Please briefly describe your work and intended use. Your responses help us review access requests and understand future dataset needs. Do not include confidential or sensitive information.

Dataset Viewer

https://cc-task-oriented-preview.vercel.app/

otoSpeech-full-duplex-task-oriented-v2

Contact

This sample dataset is provided for research purposes. We maintain larger and more diverse datasets.

For collaborations, inquiries, custom data collection, or joint research, contact consome@oto.earth.

Dataset Summary

otoSpeech-full-duplex-task-oriented-v2 contains 17.376131 hours of English, full-duplex, two-speaker task-oriented conversations. It includes 47 sessions covering eleven collaborative task types, with 48 kHz channel-separated stereo audio, 8,301 synchronized interaction events, versioned task stimuli, and ten participant-created artifacts.

This release contains sessions collected after the latest session in otoSpeech-full-duplex-task-oriented-20h. The 58 sessions from that release are not duplicated in v2.

Both releases extend otoSpeech-full-duplex-280h with task state and interaction data aligned to the recording timeline.

Task Coverage

Task	Sessions
Photo Talk	13
Spot the Difference	9
Consensus Ranking	6
Word Games	5
Ambient AI in Chat	3
Voice-to-Form Entry	2
Hiring Decision	2
Tangram Direction	2
Describe-and-Draw Portrait	2
Canvas Co-Draw	2
AI-Assisted Task	1

Sessions may be complete or partial. task.completion_status, task.rounds_completed, and the event stream preserve that distinction. AI prompt and response text is included so the role-scoped interaction can be reproduced. During collection, Ambient AI messages were visible only to the participant who sent the request; the partner did not see that private chat. Redundant full conversation-history arrays are omitted from individual events.

How to Use

from datasets import Audio, load_dataset

dataset = load_dataset(
    "webdataset",
    data_files={"train": "data/train/shard-*.tar"},
    split="train",
    streaming=True,
)
dataset = dataset.cast_column("flac", Audio(decode=False))
sample = next(iter(dataset))

print(sample["json"]["task"]["task_id"])
print(sample["json"]["events"][0])

Each WebDataset sample consists of matching <session_id>.flac and <session_id>.json members. Audio channel 0 and channel 1 correspond to the speaker entries with the same channel values.

Dataset Contents

data/train/shard-*.tar: stereo FLAC audio and session JSON
artifacts/train/artifact-shard-*.tar: participant-created task artifacts
resources/: versioned task stimuli required to reproduce task state
schemas/*.json: JSON Schemas for sessions, events, and artifacts
export-manifest.json: release totals and task-level coverage

Speaker metadata follows the public profile fields used by otoSpeech-full-duplex-280h. Profile fields are opt-in and may be empty. The name field is intentionally always empty. speaker_id values are anonymized stable identifiers, while task_role and task_role_name describe each speaker's role in the selected task.

Event timestamps use seconds relative to recording start and are derived from the backend server_t clock. The included JSON Schemas define the machine-readable session, event, and artifact structures.

YFCC/Flickr images used by Photo Talk are not redistributed in this repository. resources/photo-talk/*/yfcc-index.json contains their source URLs, Flickr page URLs, attribution, and source licenses. Other task resources, including images created for this dataset, remain included as files.

Considerations for Using the Data

By using this dataset, you agree to:

Not attempt to identify speakers, de-anonymize participants, or link them to external identities.
Respect redaction segments and avoid reconstructing or inferring removed content.
Use the dataset in accordance with the license and applicable laws and regulations.

Licensing Information

CC BY 4.0 (Creative Commons Attribution 4.0 International)

Please review the full license text: https://creativecommons.org/licenses/by/4.0/

For commercial inquiries, collaboration, or a separate commercial agreement, contact consome@oto.earth.

Citation

@misc{otospeech_full_duplex_task_oriented_v2,
  title        = {otoSpeech-full-duplex-task-oriented-v2: Full-Duplex Task-Oriented Conversations},
  author       = {otoearth},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/otoearth/otoSpeech-full-duplex-task-oriented-v2}},
  note         = {License: CC BY 4.0}
}

Downloads last month: 22

Total file size:

3.43 GB