CEB: A collection of strongly-labeled bird soundscapes and Xeno-Canto recordings from Central Europe, with bounding boxes and vocalization types

Overview

This collection Central European Birds (CEB) contains 92,857 annotations across 20,826 audio files, covering 256 eBird species codes (266 scientific names) and 21 distinct vocalization types. It is organized into three subsets that differ in how the audio was recorded and how thoroughly it was annotated:

• train_xenocanto:  focal recordings from Xeno-Canto for the project's target species (covering 57 of the 59 evaluation species), re-annotated with bounding boxes and vocalization types
• train_soundscape:  point-based annotated PAM soundscape recordings, partially with vocalization type annotations
• test_soundscape:  strongly-labeled PAM soundscapes for evaluation with vocalization type annotations

All audio is provided as FLAC. Published file paths and coordinates are anonymized. This collection supports scientific research and benchmark development in avian bioacoustics. It enables the development and evaluation of models for species detection, vocalization-type classification, and sound-event localization across focal recordings and passive acoustic monitoring soundscapes

Subsets

1. train_xenocanto
   • 2,210 audio files, 15,495 annotations, 186 species
   • Focal recordings originating from Xeno-Canto.
   • These were originally only weakly labeled. For this dataset they were downloaded and re-annotated: bounding boxes were drawn, labels were corrected, and vocalization types were assigned.
   • Background species are also newly annotated.
2. train_soundscape
   • 18,469 audio files, 62,298 annotations, 216 species
   • Weakly-labeled PAM soundscape recordings.
   • Each annotation marks an event by its center point with a fixed ±2.5 s window around it (a 5 s span).
   • Vocalization types are present for some annotations but not all (34k annotations have no vocalization type assigned).
   • Some annotations are multi-label: when several species overlap, the species and label columns (scientific_name, common_name, ebird_code_multilabel, ebird#voc_type) list the values separated by "|" (5,070 annotations).
3. test_soundscape
   • 147 audio files, 15,064 annotations, 59 species
   • Strongly-labeled soundscape recordings.
   • Every sound event is annotated with a time-frequency bounding box (start/end time, low/high frequency) and a vocalization type.
   • This subset is a meticulously annotated subset of the soundscape recordings and is intended as the evaluation set.

Geographic Coverage

Recordings are predominantly from Germany, with a small subset from Greece. Coordinates in the metadata are rounded to whole degrees to reduce spatial precision, so the locations in the metadata are approximate regions rather than exact sites. (Focal train_xenocanto recordings do not carry coordinates.)

Vocalization Types

Vocalization types are encoded in the ebird#voc_type column. The dataset uses 21 distinct types; the companion file voc_types.json lists all values with their per-subset counts. Not every annotation carries a vocalization type — in the weakly-labeled train_soundscape subset many have none, and these are counted as (missing) in voc_types.json. The explicit label something is equivalent in meaning to (missing). The definitions below cover the 21 types together with (missing); labels are given exactly as they appear in the column.

• (missing): No vocalization type was specified or annotated. This may occur, for example, when an artificial or non-biological sound cannot be classified more precisely, or when no more specific label was assigned during data annotation.
• something: Equivalent to (missing).
• song: Longer and more complex sequences of vocalizations used primarily for territorial defense or mate attraction. These are likely to be produced predominantly by males, although female song is surely also included.
• secondary song: A vocalization that serves functions similar to those of the main song but differs from it in structure or delivery.
• sub song: Quiet, often rapid and not yet fully developed song, frequently produced by young birds or outside the main singing period.
• drumming: A drumming sound produced by rapidly striking the bill against a substrate, usually wood. Typical of woodpeckers.
• call: A general call that has not been assigned to a more specific call category.
• contact call: A call used by a species to maintain contact with conspecifics, typically while perched.
• flight call: A call typically given in flight, e.g., during migration. Flight calls may occasionally also be produced while the bird is perched.
• begging call: A begging vocalization, usually produced by young birds soliciting food from their parents, but occasionally also by, e.g., females soliciting food from males.
• alarm call: A warning call produced in response to a perceived threat, typically serving to alert other animals to danger.
• ground alarm call: A specific alarm call indicating a threat originating from the ground, for example a terrestrial mammalian predator.
• air alarm call: A specific alarm call indicating an aerial threat, for example a bird of prey.
• excitement call: A call produced during a state of heightened arousal or excitement, in some cases preceding or grading into an alarm call.
• interaction call: A collective category for a range of very quiet and variable vocalizations produced during direct social interactions between two or more birds.
• feather/air (with wingbeat): A non-vocal sound produced by the feathers and/or airflow during wingbeats.
• feather/air (with wingbeat) (display): A sound produced by the wings or feathers during courtship or other display behavior.
• pecking: A tapping or pecking sound produced by the bill while striking or working on a surface.
• attack call: A call associated with aggressive behavior or an attack, for example directed towards a bird of prey.
• bill snapping (series): A series of audible bill-snapping or bill-clattering sounds.
• call in hand: A call produced while a bird is being held by a person, for example during capture, ringing, or examination.
• whistle: A whistling vocalization for which no more specific functional category is provided in the dataset.

