besimple-ai/duplex-cue

Duplex Cue

Duplex Cue is an audio benchmark for evaluating how an ongoing speaker responds when another speaker contributes during the turn. It separates the listener's local intent from the ongoing speaker's observed behavior, allowing systems to be evaluated for both content uptake and floor management.

The accompanying manuscript reports the complete study of 80 conversations and 300 canonical trials. This public repository preserves the manuscript and the full research code while releasing data-dependent artifacts for eight natural two-person conversations and the 15 canonical trials sourced from them.

Dataset contents

Item	Count
Conversations	8
Recorded conversation time	2:00:05.9
Pseudonymous speakers	12
Canonical evaluation trials	15
Cue classes represented	3
Evaluated conditions	Recorded source, PersonaPlex

The included conversations are:

Conversation	Approximate topic	Duration
conversation_014	Films, fictional characters, and hobbies	15:00.4
conversation_032	Photography, stock media, and generative AI	15:00.6
conversation_043	Met Gala fashion and charity events	15:00.9
conversation_047	Everyday technology and AI	15:00.4
conversation_052	Oaxacan festivals, food, and community traditions	14:59.2
conversation_067	Retro games, consoles, and video stores	15:02.0
conversation_069	Seasons, outdoor activities, sports, and films	15:00.9
conversation_072	News habits, meditation, exercise, and creative work	15:01.5

sample_manifest.json provides a machine-readable inventory of the conversations and trials.

Repository structure

conversations/
  conversation_NNN/
    metadata.json
    audio/
      mix.wav
      track_0.wav
      track_1.wav
      speech_altered/
    transcription/tracks/
    annotate/
    candidate_trials.jsonl
trials/
  trial_NNNN/
    metadata.json
    source/
    personaplex/
  selection.json
  results.json
app/
scripts/
paper/
  duplex-cue.tex
  duplex-cue.pdf
  provenance/
sample_manifest.json
ETHICS_AND_CONSENT.md
LICENSE_CODE.md
LICENSE_DATA.md
THIRD_PARTY_NOTICES.md

Each conversation directory includes synchronized source tracks, a stereo mix, fixed-stock-voice conversions, word- and segment-timed transcripts, cue annotations and reviews, candidate trials, and metadata. Each canonical trial includes recorded-source and PersonaPlex audio, its result review, and runner provenance.

The trial subset contains 11 backchannels, one collaboration, and three interruptions. conversation_069 and conversation_072 contribute conversation data but no canonical trials. This small, intentionally selected release should not be used to reproduce full-study aggregate results.

Paper, code, and provenance

The NeurIPS 2026 RTCA camera-ready paper is available as paper/duplex-cue.pdf with its matching LaTeX source in paper/duplex-cue.tex. It reports the complete study, not only the public data subset.

The full local explorer is under app/, and the annotation, generation, voice-conversion, scoring, and release utilities are under scripts/. The public provenance snapshots in paper/provenance/ and the global files under trials/ are recalculated or filtered to the eight-conversation release. See paper/README.md for the scope boundary and artifact inventory.

Privacy and participant protections

The conversations were screened automatically and then reviewed in full before release. The release excludes conversations containing direct contact details or account identifiers, and the transcripts do not name Besimple or Pila8.

The dataset is pseudonymized, not anonymous. Human voices can be identifying. Speaker metadata contains age-range buckets, sex, accent labels, and stable pseudonymous speaker IDs; natural conversations may contain contextual biographical information.

Users are responsible for complying with applicable privacy, publicity, biometric-data, and other laws. Besimple AI asks users not to identify speakers, link aliases to external identities, construct biometric profiles, clone or impersonate voices, or contact people mentioned in the recordings. See ETHICS_AND_CONSENT.md for the release review and safeguards.

Intended uses

Duplex Cue is intended for:

• evaluating full-duplex spoken-dialogue systems;
• studying responses to overlapping speech;
• measuring content uptake separately from turn yielding; and
• inspecting and improving turn-taking evaluation methods.

The dataset is not designed for speaker identification, biometric inference, surveillance, employment or eligibility decisions, voice cloning, impersonation, or attempts to recover participant identities.

Label schema

Cue intent describes Speaker B's contribution while Speaker A holds the floor:

• backchannel: supports A's ongoing speech without requesting a content change or taking the floor;
• collaboration: supplies an answer, correction, clarification, constraint, or completion without claiming the floor;
• interruption: attempts to take or retain the floor.

Observed response describes Speaker A's behavior:

• continued: proceeds without observable uptake of B's cue;
• adapted: acknowledges or incorporates B's contribution within the ongoing turn;
• yielded: hands over the floor in response to B.

A response may instead be excluded as silent or unusable.

Collection and processing

The source material consists of natural, unscripted two-person English conversations. Energy-based voice activity detection proposed overlapping regions, and ElevenLabs Scribe v2 supplied word-timed transcripts. Candidate cues were individually screened and labeled, then a human-confirmed cohort was selected for evaluation.

For the model condition, the ongoing speaker's full source track was converted with ElevenLabs Voice Changer to a fixed stock voice before conditioning nvidia/personaplex-7b-v1. The recorded listener channel remained on its original timeline. This conversion did not create a participant-specific voice model and is not presented as anonymization.

Generated and processed artifacts are clearly identified in their paths and provenance files. See THIRD_PARTY_NOTICES.md for provider attribution and terms.

Consent and ethics

All participant-recording pairs in this release are covered by recorded, versioned pre-capture acceptance of the Terms of Service, Privacy Policy, and Content Upload Agreement used for collection. Those agreements authorize research, model development and testing, derivative works, public distribution, and third-party licensing and sharing. Besimple AI has confirmed that these grants cover this public noncommercial release.

Besimple AI completed a documented consent, privacy, and release-risk review on September 11, 2026. The review included full transcript screening, data minimization, age bucketization, direct-identifier screening, provider-terms review, and confirmation that no known deletion-request conversation is included. The collection was not submitted to an institutional review board, and this release does not claim IRB approval or an IRB exemption. Details are in ETHICS_AND_CONSENT.md.

Automatic transcripts, activity detection, and annotations can be wrong. Human validation does not provide fully independent multi-reviewer coverage of every item. The sample is small, intentionally selected for lower content risk, demographically unrepresentative, and unsuitable for reproducing full-corpus aggregate results.

Loading

The repository uses Git LFS for audio:

git lfs install
git clone https://huggingface.co/datasets/besimple-ai/duplex-cue
cd duplex-cue

The nested research-artifact layout is not exposed as a standard datasets.load_dataset() configuration. Use the JSON, JSONL, and WAV files directly.

Licenses

Dataset and research materials are released under the Creative Commons Attribution-NonCommercial 4.0 International license. See LICENSE_DATA.md. Software under app/ and scripts/ is released under the MIT License. Third-party components and generated artifacts remain subject to the notices in THIRD_PARTY_NOTICES.md.

Citation

@misc{lu2026duplexcue,
  title  = {Continue, Adapt, or Yield: In-Turn Adaptation to Overlapping Speech in Full-Duplex Agents},
  author = {Lu, Yunqi and Baumgartner, Tyler and Johri, Nikhil and Tai, Brandon and Fan, Candice and Debaupte, Luc and Aguilar, Ruben and Wang, Bill and Zhong, Yi},
  year   = {2026}
}