Skip to content
CrowLingo

Methods · Sub-page

NatureLM-audio.

An audio-language model for bioacoustics: what the published study evaluates, and what its results do not establish for crow calls.

Methods · NatureLM-audio

NatureLM-audio is an audio-language model for bioacoustics described by Earth Species Project. This methods page introduces the research; CrowLingo does not run the model on its atlas corpus or provide model-backed similarity search.

Audio narration is unavailable while its source claims are corrected.

What it is

is an audio-language model from Earth Species Project. Robinson and colleagues describe it in NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics. It takes audio with an instruction and produces text.

How the paper's model is trained

The reported architecture connects a BEATs audio encoder to Llama-3.1-8B-Instruct through a window-level Q-Former. It learns from paired audio-text examples. The language model's base weights remain frozen while LoRA adapters are updated; the audio encoder and Q-Former remain trainable.

Training starts with focal-species classification, then introduces broader tasks including detection, captioning, life-stage and call-type prediction, with speech and music data. This is the configuration described in the paper, not a guarantee that every released checkpoint behaves identically.

What BEANS-Zero evaluates

  • Species generalization. Classification tasks include species and taxa held out from training.
  • Bird life stage and call type. The described call-type task distinguishes song from call; it is not a crow territorial-intent test.
  • Captioning. The benchmark evaluates generated descriptions of bioacoustic audio.
  • Multiple individuals. One task uses generated mixtures of one to four zebra finches. It does not establish accurate counting of crows in field recordings.

Results depend on the task, data split, model configuration and evaluation measure. Generalization in these experiments does not establish reliable crow-to-raven transfer, territorial-call interpretation or a translation of animal meaning.

Using an answer responsibly

A caption or classification is a model output to check against suitable recordings and independently established labels. Fluent wording is not evidence of a bird's intention. A new research question still needs a validation design; this page makes no promise that prompts replace weeks of model development.

CrowLingo does not run NatureLM-audio on its Atlas corpus and does not offer model-backed similarity search. The Atlas remains a listening and learning interface with illustrative coordinates. Its recordings are not outputs of the model described here.

Primary sources

Sources checked November 11, 2026. This is a reading of the paper and documentation, not an independent reproduction or a CrowLingo model benchmark.

Frequently asked

What people ask about this.

What is NatureLM-audio?

NatureLM-audio is Earth Species Project's audio-language model for bioacoustics. The published configuration connects a BEATs audio encoder through a Q-Former to Llama-3.1-8B-Instruct, trains on paired audio-text examples and produces text from audio and an instruction. Performance depends on the model configuration and task.

What can NatureLM-audio actually do?

The paper evaluates species generalization, bird life-stage and song-versus-call classification, captioning, and other tasks. Its multiple-individual task uses generated mixtures of one to four zebra finches; that does not establish reliable crow counting or territorial-intent interpretation. CrowLingo does not run this model on its Atlas corpus.