BirdNET, Perch and NatureLM-audio: choose by the task.
Detecting a species, comparing sound patterns and asking a question about a clip are different jobs. Choose an output that fits your research question, then validate it on recordings that represent your setting.
CrowLingo Atlas · Documentation review: · Corrections
How this comparison was made
This is a documentation comparison of the developers’ published projects, not a head-to-head performance benchmark. BirdNET has multiple model versions; a figure for v2.4 must not be treated as a specification for every release. Perch 2.0 and NatureLM-audio likewise require an explicit checkpoint and setup before comparing results.
The comparison does not establish a best model for corvid call meanings, hardware latency, accuracy on your recordings or a universal inference price. Code, weights and source recordings can have different terms.
Compare the output and its limits
| Criterion | BirdNET | Perch | NatureLM-audio |
|---|---|---|---|
| Task to investigate | Species detection and classification | Audio embeddings and downstream classification | Questions about audio in natural language |
| What to inspect | Model version, species coverage and detection thresholds | Embedding model, downstream task and evaluation data | Model card, question format and grounding of the answer |
| What the output cannot establish alone | A bird's intent or message | The meaning of a nearby point in an embedding space | That a fluent answer is correct or a bird's message is translated |
| License check | Check the specific code and weight licenses separately | Check the repository and selected model's terms | Check the repository and selected checkpoint's terms |
Before using an output as evidence
- Identify the exact model, checkpoint, preprocessing and output you used.
- Choose a labeled evaluation set that represents the species, microphones, noise and geography you care about.
- Separate a recordist’s observation, a classifier label, an embedding pattern and your behavioral interpretation.
- Report errors and uncertainty. Similar sounds, a confidence score or fluent language do not establish a call’s meaning.
Inspect the original documentation
Compare actual reference recordings and their source notes →