Chapter 05

Birds

Resident and migratory species of the humid Andean slopes.

By Camila Orozco-Bernal, Yoshi Orozco, María Josefita Alcibíades and Andrés Montoya-López

This chapter was assembled by listening rather than looking. A recorder left running in the forest hears the birds a person walking through it never will — the ones that call before dawn, the ones that fall silent at the sound of footsteps, the ones that live too high in the canopy to see.

What follows is a reference library: one recording for each species identified so far, with the sonogram that shows the shape of its call.

How these were recorded

Recordings come from a Song Meter Micro 2 left running continuously at fourteen sites across La Honda, from 1,514 m at Angurria to 2,593 m at El Tambo — La Linda. Between December 2025 and April 2026 the unit collected roughly 1,170 hours of audio in one-hour files, mono, 48 kHz.

Identification runs in two passes. BirdNET makes the first proposal, scanning the recordings in three-second windows and returning a species and a confidence score for each detection. Every proposal is then corroborated by ear by Camila Orozco-Bernal, who confirms, corrects or rejects it. The model proposes; a person decides. Where the two disagree, the human identification is the one recorded here.

Each species below is represented by a single seven-second clip, cut from two seconds before the detection to five seconds after. The two seconds of lead-in are not padding — they are a sample of the same forest without the bird in it, which is what lets each sonogram be exposed for the call rather than for the stream or the cicadas.

How reliable is this?

Rather than assert that the method works, we tested it. All 177 verified reference clips were put back through BirdNET a second time, to see whether the model would independently reach the same species. It agreed on 173 of 177 — 97.7%.

More usefully, the disagreements are not scattered. They fall entirely in the low-confidence tail:

BirdNET confidence	recordings	agreement
0.90 and above	77	100%
0.70 – 0.90	34	100%
0.50 – 0.70	23	100%
0.30 – 0.50	29	96.6%
below 0.30	14	78.6%

Above a confidence of 0.50, across 134 recordings, the model and the second pass have never disagreed. Every disagreement lies below it. That is where listening still earns its keep, and it is why the confidence score is printed under each recording below — not as a grade, but as a marker of how much of an identification rests on a person rather than a model.

Four identifications remain genuinely contested and are marked pending review. A few others differ from BirdNET only because the model's checklist is older than current taxonomy: it cannot name Driophlox cristata, Strix virgata, Cyphorhinus dichrous or Pogonotriccus ophthalmicus at all, only the names those birds used to carry.

What this chapter is not

A recorder documents presence, not abundance. A species heard once and a species heard nightly appear here identically. Nor is the list closed — ten of the fourteen deployments have yet to be analysed, and the species below come from four sites over eleven days of listening.