Can a bird-sound model hear threatened fish spawning? It caught only about half the events
Researchers reused embeddings from seven pretrained sound models to detect twaite shad splashes in UK and Belgian river recordings. BirdNET transferred best, but the leading classifier reached F1=0.69 and produced more false alarms in a preliminary continuous-recording check.
The short version
- What happened
- Researchers manually annotated 28 hours, 40 minutes and 38 seconds of recordings from UK and Belgian rivers, identifying 351 twaite shad splashing events collected between 2022 and 2024.
- Who it affects
- For river managers and field ecologists, a reliable detector could reduce the time spent listening to long recordings and make more sites practical to monitor. At the current error level, human review remains necessary.
- What remains unproven
- 351 positive events from recordings collected in three rivers during 2022–2024
Keep in mind
351 positive events from recordings collected in three rivers during 2022–2024
Selected thresholds maximised F1 on the available test data
Only a qualitative five-hour check on continuous recordings
Elevated false positives in more realistic rare-event audio
No new-river external validation or population-level outcome
No measured staff-time saving or conservation decision
Scholarly book chapter; journal peer-review status is not described
Source analysis · 1 linked source. Primary analysed source date: 28 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
A model trained to hear birds supplied the most useful sound features
Twaite shad males make distinctive surface splashes while spawning, giving conservation teams a non-invasive way to locate activity—if they can separate rare splashes from rain, boats, birds and moving water. Researchers recorded known spawning rivers in the UK and Belgium and manually labelled 351 events across 28 hours, 40 minutes and 38 seconds of audio.
They tested embeddings—compact numerical sound features—from seven pretrained models and trained simple classifiers on top. BirdNET, originally built for bird calls, supplied the strongest transferable representation. The best training-data configuration reached an F1 score of 0.69 on a held-out test set with ten noise clips for every splash clip. F1 balances missed events and false alarms; it does not say the detector found 69% of all spawning activity in a river.[1]
The rare-event problem remains visible in the errors
The selected classifier recorded 41 true positives and eight false positives at its chosen threshold. Other configurations caught 60 or 64 events but produced 79 or 164 false positives. Across the embedding comparison, even BirdNET correctly identified only a little over half of positive sound events. Noise from water and distant or masked splashes drove both kinds of error.
The team also ran the classifiers over five hours of continuous field audio. The relative ranking looked similar, but false positives increased; the check was qualitative rather than a separate labelled validation set. The chapter therefore supports a promising annotation aid, not autonomous population monitoring. Tests on new rivers, seasons and equipment—with a pre-set threshold and a complete human-reviewed denominator—would show whether the tool can reduce staff time without hiding real spawning activity.[1]
The distinction that matters
Three evidence layers behind the detector
Each layer answers a different question; the final operational claim is still unresolved.
Annotated audio
351 splashes were labelled across 28 h 40 min of recordings
A small, real river dataset for training and testing
Held-out clips
Seven embedding models and seven dataset variants were compared
Best reported F1 was 0.69
Continuous field audio
Five unedited hours were checked qualitatively
More false alarms; no deployment estimate yet