japht/inatspectro-bioacoustics

iNatSpectro Bioacoustics

A community-contributed dataset of bioacoustic spectrograms derived from iNaturalist audio observations. Contributions are made via the iNatSpectro browser extension, which renders spectrograms directly in the browser and lets users submit species annotations and recording metadata for ML training. 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 iNaturalist audio file. The spectrogram is rendered in-browser by iNatSpectro, and the contributor submits the recording metadata and (optionally) an annotation marking where the species call occurs. Entries may include a strong label (an annotated time/frequency bounding box identifying where a species call occurs) or a weak label (species is known to be present in the audio, but the exact location is not annotated).

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.

Dataset structure

Each NDJSON file contains a single JSON object. All annotation fields are top-level (flat) — there is no nested object.

Field	Type	Description
schema_version	number	Schema version of this record (currently 2)
observation_id	string	iNaturalist observation ID
sound_id	string	iNaturalist sound ID
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 undetermined
audio_url	string	Source audio URL on iNaturalist
sample_rate	number | null	Sample rate of the audio file in Hz (file rate, not effective real-world rate for TE recordings)
time_expansion	number | null	GUANO time-expansion factor (null = direct recording). For bat detectors that slow recordings; annotation bounding boxes are already in real-world units.
duration	number | null	Duration of the source audio in seconds
profile	string | null	iNatSpectro analysis profile used (e.g. "Bat", "Bird", "Frog")
time_start	number | null	Annotation box start, seconds from start of recording. null = weak label.
time_end	number | null	Annotation box end, seconds from start of recording. null = weak label.
freq_low	number | null	Lower frequency bound of annotation box, Hz. null = weak label.
freq_high	number | null	Upper frequency bound of annotation box, Hz. null = weak label.
annotation_type	string | null	Sound event category from the OFV vocabulary (e.g. "echolocation", "song", "feeding-buzz"). null if not specified.
taxon	string | null	Per-event taxon name override (for multi-species recordings). null if not specified.
conf	number | null	Annotator confidence, 0–1. null if not specified.
note	string | null	Free-text note on the annotation event. null if not specified.
inat_username	string | null	Contributor's iNaturalist username (opt-in). null when contributor did not opt in.
contributed_at	string	ISO 8601 timestamp of contribution

Strong vs weak labels

Records where time_start is null are weak labels — the species is confirmed present somewhere in the audio but the exact location is not annotated. Records with a non-null time_start are strong labels with a time/frequency bounding box.

Source data

Observations are sourced from iNaturalist, a global citizen science platform. Contributions are accepted across all quality grades; the quality_grade field records the grade at contribution time ("research", "needs_id", "casual") — only research-grade observations have community consensus on species identification, so it is common to filter on this field for training.

Audio files remain hosted on iNaturalist and are referenced by audio_url. The spectrogram was rendered locally in the contributor's browser using iNatSpectro; this dataset stores the resulting metadata and annotations.

How to use

from datasets import load_dataset

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

# Example: inspect an entry
entry = ds[0]
print(entry["scientific_name"], entry["annotation_type"], entry["time_start"], entry["time_end"])

Deduplicate by recording (keep latest contribution per observation + audio 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)

Filter to research-grade observations only:

research = ds.filter(lambda x: x["quality_grade"] == "research")

License

This dataset is released under CC BY 4.0. Source audio observations on iNaturalist are subject to their individual observation licenses as set by their authors. Spectrogram float data and metadata in this dataset are contributed under CC BY 4.0 by iNatSpectro users.

Contributing

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