for every call; a second species is shown when two bats overlap | Confidence score for each species; multi-label, so overlapping bats can each be named [1] |
| Reviewing results | Spectrogram review in the browser; confirm or correct each call, every change recorded | No review interface; you check the output in your own tools |
| Reports and maps | Species tables, activity charts, maps and a Word report from the project | Raw predictions; tables and reports are up to you |
| Team working | Whole team on one organisation account, shared projects | Shared through your own folders and scripts |
| Openness | Proprietary model | Model, code and training annotations open; CC BY 4.0 licence [1][2] |
| Cost | Package of analysis GB, lower price per GB for bigger packages; recordings kept for 9 years | Free; you provide the computer and the time [2] |
| Published accuracy | 98.7% F1 in a third-party benchmark by Wilder Sensing, against 83.0% for Kaleidoscope Pro [6] | Highest average AUC of six tools on the authors' own 1,984-clip test set [1] |
| Head-to-head test (our own) | 94% correct species and 1% of noise reported as bats, on 3,159 verified Danish recordings | 66% correct species (87% on the species it covers) and 95% of noise reported as bats, on the same recordings |

## How each tool works

### BSG-BATS: an open model you run yourself

BSG-BATS has two parts. The first is a web portal, part of the Bird Sounds Global system, where bat experts label calls in 10-second recordings at species or phonic group level [1][3]. The second is a convolutional neural network trained on those labels: six convolutional layers and two fully connected layers, written in Python with PyTorch and librosa [1].

The model slides a one-second window along each WAV file in half-second steps and gives a confidence score for every species, plus a background class for noise. It is multi-label, so it can name more than one bat in the same window [1]. The authors report that a 10-second file takes under a second on one laptop CPU core, and the full 1,984-file test set about two minutes on four cores [1]. The model, code and documentation are on Zenodo under CC BY 4.0 [2], and the authors invite researchers to add recordings and annotations so the next version covers more species and regions [1].

### BioSonic: the whole survey workflow in the browser

You upload a season of recordings in the browser, from 1 GB to thousands of GB, and the analysis runs on BioSonic's servers. A deep-learning model trained on more than 2.5 million verified bat calls and 3.5 million noise files from across Europe filters out noise and gives every call a species and a confidence score. A new model goes live every weekend.

The results land in a shared project. Your team filters by species and confidence, listens and looks at the spectrogram, confirms or corrects each call, and approves the night. Every change is recorded against the person who made it. Species tables, activity charts, maps and a Word report come out of the same project.

## Accuracy and evidence

### Head-to-head on Danish recordings

In November 2026 we ran BSG-BATS (version 0.21, the released six-model ensemble with its own code, as published) and BioSonic on exactly the same 3,159 manually verified recordings from Denmark, made with six detector types. None of the recordings was used to train either model. These are the same recordings and the same scoring as in our [BatDetect2 comparison](https://biosonic.io/compare/batdetect2).

| Same 3,159 recordings | BioSonic | BSG-BATS |
|---|---|---|
| Bat or noise, balanced accuracy (higher is better) | 99.2% | 77.0% |
| Noise reported as bat, when set to find 99% of bats (lower is better) | 1% | 95% |
| Correct species, all species in the test (higher is better) | 94% | 66% |
| Correct species, only species BSG-BATS has a class for | 94% | 87% |
| Correct species, worst to best detector type | 92–99% | 60–91% |
| Bats missed, best to worst detector type (lower is better) | 0–2% | 9–52% |
| Feeding buzzes and social calls | Detected (9% and 2% missed) | Not part of the model |
| Zero-crossing: bat or noise, balanced accuracy | 95.0% | 68.6% |
| Zero-crossing: correct species | 89% | 4% |

Balanced accuracy is the average of the share of bat recordings and the share of noise recordings each tool got right. Version 0.21 of BSG-BATS has no class for three of the Myotis species in the test, so the second species row leaves those 347 recordings out and scores the other 398. Zero-crossing means the same recordings converted to zero-crossing format, which holds far less information.

