japht/inatspectro-bioacoustics

iNatSpectro Bioacoustics

A community-contributed dataset of bioacoustic call annotations pointing at iNaturalist audio observations. Contributions are made via the iNatSpectro browser extension: a user listens to and views the spectrogram of an iNaturalist recording, marks a call, and submits a pointer to the source recording plus the marked time/frequency bounding box. Each record carries the observation's quality_grade, so consumers can scope training to research-grade data (or any grade mix) as needed.

Dataset description

Each entry in this dataset corresponds to one annotation contribution against one iNaturalist audio file. No audio bytes and no rendered spectrogram pixels are stored — each record is a pointer (audio_url, observation_id, sound_id) to the source recording on iNaturalist, plus a human-placed annotation. Entries may carry a strong label (time_start/time_end/freq_low/freq_high identifying where a species call occurs) or be a weak label, where the recording is contributed without a call location marked. See "Weak vs. strong labels" below for how to detect which you have.

The dataset is append-only. Each contribution is stored as a separate NDJSON file at data/{observation_id}_{sound_id}_{unix_timestamp_ms}.ndjson. Because some iNaturalist observations have multiple audio files, sound_id (the iNaturalist sound ID) is needed alongside observation_id to uniquely identify a recording.

This is the v3 payload shape — a flat annotation pointer with no nested objects and no spectrogram render parameters. See schema/README.md in the project repo for the full history of how the schema evolved away from storing rendered pixels.

Dataset structure

Each NDJSON file contains a single JSON object with the following fields:

Field	Type	Description
schema_version	number	Payload contract version. Always 3 for records in this shape.
observation_id	string	iNaturalist observation ID (digits only)
sound_id	string	iNaturalist sound/media file ID (digits only) — identifies which recording on the observation this contribution refers to
audio_url	string	Canonical URL of the source recording on iNaturalist. Source of truth for re-fetching full-fidelity audio
contributed_at	string	ISO 8601 timestamp of contribution
time_start	number	Box start, seconds from start of recording
time_end	number	Box end, seconds from start of recording
freq_low	number	Lower frequency bound of the annotation box, Hz
freq_high	number	Upper frequency bound of the annotation box, Hz
annotation_type	string | null	Sound Annotation event type (e.g. echolocation, feeding-buzz, song), from the iNaturalist Sound Annotation observation field vocabulary. null when not specified
taxon	string | null	Per-event taxon name override for multi-species recordings. null when not specified
conf	number | null	Annotator confidence, 0-1. null when not specified
note	string | null	Free-text note attached to this annotation event. null when not specified
common_name	string | null	Common name of the observed species
scientific_name	string | null	Scientific name
iconic_taxon_name	string | null	Broad taxonomic group (e.g. "Mammalia")
quality_grade	string | null	iNaturalist quality grade at contribution time: "research", "needs_id", "casual", or null if unknown. Use to scope training by grade
inat_username	string | null	Contributor's iNaturalist username (opt-in), null if the contributor didn't opt in
sample_rate	number | null	Source audio sample rate in Hz, as stored in the audio file
time_expansion	integer | null	GUANO TE field: recording time-expansion factor. 1 = direct recording. Annotation box values are already corrected to real-world units
duration	number | null	Recording duration in seconds
profile	string | null	iNatSpectro analysis profile the extension applied at annotation time (e.g. "Bat") — a weak taxonomic-group signal, not sound metadata

Weak vs. strong labels

A record is a strong label when time_start is present with a real box (time_start/time_end/freq_low/freq_high all set) — a human placed a bounding box around a specific call. annotation_type being non-null additionally tells you the event category from the OFV vocabulary.

time_start/time_end/freq_low/freq_high are required fields on every v3 record, so in practice every record here is a strong label; a future weak-label path (recording contributed with no box marked) would be represented by these fields being absent. Check annotation_type, taxon, conf, and note for null when you need to know whether a per-event refinement was supplied on top of the box.

Source data

Observations are sourced from iNaturalist, a global citizen science platform. Contributions are accepted across all quality grades; each record's quality_grade field records whether the observation was research, needs_id, or casual at contribution time (only research-grade observations have community consensus on species identification). Filter on quality_grade to scope training accordingly.

No audio bytes or derived spectrogram pixels are stored in this dataset. Audio files remain hosted on iNaturalist and are referenced only by audio_url. To train or otherwise use the audio itself, fetch it from audio_url and compute spectrograms (or any other featurization) at use time — the dataset intentionally leaves that choice to the consumer rather than baking in one renderer's display parameters.

How to use

from datasets import load_dataset

ds = load_dataset("japht/inatspectro-bioacoustics", split="train")

# Example: fetch the source audio for one record and re-featurize it yourself
entry = ds[0]
audio_url = entry["audio_url"]
box = (entry["time_start"], entry["time_end"], entry["freq_low"], entry["freq_high"])

# import requests
# audio_bytes = requests.get(audio_url).content
# ... decode + compute your own spectrogram/features from audio_bytes,
#     using `box` to crop or label the region of interest

Deduplicate by recording (keep latest contribution per observation + sound file):

import pandas as pd

df = ds.to_pandas()
df = (
    df.sort_values("contributed_at")
      .drop_duplicates(subset=["observation_id", "sound_id"], keep="last")
)

Filter to strong labels only:

strong = ds.filter(lambda x: x["time_start"] is not None)

License

This dataset is released under CC BY 4.0. The annotation and metadata records in this dataset — the fields listed above — are contributed under CC BY 4.0 by iNatSpectro users.

Licensing of referenced audio

This dataset stores annotation pointers and metadata only — it never contains audio files or derived spectrogram data. Each source iNaturalist observation carries its own license, set by the original observer, and iNaturalist's own Terms of Service govern use of the platform's content.

The CC BY 4.0 license above covers the annotation/metadata layer only. It is the responsibility of anyone using this dataset to resolve audio_url back to the source observation and check that observation's current license before using the referenced audio (for training, redistribution, or any other purpose) — the same responsibility that already applies to any other reuse of iNaturalist- or GBIF-sourced data. Providing a pointer to publicly-viewable iNaturalist content does not itself grant or override the license terms on that audio.

This dataset does not include a license field, so license-based filtering must be done by the consumer against the observation's current license. For bulk audio download use cases that need this, the project's separate inatspectro-bulk-tools ingest tooling already implements license-based filtering (is_redistributable/is_commercial_use_ok in fetch.py, exposed via the --licensing {any,redistributable,commercial} CLI flag) as a reference for how to do this safely.

Contributing

Contributions are made via the iNatSpectro browser extension. Open any iNaturalist observation with audio, view the spectrogram, mark a call, and click Contribute to submit. The observation's quality grade is recorded with the contribution.