Skip to content
CrowLingo
← Methods

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.

· 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

Documentation-level comparison; no common-dataset accuracy result is asserted.
CriterionBirdNETPerchNatureLM-audio
Task to investigateSpecies detection and classificationAudio embeddings and downstream classificationQuestions about audio in natural language
What to inspectModel version, species coverage and detection thresholdsEmbedding model, downstream task and evaluation dataModel card, question format and grounding of the answer
What the output cannot establish aloneA bird's intent or messageThe meaning of a nearby point in an embedding spaceThat a fluent answer is correct or a bird's message is translated
License checkCheck the specific code and weight licenses separatelyCheck the repository and selected model's termsCheck the repository and selected checkpoint's terms

Before using an output as evidence

  1. Identify the exact model, checkpoint, preprocessing and output you used.
  2. Choose a labeled evaluation set that represents the species, microphones, noise and geography you care about.
  3. Separate a recordist’s observation, a classifier label, an embedding pattern and your behavioral interpretation.
  4. 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 →

Explore the research library

Frequently asked

What people ask about this.

What is the difference between BirdNET and Perch 2.0?

BirdNET supports species detection and classification. Perch supports audio embeddings and downstream classification. Compare the exact model versions, inputs and output requirements before selecting either; the distinction does not establish comparative accuracy on your recordings.

Which bioacoustic model is best for crows?

This documentation comparison does not establish a best model for crow recordings or their meaning. A defensible performance comparison needs a labeled evaluation set, exact model versions and a stated task. A nearby embedding point or fluent answer is not evidence of a call translation.

Is BirdNET still relevant in 2026?

BirdNET remains an option to evaluate for species detection. Check its current model documentation and validate performance in your recording conditions. Do not assume that an older version specification describes every release or that another model is always more accurate or more expensive.