### Results by detector

Many surveys use one detector model for a whole season, so what matters is how a classifier does on your detector, not only on average. We split the test by detector type, with up to 100 single-species bat recordings and up to 100 noise recordings per type, all held out from training.

| Detector | BioSonic: bat or noise | BSG-BATS: bat or noise | BioSonic: correct species | BSG-BATS: correct species |
|---|---|---|---|---|
| Song Meter Mini Bat (Wildlife Acoustics) | 100% | 83.0% | 93% | 60% |
| Pettersson D500X | 100% | 73.5% | 93% | 64% |
| Song Meter SM4BAT (Wildlife Acoustics) | 99.5% | 84.9% | 99% | 81% |
| Titley Chorus | 96.6% | 89.4% | 99% | 87% |
| Elekon Batlogger | 98.1% | 80.0% | 92% | 73 of 100 |
| Titley Swift | – | – | 94% | 91% |
| **Range across detectors** | **96.6–100%** | **73.5–89.4%** | **92–99%** | **60–91%** |

Bat or noise = balanced accuracy, the average of the share of bat and noise recordings each tool got right. Correct species = share of bat recordings given the right species, including the species BSG-BATS has no class for; counting only the species it covers, BSG-BATS scored 68–94% per detector type. Higher is better in every column. From Titley Swift there were only bat recordings, so there is no bat-or-noise score for it.

BSG-BATS's weakest detector is Pettersson D500X, where it found 48% of bat recordings and named the right species for 64%, against 100% and 93% for BioSonic. BioSonic stays at 92% correct species or higher on every detector type.

**What this means in practice.** To find as many bats as BioSonic, BSG-BATS had to report 95% of noise recordings as bats. At the setting that balanced bats and noise best, it still missed 38% of bat recordings. So either a night of wind, rain and insects turns into bat activity that someone has to clear by hand, or a large share of real bat passes is never seen. Its results also swung with the detector: it missed 9% of bats on one detector type and 52% on another, while BioSonic missed at most 2% on any. Most surveys use one detector throughout, so a swing like that shows up as differences between sites that are not really there.

Species identification was its stronger side. On the species it covers it named the right one for 87% of bat recordings, ahead of BatDetect2 (80%) on the same recordings but behind BioSonic (94%). It was best of all three on Daubenton's bat, where it named every recording correctly (BioSonic 88%), partly because it leans towards that species and also called many other Myotis recordings Daubenton's. It matched BioSonic on serotine and soprano pipistrelle. It has no output for feeding buzzes or social calls, and it is built for full-spectrum WAV: on zero-crossing recordings it fell to 4% correct species.

To be fair to BSG-BATS: it was trained on 4,042 annotated segments, much of it from Finland, Spain and Italy [1], and these were Danish recordings from detectors and settings it may not have seen. We gave it the threshold that worked best on held-out recordings; even a threshold tuned on these very test recordings, which no user could know in advance, only raised its bat-or-noise score to 77.4%. The authors invite others to add annotations, and a version trained on more local data could do better. This is our own test; the independent benchmark of BioSonic is the one by Wilder Sensing below.

### Other evidence

**BSG-BATS.** The authors tested their model against Kaleidoscope Pro, SonoChiro, the BTO Acoustic Pipeline, BatDetect2 and Tadarida on 1,984 one-second clips, about 100 per species, kept apart from the training data. BSG-BATS had the highest AUC averaged over all 21 species and was best for most single species; BatDetect2 did well on the 14 species it could name, and the BTO Pipeline came third [1]. ecoObs bcAdmin4 was left out because many of its answers stayed at group level [1]. The authors are open about the limits: the test clips come from the same annotation pool as the training data, many Northern bat recordings came from one contributor, and for a few rare species several clips came from the same recording, which may make some scores look better than they would be in the field [1].

**BioSonic.** In a third-party benchmark by Wilder Sensing (5,325 predictions), BioSonic scored **98.7% F1** against 83.0% for Kaleidoscope Pro on the same data, with 13.4 times fewer false positives and 16.3 times fewer missed bats. The method and confusion matrices are on our [benchmarks page](https://biosonic.io/benchmarks) [6].

**Why the published figures do not compare.** AUC asks how well a model ranks the right species above the wrong ones across all confidence levels. F1 asks how many identifications were right and how many bats were missed at the setting you actually use. The paper's AUC and Wilder Sensing's F1 use different metrics, test sets and species, so they cannot be put side by side; the Danish test above is the one that runs both tools on the same recordings. To check on your own sites, take recordings an expert has already identified, run them through both, and compare false positives, missed bats and checking time.

