This dataset contains expert annotations and BirdNET-generated species detections from one-minute audio recordings collected at a single urban site in Laim, Munich, Germany, between May and October 2021. Recordings were made in timed intervals during the early morning hours to focus on peak bird vocal activity, resulting in 15.5 hours of audio from 60 recording days. Each recording was manually reviewed by two experienced ornithologists for bird vocalisations, and the same audio was processed using BirdNET-Analyzer (v2.2, April 2023) across a range of parameter combinations, including variations in sensitivity, overlap, week of year, and confidence threshold. The dataset includes raw BirdNET outputs for all parameter combinations as well as expert-verified species lists for each clip.

Acoustic recordings

Between May and October 2021, we deployed a single Frontier Labs BAR recorder on the roof of a residential building in Munich, Laim, Germany. Recordings were made in week-long sessions, with at least a one-week break between each session. During each session, the recorder was programmed to capture one-minute clips every ten minutes, starting two hours before sunrise and continuing until three hours after sunrise. All recordings were made at 48 kHz sampling rate, 16-bit depth, and with a gain setting of 40 dB.

Species Identification

Two experienced ornithologists reviewed each one-minute clip to identify bird vocalisations, either by listening or by inspecting the spectrograms. This was done using Wildlife Acoustics Kaleidoscope Pro (version 5.6.8), with default settings (FFT size: 256, Window size: 128, Max cache: 256 MB). Each recording was then annotated with the species present. In addition, we processed the same audio files using BirdNET-Analyzer (April 2023 release, model v2.2), generating a list of species detected by the algorithm for each clip.

BirdNET parameter settings

We generated BirdNET detections using different parameter combinations. First, we included or excluded the week of the year, which allows BirdNET to filter expected species based on date and location using eBird data. Second, we varied the sensitivity parameter, which affects how likely BirdNET is to detect quieter or background vocalisations. We used three levels: 0.5 (low), 1.0 (default), and 1.5 (high). Third, because BirdNET processes audio in three-second segments, we used four different overlap settings (0, 1, 2, and 2.9 seconds) to control how much of the previous segment is reused. Finally, BirdNET assigns a confidence score to each detection. We set the minimum confidence threshold to the default (0.1) to retain all detections.

The code for producing the analyses found in the associated manuscript are on GitHub:

https://github.com/AFairbairn/BirdNET_study_PLOS_one