ULM-DS-Lab/indonesian-bird-sounds

Indonesian Bird Sounds (Few-Shot)

Dataset Description

A curated dataset of bird sound recordings from Indonesia, designed for few-shot audio classification research on under-represented tropical bird species.

Dataset Summary

• Source: Xeno-Canto — a citizen-science repository of bird sounds worldwide
• Task: Few-shot bird species classification (audio)
• Region: Indonesia (Eastern Indonesia focus — Maluku, Papua, Sulawesi, Java, Sumatra, Borneo)
• Format: Raw MP3 audio files organized in Species_name/xc_id.mp3 folders
• Species: 26 Indonesian bird species
• Total Recordings: 1,919 audio files
• Recordings per Species: 51–136 (mean: 73.8)
• Metadata: metadata.csv with 18 columns per recording (species, location, coordinates, quality, recordist, date, license, etc.)

Supported Tasks

Primary task:

• Few-shot audio classification: Classifying bird species from audio recordings with limited labeled examples per class

Secondary methodological tasks:

• Bioacoustic representation learning in tropical ecosystems
• Transfer learning from pre-trained audio models (e.g., BirdNET, Perch, YAMNet) to low-resource species
• Active learning strategies for wildlife monitoring
• Cross-domain evaluation of general-purpose audio embeddings on tropical bird calls

Note: This dataset contains raw, unprocessed field recordings of varying length and quality. For general bird sound classification at scale, consider larger datasets such as BirdCLEF. This dataset specifically targets the few-shot regime for under-represented Indonesian species absent from most training corpora.

Dataset Structure

Data Files

indonesian_birds_fewshot/
├── README.md
├── metadata.csv
├── Brachypteryx_leucophris/
│   ├── 123456.mp3
│   └── ...
├── Cacomantis_sepulcralis/
│   └── ...
└── ... (26 species folders)

Each species folder is named using the scientific name with underscores (e.g., Megapodius_freycinet/). Audio files are named by their Xeno-Canto recording ID (e.g., 1059986.mp3).

Species List (26 species)

#	Scientific Name	English Name	Recordings
1	Brachypteryx leucophris	Lesser Shortwing	66
2	Cacomantis sepulcralis	Rusty-breasted Cuckoo	85
3	Cinnyris jugularis	Olive-backed Sunbird	120
4	Corvus enca	Slender-billed Crow	58
5	Dicrurus bracteatus	Spangled Drongo	100
6	Dicrurus hottentottus	Hair-crested Drongo	90
7	Eclectus roratus	Eclectus Parrot	56
8	Geoffroyus geoffroyi	Red-cheeked Parrot	69
9	Locustella castanea	Chestnut-backed Bush Warbler	79
10	Lycocorax pyrrhopterus	Paradise-crow	52
11	Macropygia doreya	Sultan's Cuckoo-Dove	58
12	Malacocincla sepiaria	Horsfield's Babbler	67
13	Megapodius freycinet	Dusky Megapode	54
14	Origma murina	Papuan Scrubwren	56
15	Orthotomus sepium	Olive-backed Tailorbird	51
16	Otus magicus	Moluccan Scops-Owl	64
17	Phyllergates cucullatus	Mountain Tailorbird	136
18	Phylloscopus poliocephalus	Island Leaf Warbler	68
19	Pitta maxima	Ivory-breasted Pitta	77
20	Pitta sordida	Hooded Pitta	61
21	Pnoepyga pusilla	Pygmy Cupwing	103
22	Psilopogon armillaris	Flame-fronted Barbet	60
23	Rhipidura rufiventris	Northern Fantail	58
24	Rhyticeros plicatus	Papuan Hornbill	56
25	Tanysiptera galatea	Common Paradise Kingfisher	123
26	Tesia superciliaris	Javan Tesia	52

Total: 1,919 recordings

Data Fields

metadata.csv contains one row per recording with the following columns:

