Skip to content
CrowLingo
Back to the pipeline

Stage 03

Detect & preprocess — preserve the evidence.

Review candidate detections, retain the original recording, and prepare derivatives for a specific model and task. There is no validated universal cleanup recipe for this crow corpus.

Stage 3 — Preprocess

Preserve the original recording and prepare separately identified derivatives for a selected model version and task. Follow its documented sample rate, channel, window and amplitude requirements, and record every transformation. Per-clip peak normalization changes relative recording levels; filtering or denoising does not guarantee better results. Compare transformed recordings with originals and evaluate the intended task. This is a proposed workflow, not a validated preprocessing pipeline for CrowLingo’s atlas.

Audio narration is unavailable while its source claims are corrected.

First, review candidate detections

Stage 2 asks where a target sound may occur. A model score or generated answer is a candidate to check against the recording, not a confirmed crow call or a calibrated probability by default. Keep clip boundaries and a link to the unmodified source.

Use independently checked examples to evaluate false detections and missed calls for your task. Document the selection process and conditions represented in the sample. CrowLingo has not measured a crow-detection error rate or comparative model speed for this atlas.

Prepare inputs for the selected version

Input requirements differ. The BirdNET model documentation describes V2.4 with 48 kHz input handling, automatic resampling and normalization before spectrogram conversion. Its V3.0 preview specifies 32 kHz and three-second segments. These are versioned requirements, not a general rule for every encoder.

The NatureLM-audio README shows configurable inference windows and prompts. Follow the selected interface rather than assuming that every model expects the same sample rate, window, channels or output.

  1. Preserve the original file, sample rate, channels and recording notes.
  2. Choose a model version and task, then check its documented input handling.
  3. Create separately identified derivatives only as needed; log trimming, resampling, channel conversion and amplitude changes.
  4. Compare results against the originals and held-out annotations before adopting a transformation.
  5. Keep the source-to-clip mapping, parameters, software versions and failed or excluded examples.

These are proposed research steps. They do not establish a validated preprocessing pipeline for CrowLingo's atlas. Continue to model outputs and extraction provenance before generating representations.

Normalization changes the amplitude evidence

Independently peak-normalizing every clip does not preserve relative recording levels. For example, two otherwise identical waveforms with peaks of 0.5 and 1.0 become identical if each is scaled to a peak of 0.707. This is an illustrative arithmetic example, not a measurement of crow behavior.

Model-required amplitude handling and analysis of original recording levels are separate decisions. Preserve the unnormalized recording and transformation history. A normalized waveform alone does not establish urgency, proximity or behavioral meaning.

Evaluate filtering and denoising

This guide does not prescribe a bandpass, denoising strength or normalization target for all crow recordings. BirdNET's documented input band extends from 0 to 15 kHz; cutting it to 200 Hz–8 kHz removes part of that band, without establishing that the model will perform better.

If a transformation is proposed, retain the original, compare recordings and spectrograms before and after, and test its effect on the intended task. Document uncertainty. A cleaner-sounding recording is not, by itself, evidence of better classification or more meaningful clusters.