Files in this Collection

• *.csv contain the metadata for the subset.
• *.tar.gz contain the audio (.flac) files for the subsets.
• voc_types.json contains the vocalization-type values with their per-subset counts.
• scientific_name_counts.json contains the species (scientific_name) values with their per-subset counts.
• schema.json contains the data dictionary for the CSV columns.
  
  

Within each audio archive, files use anonymized paths of the form <subset>_project_<NNN>/audio_<NNNNNN>.flac
e.g. train_soundscape_project_001/audio_000001.flac

Licenses

This record bundles material under more than one license, choose terms by component.

• Soundscape audio (train_soundscape and test_soundscape) and ALL dataset metadata and annotations are released under Creative Commons Attribution 4.0 International (CC-BY-4.0). This is the dataset's own contribution and the record-level license.
• Focal audio (train_xenocanto) was collected from Xeno-Canto. Each recording is redistributed under its own original license, given per row in the license column, with attribution to the recordist (xc_recordist) and the source (xc_url). It is not relicensed. Several recordings are non-commercial (NC) and/or share-alike (SA), so use of the focal audio is subject to those per-recording terms. License breakdown for the 15,495 train_xenocanto annotations:
  • CC-BY-NC-SA 4.0: 14,299
  • CC-BY-NC-SA 3.0: 870
  • CC-BY 4.0: 131
  • CC-BY-SA 4.0: 73
  • CC0 1.0: 58
  • CC-BY-NC 4.0: 10
       

Acknowledgements

This work was carried out as part of the DeepBirdDetect project (Fkz. 67KI31040E) funded by the German Federal Ministry for the Environment, Nature Conservation, Nuclear Safety and Consumer Protection (BMUV). All annotation and validation was done by Ralph Martin.

Disclaimer

We carefully selected and composed this dataset's content. If you believe that any of this content violates licensing agreements or infringes on intellectual property rights, please contact us immediately. In such a case, we will promptly investigate the issue and remove the implicated data records from our dataset if necessary. Users are responsible for ensuring that their use of the dataset complies with all licenses, applicable laws, regulations, and ethical guidelines. We make no representations or warranties of any kind and accept no responsibility in the case of violations.

Changes in v1.1: metadata only, audio unchanged.

1. The 40 annotations flagged invalid_time_removed (18 test_soundscape, 22 train_xenocanto) now keep their point in time. Their times were never unknown: the pre-release annotation tables show them as single-point mis-clicks, with start_time == end_time and low_freq == high_freq. start_time/end_time are restored and the rows are relabelled label_status=zero_area_box. Species and vocalization type were always valid. The boxes still have no extent: filter them out for detection tasks.
2. Two genus-only scientific_name values completed, from their common name and eBird code: Thinornis -> Charadrius dubius (Little Ringed Plover, lirplo, 104 annotations) and Botaurus -> Botaurus minutus (Little Bittern, litbit1, 44). Botaurus stellaris is unaffected. schema.json and scientific_name_counts.json regenerated to match. No rows added or removed.

Known issue: XC667540 (XC001/001406.flac) is 19.99 s but has 4 annotations out to 26.43 s. The audio is truncated, not the metadata.