AI bioacoustic monitoring
BirdNET (6,000+ species claimed), Merlin Sound ID (2021) and Perch 2.0 (2025) sort field recordings automatically; in one test humans still detected more birds.
Open in the interactive tree →Cheap recorders produce more audio than people can review, so classifiers screen it in short windows and return species lists with confidence scores. This is where AI already replaces manual sorting of animal sound, although accuracy varies by species and settings.
As of October 2026
BirdNET, from the Cornell Lab and Chemnitz University of Technology, says it recognises over 6,000 species. In 144 paired point counts, humans made 382 detections against 222 for Merlin Sound ID, with precision 92% against 86%, and Merlin returned different species lists on two devices at the same time in 57% of counts (study summarised October 2025). A May 2026 summary of a global BirdNET assessment says performance varies among species and settings and cites the WABAD benchmark of over 90,000 annotated vocalisations from more than 1,100 species. Perch 2.0 (Google DeepMind, August 2025) reports leading BirdSet and BEANS scores and transfers to marine tasks with almost no marine training data (authors' evaluation).
Open steps
- Sort years of recordings High AI leverageRun detectors over recorder archives to get species lists by site and date; BirdNET splits audio into 3-second windows and can be run in batches.
- Cut false alarms Medium AI leverageMeasure and reduce false detections per species and device: in a 144-count test Merlin detected 222 birds to humans' 382 and listed 12 false-positive species.
- Extend to non-bird sound Medium AI leverageExtend detectors beyond birds: Perch 2.0 expanded from birds to a multi-taxa training set and transferred to marine mammal tasks with almost no marine training data.
Where AI could help
High AI leverage. Detectors already do the sorting people cannot: thousands of recordings become species lists, though humans still catch more birds and must check the output.
- Turn recorder archives into species lists by site and date
- Filter detections by confidence score and local species ranges
- Reuse embeddings for new species with few labels
- Run on phones and low-power field devices
Shown so far
- A May 2026 BOU summary says free tools such as Merlin and BirdNET translate thousands of audio files into bird predictions, with performance varying among species and settings. source
- In 144 paired point counts, humans made 382 detections and Merlin Sound ID 222; precision was 92% for humans and 86% for Merlin (American Ornithological Society summary, October 2025). source
- Perch 2.0 (Google DeepMind) reports state-of-the-art scores on BirdSet and BEANS and outperforms specialised marine models on marine transfer tasks despite almost no marine training data (authors' evaluation). source
Prerequisites
- Songbirds learn their songs1954-1970
- Deep Learning2012