## When to choose which

BSG-BATS is one of the first open-source classifiers for European bats and a good base for research. BioSonic is built for survey teams who need answers, checks and reports on a deadline.

### Choose BioSonic if you

- Want results without installing Python, managing models or writing scripts
- Survey in the UK or the Nordics and need species outside BSG-BATS's 21, such as Natterer's, whiskered or Brandt's bat
- Work as a team and want everyone in the same projects, with a record of who checked what
- Need species tables, charts, maps and a Word report straight from the project
- Process whole seasons of static detector data and do not want to size your own computers for it
- Want recordings stored in the EU, mainly in Sweden, and kept for 9 years

### Choose BSG-BATS if you

- Want an open model you can inspect, cite and retrain
- Are comfortable in Python and want to run everything on your own computer, offline
- Work on research or citizen science and want to contribute annotations back
- Survey in southern Europe, where its species list includes Miniopterus schreibersii, Tadarida teniotis, Hypsugo savii and Rhinolophus euryale
- Have no software budget and time to build your own review and reporting

### Using both

Both read standard WAV files, so you can run the same recordings through each and compare. A research group might use BSG-BATS for open, reproducible analyses and BioSonic for client surveys that several people need to check and sign off.

## The 21 species in BSG-BATS version 0.21

From Table 1 of the paper [1]: Barbastella barbastellus, Eptesicus nilssonii, Eptesicus serotinus, Hypsugo savii, Miniopterus schreibersii, Myotis alcathoe, Myotis crypticus, Myotis daubentonii, Nyctalus leisleri, Nyctalus noctula, Pipistrellus kuhlii, Pipistrellus nathusii, Pipistrellus pipistrellus, Pipistrellus pygmaeus, Plecotus auritus, Plecotus austriacus, Rhinolophus euryale, Rhinolophus ferrumequinum, Rhinolophus hipposideros, Tadarida teniotis and Vespertilio murinus.

Three Myotis species are included. Other Myotis common in surveys, such as Natterer's bat (M. nattereri), whiskered bat (M. mystacinus), Brandt's bat (M. brandtii), greater mouse-eared bat (M. myotis) and pond bat (M. dasycneme), are not in this version. Check the list against the species you expect on your sites.

## Frequently asked questions

### What is BSG-BATS?

BSG-BATS is an open-source annotation portal and convolutional neural network classifier for European bat calls, built in the LIFEPLAN programme coordinated at the University of Helsinki and published by Meramo, Somervuo and colleagues in Methods in Ecology and Evolution (volume 17, 2026) [1]. Version 0.21 identifies 21 European bat species. The model and code are free to download from Zenodo under a CC BY 4.0 licence and run in Python [2].

### Which is more accurate, BioSonic or BSG-BATS?

In our own test, BioSonic. In November 2026 we ran BSG-BATS version 0.21 and BioSonic on the same 3,159 manually verified recordings from Denmark, none of which was used to train either model. BioSonic named the correct species for 94% of bat recordings against 66% for BSG-BATS; counting only the species BSG-BATS has a class for, it was 94% against 87%. When both were set to find 99% of bats, BSG-BATS reported 95% of noise recordings as bats against 1% for BioSonic. BSG-BATS was trained mostly on recordings from Finland, Spain and Italy, so part of the gap may come from running it on Danish data. See the [full results](#accuracy-and-evidence). The BSG-BATS paper reports the highest average AUC of six tools on its own test set, which did not include BioSonic [1]. Separately, in an independent benchmark by Wilder Sensing, BioSonic scored 98.7% F1 against 83.0% for Kaleidoscope Pro [6].

### Which bat species does BSG-BATS identify?

Version 0.21 covers the 21 species [listed above](#the-21-species-in-bsg-bats-version-021) [1]. Natterer's bat, whiskered bat, Brandt's bat and greater mouse-eared bat are not among them.