Column	Type	Description
species	string	Scientific name (e.g., "Megapodius freycinet")
id	int	Xeno-Canto recording ID
genus	string	Genus name
specific_epithet	string	Species epithet
subspecies	string	Subspecies name (if identified; may be empty)
english_name	string	Common English name
country	string	Country of recording (all "Indonesia")
location	string	Locality description (e.g., "Waisai, Raja Ampat Regency, Southwest Papua")
latitude	float	Recording latitude (may be empty)
longitude	float	Recording longitude (may be empty)
quality	string	Xeno-Canto quality rating: A (best) through E (worst); may be empty if not rated
length	string	Recording duration in M:SS or H:MM:SS format
recordist	string	Name of the person who made the recording
date	string	Recording date (YYYY-MM-DD)
time	string	Recording time (HH:MM); ? if unknown
file	string	Direct MP3 download URL
url	string	Xeno-Canto recording page URL
license	string	Creative Commons license URL for the recording

Source Data

Original Source

All recordings are sourced from Xeno-Canto, a citizen-science platform where volunteer recordists contribute bird sound recordings from around the world.

Xeno-canto Foundation and Naturalis Biodiversity Center (2005–2025). xeno-canto — Bird sounds from around the world. https://xeno-canto.org/

Data Collection

• Curation method: Species were selected based on geographic range (Indonesian endemics and near-endemics) and data availability on Xeno-Canto, targeting species with 50–150 recordings to create a naturally few-shot scenario
• Recording conditions: Field recordings made by citizen-science contributors across Indonesia, varying in equipment, environment, and recording quality
• Geographic coverage: Recordings span major Indonesian biogeographic regions including Maluku, Papua, Sulawesi, Java, Sumatra, and Borneo
• Temporal range: Recordings collected across multiple decades by different recordists
• Quality ratings: Xeno-Canto community-assigned quality ratings (A–E) are preserved in the metadata; not all recordings have been rated
• Licenses: Individual recordings carry Creative Commons licenses as specified by each recordist (predominantly CC BY-NC-ND 4.0 and CC BY-NC-SA 4.0)

Usage Example

import pandas as pd
import librosa
from pathlib import Path

# Load metadata
metadata = pd.read_csv("indonesian_birds_fewshot/metadata.csv")
print(f"{len(metadata)} recordings across {metadata['species'].nunique()} species")

# Load a single audio file
species = "Pitta maxima"
species_folder = species.replace(" ", "_")
sample_id = metadata[metadata['species'] == species]['id'].iloc[0]

y, sr = librosa.load(f"indonesian_birds_fewshot/{species_folder}/{sample_id}.mp3", sr=22050)
print(f"Loaded {species} recording {sample_id}: {len(y)/sr:.1f}s at {sr} Hz")

# Build few-shot episodes
from sklearn.model_selection import train_test_split

species_list = metadata['species'].unique()
support_ids, query_ids = [], []

for sp in species_list:
    sp_ids = metadata[metadata['species'] == sp]['id'].values
    support, query = train_test_split(sp_ids, train_size=5, random_state=42)
    support_ids.extend([(sp, sid) for sid in support])
    query_ids.extend([(sp, qid) for qid in query])

print(f"Support set: {len(support_ids)} (5-shot), Query set: {len(query_ids)}")

Citation

If you use this dataset in your research, please cite:

@dataset{ulm_indonesian_bird_sounds_2025,
  title     = {Indonesian Bird Sounds (Few-Shot)},
  author    = {{ULM Data Science Lab}},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ULM-DS-Lab/indonesian-bird-sounds}
}

License

All audio recordings are sourced from Xeno-Canto and are shared under individual Creative Commons licenses as specified by each recordist (see the license column in metadata.csv). The majority of recordings are shared under CC BY-NC-ND 4.0 and CC BY-NC-SA 4.0. The dataset as a whole is distributed under CC BY-NC 4.0.

Contact

F Indriani - f.indriani@ulm.ac.id Lambung Mangkurat University

For questions about the original recordings, please contact the individual recordists via Xeno-Canto.