kennethli319/seamless-interaction-jefferson-annotations

Seamless Interaction Jefferson-Style Annotations

An automatic, turn-oriented annotation layer for the Meta Seamless Interaction Dataset. It compares the dataset's traditional transcript with an ASR-derived Jefferson-style condition and supplies speech-act, communicative-purpose, interactional-signal, alignment, and quality fields.

This is a derived noncommercial research dataset. It does not redistribute the source audio. Every record retains the original interaction ID, split, speaker, timestamps, and canonical Meta audio URLs.

Coverage

• Conversations: 5,120
• Train conversations: 4,372
• Dev conversations: 748
• Segment rows: 504,610
• Grouped turn rows: 153,379
• Accepted Jefferson segments: 278,289

All Naturalistic/dev interactions are included. Naturalistic/train uses a deterministic equal-count sample capped by the smallest category:

• Charades: 1,093
• Collaborative story: 1,093
• Conversation: 1,093
• Grounded gesture: 1,093

Sampling seed: turnwise-naturalistic-balanced-train-v1.

Configurations

• conversations: conversation metadata, source audio links, and summary counts.
• segments: traditional and Jefferson-style segment rows with timestamps, existence status, cue counts, and quality reasons.
• turns: aligned transcript conditions, limited-taxonomy speech acts, one-to-two-sentence attention/purpose/implication fields, Jefferson signals, alignment, and quality metadata.

The repository also includes compressed per-conversation records in browser/ and exact release-level summaries in analysis/. The browser records preserve both transcripts, every grouped turn, annotations, quality evidence, canonical source-audio URLs, and a split field without redistributing source audio.

Speech-act taxonomy

• INFORM — Present an experience, fact, plan, or account as new content.
• ELABORATE — Add a reason, example, consequence, or detail.
• RESPOND — Give a substantive answer after a partner's invitation.
• INVITE — Solicit the partner's answer, view, or next contribution.
• ALIGN — Acknowledge, affirm, encourage, or display attention.
• CO_CONSTRUCT — Advance content the speakers are jointly building.
• EVALUATE — Express a judgment, preference, stance, or assessment.
• MANAGE_FLOOR — Coordinate speaking order, timing, or completion.
• REPAIR — Correct, clarify, restart, or resolve speaking trouble.
• SOCIAL — Perform a courtesy, greeting, thanks, or closing move.

Interactional notation fields

The Jefferson-style condition exposes coded evidence for hesitation (micropauses, timed pauses, filled pauses), emphasis (prominence, lengthening, pitch movement), overlap, relatively quiet delivery, and uncertain hearing. These marks are automatically inferred from ASR timestamps and token/audio signals; they should not be treated as human conversation-analysis judgments.

Method

1. Decode both isolated speaker channels with Whisper large-v3-turbo Q5 and a disfluency-preserving initial prompt.
2. Derive deterministic crosstalk suppression from the prompted decode.
3. Run an independent unprompted decode as an existence check.
4. Retain confirmed and probable Jefferson candidates in the primary condition; keep uncertain and rejected candidates in the segment table for inspection.
5. Group adjacent accepted same-speaker segments into turns at gaps of 2.4 seconds or less, align nearby same-speaker reference material, and apply the ten-label deterministic speech-act methodology.

Full hashes, runtime settings, sampling provenance, and file checksums are in provenance/ and release-manifest.json.

Intended and out-of-scope uses

Intended uses include noncommercial research on transcript representation, conversation understanding, speech-act inference, interactional-signal distributions, and human/audio-grounded evaluation of automatic annotations.

Do not use these annotations as verified psychological states, clinical or employment evidence, identity or demographic inference, surveillance labels, or a substitute for listening to the source audio. Aggregate rates describe this constructed sample. Because train is sampled equally across four interaction categories, they are not population estimates for the original train split.

Important limitations

• These are automatic Jefferson-style transcripts, not human-produced Jefferson transcripts.
• Communicative-purpose and implication fields are deterministic model-derived interpretations, not ground-truth claims about speakers' private mental states.
• The unprompted existence pass uses the same ASR family and is not an independent human label.
• Source timestamp noise, speaker bleed, ASR errors, omissions, and hallucinated wording can remain. Use the audio and quality fields for consequential review.
• Source channels with zero audio frames are retained as empty Jefferson tracks; any available reference material remains visible as traditional-only turns.
• The dataset is intended for noncommercial research and analysis.
• The source contains recordings of human participants. Follow the source dataset's consent, access, privacy, attribution, and acceptable-use terms; do not attempt to identify participants or infer sensitive traits.

License and attribution

The source Seamless Interaction Dataset is licensed under CC BY-NC 4.0. This derived release uses the same license, credits Meta, links the source, and marks the material as modified. No endorsement by Meta is implied.

Please cite the original Seamless Interaction publication and identify this release and its version when using these annotations.

@article{agrawal2025seamlessinteraction,
  title={Seamless Interaction: Dyadic Audiovisual Motion Modeling and Large-Scale Dataset},
  author={Agrawal, Vasu and others},
  journal={arXiv preprint arXiv:2506.22554},
  year={2025}
}