---
canonical: https://biosonic.io/bat-classifier-accuracy
---

# Bat call classifier accuracy: what independent benchmarks show

Last updated: 2026-11-21. Written by the BioSonic team.

Which automatic bat classifiers have been independently benchmarked? Published tests on recordings of known species include Brabant et al. 2018 in coastal Western Europe (batIdent 77%, Kaleidoscope 71%, SonoChiro 65% and BatExplorer 31% correct species), Rydell et al. 2017 in Sweden (SonoChiro, Kaleidoscope and BatClassify, with large differences between species), Solick et al. 2024 in the northeastern USA (Kaleidoscope Pro, SonoBat and BCID), and the Wilder Sensing benchmark on about 5,300 UK recordings, where BioSonic scored 98.7% F1 and Kaleidoscope Pro 83.0%. Every study says an expert should still check automatic identifications.

We make BioSonic, one of the classifiers below. To keep this page useful, we list every published accuracy test we could open and check, say who ran each test, and link the paper. Numbers are quoted as the authors give them.

## Independent tests on recordings of known species

| Study | Region | Tools tested | Headline result |
|---|---|---|---|
| Brabant R et al. 2018 | Coastal Western Europe | batIdent 1.5, Kaleidoscope Pro 4.5.4, SonoChiro 4.0, BatExplorer 1.11.4 (default settings) | Correct species: batIdent 77%, Kaleidoscope 71%, SonoChiro 65%, BatExplorer 31% |
| Rydell J et al. 2017 | Sweden | SonoChiro, Kaleidoscope, BatClassify | All three separated rain from bats; group level ID 68 to 100%; species ranged from 83 to 99% (soprano pipistrelle, barbastelle) down to 16 to 26% (parti coloured bat) and 4 to 93% (Myotis) |
| Solick DI et al. 2024 | Northeastern USA | Kaleidoscope Pro 5.4.7, SonoBat 30.0, BCID 2.8b | Kaleidoscope Pro passed 0.80 on all five metrics for five species groups, SonoBat for four, BCID for none; sensitivity for two Myotis species only 0.01 to 0.52 |
| Wilder Sensing benchmark | United Kingdom | BioSonic, Kaleidoscope Pro and other classifiers | F1: BioSonic 98.7%, Kaleidoscope Pro 83.0%; BioSonic had 13.4 times fewer false positives and 16.3 times fewer missed bats |

### Coastal Western Europe: batIdent 1.5 and others

**Data:** 148 recordings of known species, 9 species, recorded with batcorders. **Tools:** batIdent 1.5, Kaleidoscope Pro 4.5.4, SonoChiro 4.0, BatExplorer 1.11.4 (default settings).

**Result:** Correct species: batIdent 77%, Kaleidoscope 71%, SonoChiro 65%, BatExplorer 31%. batIdent and Kaleidoscope were relatively reliable and BatExplorer poor, and the authors say a trained bat expert must cross check automatic identifications.

Brabant R, Laurent Y, Dolap U, Degraer S, Poerink BJ (2018). Comparing the results of four widely used automated bat identification software programs to identify nine bat species in coastal Western Europe. Belgian Journal of Zoology 148(2): 119 to 128. https://doi.org/10.26496/bjz.2018.21

### Sweden: SonoChiro and others

**Data:** Calls of 7 common species plus rain noise, Pettersson full spectrum detectors. **Tools:** SonoChiro, Kaleidoscope, BatClassify.

**Result:** All three separated rain from bats; group level ID 68 to 100%; species ranged from 83 to 99% (soprano pipistrelle, barbastelle) down to 16 to 26% (parti coloured bat) and 4 to 93% (Myotis). Each program had major but different weaknesses, so output should be validated and used only where confidence is high.

Rydell J, Nyman S, Eklof J, Jones G, Russo D (2017). Testing the performances of automated identification of bat echolocation calls: A request for prudence. Ecological Indicators 78: 416 to 420. https://doi.org/10.1016/j.ecolind.2017.03.023

### Northeastern USA: Kaleidoscope Pro 5.4.7 and others

**Data:** 1,500 recordings: 1,120 known species reference calls of 9 species plus 380 noise files. **Tools:** Kaleidoscope Pro 5.4.7, SonoBat 30.0, BCID 2.8b.

**Result:** Kaleidoscope Pro passed 0.80 on all five metrics for five species groups, SonoBat for four, BCID for none; sensitivity for two Myotis species only 0.01 to 0.52. The programs rarely give false positives but often miss true calls, especially Myotis, so an expert should verify files before regulatory decisions.

Solick DI, Hopp BH, Chenger J, Newman CM (2024). Automated echolocation classifiers vary in accuracy for northeastern U.S. bat species. PLoS ONE 19(6): e0300664. https://doi.org/10.1371/journal.pone.0300664

### United Kingdom: BioSonic and others

**Data:** About 5,300 expert verified UK recordings (5,325 predictions). **Tools:** BioSonic, Kaleidoscope Pro and other classifiers.

