Skip to content
CrowLingo
Back to the pipeline

Stage 04

Embed — choose a representation you can evaluate.

An embedding depends on the recording, preprocessing and model version. First check what the model actually returns; a species score or a language answer is not automatically an audio embedding.

Stage 4 — Embed

An audio embedding depends on a specific model version and preprocessing recipe. Check the documented output before extraction: species scores, embeddings and generated language answers are different products. Record the checkpoint, input handling, vector interface and clip mapping, then evaluate the representation for your task. NatureLM-audio documents prompted text inference; this guide does not establish a fixed-size embedding API for it. CrowLingo has not extracted embeddings for the current atlas corpus, whose coordinates and groups remain illustrative.

Audio narration is unavailable while its source claims are corrected.

Check the output before choosing the model

This is a proposed research workflow. CrowLingo has not extracted model embeddings for its atlas corpus, and has not benchmarked these models against one another on crow calls.

  • BirdNET. Its developer documentation distinguishes species predictions from embeddings. V2.4 lists 1,024-dimensional embeddings; the V3.0 preview lists 1,280. Input requirements and output dimensions belong to a specific version, not to every audio model. Read the model specifications.
  • Perch. The developer repository describes models for bioacoustic classification and embeddings. Check the selected checkpoint, preprocessing and extraction interface, including the separate Perch 2.0 documentation. These capabilities do not establish superiority for within-crow call grouping. Read the Perch documentation.
  • NatureLM-audio. The documented inference pipeline takes audio and a prompt and generates text. Its README shows NatureLM.models.NatureLM and NatureLM.infer.Pipeline; it does not establish the fixed-size embedding API previously shown on this page. Follow the developer instructions for the supported interface. The NatureLM-audio paper reports specific audio-language evaluations, including BEANS-Zero; those results do not validate crow translation or this atlas.

Record a reproducible extraction recipe

Before generating vectors, define the task and a held-out evaluation. Then record the following alongside each output. This checklist is a proposed data contract, not executable model code or a completed CrowLingo dataset.

  1. Source recording identifier, clip boundaries and permission record.
  2. Model name, checkpoint/version, software version and a checksum of the actual weights used.
  3. Sample rate, channel handling, windowing, padding, normalization and pooling choices required by that model.
  4. Extraction layer or interface, vector dimension, normalization and intended distance metric.
  5. A stable clip-to-vector mapping, output format, run date and extraction failures.

The next stage must load the same recorded vectors and clip mapping. Do not silently switch between incompatible file formats, model versions or preprocessing recipes. Keep separate runs distinguishable so that a later comparison can be reproduced.

Distance is a hypothesis to test

Outputs from independently trained models need not share a coordinate system, even when their dimensions match. Within a single model, proximity under a chosen metric may help formulate questions; it does not by itself prove perceptual similarity, individual identity or behavioral meaning. Evaluate those claims with appropriate recordings, labels and observations, including held-out examples.

For complementary measured-feature approaches and evaluation requirements, see Measured and learned features.