### Is BSG-BATS free, and can consultancies use it?

Yes. The Zenodo record that holds the model, code and documentation is licensed CC BY 4.0, which allows commercial use as long as you credit the authors [2]. You download a 469 MB archive and run the classifier in Python on your own computer; the authors report under one second per 10-second WAV file on a single laptop CPU core [1]. You set up the installation, file handling, review and reporting yourself.

### Does BSG-BATS have a web app?

Not for classification, as of November 2026. The BSG portal is a web tool for annotating bat recordings that feed future versions of the model [3], and the authors write that a desktop application bringing their bird and bat models together is under construction [4]. Today you run the classifier from the Zenodo download [2]. BioSonic runs in the browser and does the analysis on its own servers.

### What is the best software for automatically identifying European bat species?

It depends on who is doing the work. For research teams who want an open model they can inspect and retrain, BSG-BATS and [Tadarida](https://biosonic.io/compare/tadarida) are strong choices. For consultancies that need a ready-to-use service with team review and reports, BioSonic identifies European bat species in the browser, scored 98.7% F1 in a third-party benchmark by Wilder Sensing [6], and in our own test on 3,159 verified Danish recordings named the correct species for 94% of bat recordings against 66% for BSG-BATS and 68% for [BatDetect2](https://biosonic.io/compare/batdetect2). [Kaleidoscope Pro](https://biosonic.io/compare/kaleidoscope) and [SonoChiro](https://biosonic.io/compare/sonochiro) are established desktop classifiers, the [BTO Acoustic Pipeline](https://biosonic.io/compare/bto-acoustic-pipeline) serves UK and European monitoring, and [ecoObs bcAdmin](https://biosonic.io/compare/ecoobs-batident) suits batcorder users. Run two or three on recordings you have already identified before you choose. Our [overview of bat identification software](https://biosonic.io/kaleidoscope-pro-alternatives) goes through each one.

## Other comparisons

- [BioSonic vs BatDetect2](https://biosonic.io/compare/batdetect2): the other open-source deep learning model, trained on UK bats, vs a ready-to-use service.
- [BioSonic vs Tadarida and Vigie-Chiro](https://biosonic.io/compare/tadarida): France's open-source toolbox and citizen-science programme vs a commercial service.
- [BioSonic vs SonoChiro](https://biosonic.io/compare/sonochiro): Biotope's Windows classifier vs browser-based identification and team review.
- [See all comparisons](https://biosonic.io/compare)

## References

1. Meramo, K., Somervuo, P., Rannisto, M., Lauha, P., López-Baucells, A., Gili, F., Roslin, T., Lilley, T. M. and Ovaskainen, O. (2026) BSG-BATS: An open-source data annotation portal and classifier for European bat vocalizations. Methods in Ecology and Evolution 17(3): 718–726. Open access. [doi:10.1111/2041-210x.70220](https://doi.org/10.1111/2041-210x.70220)
2. Somervuo, P. (2025) BSG-BAT model, code and documentation, version 0.21, CC BY 4.0. Zenodo: [doi:10.5281/zenodo.15495676](https://doi.org/10.5281/zenodo.15495676)
3. Bird Sounds Global, bat annotation portal and instructions: [bsg.laji.fi](https://bsg.laji.fi/bats/identification/instructions)
4. Meramo, K. (2025) Teaching models to listen to bats: the story behind BSG-BATS. Methods in Ecology and Evolution blog: [methodsblog.com](https://methodsblog.com/2025/12/11/teaching-models-to-listen-to-bats-the-story-behind-bsg-bats/)
5. Swedish University of Agricultural Sciences (2026) Researchers sharpen bat monitoring with an open-access tool: [slu.se](https://www.slu.se/en/news/2026/01/researchers-sharpen-bat-monitoring-with-an-open-access-tool/)
6. Wilder Sensing, third-party benchmark, 5,325 predictions: [biosonic.io/benchmarks](https://biosonic.io/benchmarks)

Competitor facts last checked: 2026-11-19.

## Ready to try it?

Upload a night you have already identified, with BSG-BATS or by hand, and compare the results. Nothing to install.

[Start free trial](https://bat.biosonic.se/signup) | [Book a demo](https://cal.com/biosonic/30min)

Questions? Email [josef.carlson@biosonic.se](mailto:josef.carlson@biosonic.se)