**Result:** F1: BioSonic 98.7%, Kaleidoscope Pro 83.0%; BioSonic had 13.4 times fewer false positives and 16.3 times fewer missed bats. Run by Wilder Sensing, a UK company, not by BioSonic; the only recent head to head test of a deep learning classifier against Kaleidoscope Pro on UK species we know of.

Wilder Sensing. Third party benchmark of bat classifiers on expert verified UK recordings (5,325 predictions). Results and method summarised by BioSonic at biosonic.io/benchmarks. https://biosonic.io/benchmarks

## Agreement studies and tests run by developers

These are useful, but they either compare programs with each other instead of with known species, or were run by the people who built the tool.

**Lemen et al. 2015.** On 29,782 field files from Nebraska, Bat Call ID, EchoClass, Kaleidoscope Pro and SonoBat agreed on the species only about 40% of the time. This measures agreement between programs, not accuracy against known species. Lemen C, Freeman PW, White JA, Andersen BR (2015). The problem of low agreement among automated identification programs for acoustical surveys of bats. Western North American Naturalist 75(2): 218 to 225. https://digitalcommons.unl.edu/natrespapers/495

**Mac Aodha et al. 2018 (Bat Detective).** A deep learning detector beat SonoBat, SCAN'R and Kaleidoscope at finding search phase calls in noisy UK and Eastern European transect recordings (average precision 0.87 to 0.90 against 0.20 to 0.55). This tests detection, not species identification, and was run by the detector's developers. Mac Aodha O, Gibb R, Barlow KE, Browning E, Firman M, Freeman R, et al. (2018). Bat detective: Deep learning tools for bat acoustic signal detection. PLoS Computational Biology 14(3): e1005995. https://doi.org/10.1371/journal.pcbi.1005995

**Mac Aodha et al. 2022 (BatDetect2).** On 17 UK species (2,809 files) BatDetect2 reached a mean average precision of 0.88, against 0.71 for a call parameter baseline. No commercial tools were compared, and the test was run by the developers. Mac Aodha O, Martinez Balvanera S, Damstra E, et al. (2022). Towards a General Approach for Bat Echolocation Detection and Classification. bioRxiv. https://doi.org/10.1101/2022.12.14.520490

**Tabak et al. 2021.** On recordings from 20 US states the authors' own model was correct for about 92% of calls, against 65% for Kaleidoscope Pro and 61% for BCID on the same calls. Run by the model's developers. Tabak MA, Murray KL, Lombardi JA, Bay KJ (2021). Automated classification of bat echolocation call recordings with artificial intelligence. bioRxiv. https://doi.org/10.1101/2021.06.23.449619

## Our own test

In our own test on 3,159 verified Danish recordings, BioSonic named the correct species for 94% of bat recordings, the BTO Acoustic Pipeline for 70.1% and BatDetect2 for 68%. We made this test ourselves, so treat it as a vendor result, not an independent one. Method and full results: [BioSonic benchmarks](https://biosonic.io/benchmarks).

## What the evidence does not cover

- We found no published independent accuracy test of the BTO Acoustic Pipeline, of SonoBat on European species, or of BSG-BATS, iBatsID or Anabat Insight.
- Most known species tests in Europe are from 2017 and 2018 and used software versions that have since been updated, so their numbers describe those versions, not today's.
- Studies report different metrics (percent correct, sensitivity, positive predictive value, F1), so numbers from different studies cannot be compared directly.

## Frequently asked questions

### Which automatic bat classifier is the most accurate?

No single study has tested every classifier on the same recordings. In the Wilder Sensing benchmark on about 5,300 UK recordings, BioSonic scored 98.7% F1 against 83.0% for Kaleidoscope Pro. In Brabant et al. 2018, on 148 known species recordings from coastal Western Europe, batIdent (77%) and Kaleidoscope (71%) beat SonoChiro (65%) and BatExplorer (31%).

### How accurate is Kaleidoscope Pro for European bats?

Brabant et al. 2018 found 71% correct species identification for Kaleidoscope 4.5.4 on nine species in coastal Western Europe. The Wilder Sensing benchmark measured 83.0% F1 on UK recordings. Both studies expect an expert to check the results.

### Do I still need to check automatic bat identifications by hand?

Yes. Every published evaluation recommends expert checking of automatic identifications, especially for Myotis, Plecotus and Nyctalus and Eptesicus calls, and before results support a planning or licensing decision.

### How should I test a bat classifier myself?

Run it on a set of your own recordings whose species you have confirmed, from your own detectors and habitats, and count correct identifications, false positives (noise or the wrong species) and missed bats per species. F1 combines false positives and misses into one number.

## How to cite this page

BioSonic (2026). Bat call classifier accuracy: what independent benchmarks show. https://biosonic.io/bat-classifier-accuracy (accessed DATE).

Know of a published test we missed? Tell us at support@biosonic.se and we will add it.

## Related Pages

- [BioSonic benchmarks](https://biosonic.io/benchmarks)
- [Bat call analysis software compared](https://biosonic.io/bat-acoustic-analysis-software)
- [Kaleidoscope Pro alternatives](https://biosonic.io/kaleidoscope-pro-alternatives)
- [Automatic bat species identification](https://biosonic.io/automatic-bat-species-identification)
