ROOTS

ROOTS contains 43,922,135 audio-language conversations across four taxonomy tiers. Audio is supplied by the source datasets below.

Quick start

from datasets import load_dataset

dataset = load_dataset("iclr2027anon/ROOTS", split="train", streaming=True)
row = next(iter(dataset))

Use the audio guide to locate and load each conversation's clips in order.

Columns

Columns	Meaning
id	Conversation ID: roots_ followed by 32 hexadecimal characters.
tier, category, task, format	Taxonomy tier, task family, specific task, and response format.
messages	Conversation turns with role and content.
source_dataset	Source label from the table below.
audio_paths, source_audio_ids, source_urls	Ordered source-relative paths, recording IDs, and recording-specific links.
audio_start_seconds, audio_end_seconds	Crop boundaries within each referenced file, in seconds; null means the complete file.

All audio lists have the same length. Preserve their order and repeated clips. For the three hosted audio datasets, source_audio_ids matches their metadata id exactly. Crop times are separate. Other sources retain their upstream identifiers. id_namespaces.json gives the repository and revision for each hosted dataset.

Taxonomy

category follows the paper's task families. taxonomy.json maps all 81 tasks to categories and tiers.

Tier	Capability	Categories
1	Acoustic perception	Acoustic-property estimation; acoustic description matching; acoustic captioning.
2	Biological category recognition	Species and taxonomy recognition; taxon presence detection; call-type recognition; life-stage and behavior prediction (including sex); semantic captioning.
3	Scene understanding	Event and species counting; temporal localization; co-occurrence; order reasoning; acoustic comparison across events; structured scene description.
4	In-context learning	Audio-example-conditioned classification; few-shot detection; call-type matching.

Species listing belongs to Tier 2, under species and taxonomy recognition.

Source datasets

ROOTS source	Conversations	Audio source	File selection
animal-sound-archive	895,935	Animal Sound Archive	Full recording ID, including TSA: or SUB:.
animalspeak-pseudovox	466,680	Pseudovox	Indexed WAV member in a TAR archive.
birdset	65,574	BirdSet	Subset and named test_5s clip; crop times are within that clip.
f0-bioacoustic	56,698	F0 Bioacoustic / Dryad	Taxon ZIP and case-sensitive filename; use the complete cut file.
inaturalist	12,588,369	iNaturalist	Sound ID plus observation ID; an observation can contain several sounds.
synthetic-call-type-tasks	200,078	Task-specific synthetic audio	Use the task-specific IDs and audio order.
synthetic-fewshot-detection	100,000	Task-specific synthetic audio	Use the task-specific IDs and audio order.
synthetic-sed	4,084,734	Strong detection	Indexed scene WAV and crop interval.
synthetic-sed-diarization	4,019,676	Detection and diarization	Indexed scene WAV and crop interval.
wabad	269,946	WABAD / Zenodo	Full recordings with site folders; apply the stated crop.
watkins	1,175,191	Watkins / WHOI	Original recording filename from the source ID.
xeno-canto	19,949,254	Xeno-Canto	XC recording ID.
xeno-canto+inaturalist	50,000	Xeno-Canto and iNaturalist, linked above	Resolve each audio slot separately and preserve its position.

Source licenses and attribution requirements apply to downloaded audio. Pseudovox's original recording links are in its own metadata.

Files

• data/train-*.parquet: conversations.
• roots_audio.py and audio_assets.sqlite.gz: audio resolver and index of 4,745,972 distinct references reused across conversations.
• templates/: generation templates.