Benchmark report
Duplex Cue: Does a voice agent adapt while speaking?
An audio benchmark for measuring how an ongoing speaker responds when a listener contributes during the turn—whether the speaker continues unchanged, adapts in-turn, or yields the floor.
Problem
Listening while talking is not the same as responding.
A listener may supply a missing word, offer a correction, acknowledge the story, or try to take the floor. A full-duplex agent can keep talking through the cue, incorporate it without stopping, or hand over the turn. A simple stop-or-continue score cannot tell those behaviors apart.
The ongoing speaker is mid-thought.
A backchannel, collaboration, or interruption overlaps the turn.
A continues, adapts, or yields.
Evaluation design
Separate what the listener meant from what the speaker did.
Duplex Cue assigns two independent labels to each local interaction. Cue intent describes Speaker B's contribution. Observed response describes Speaker A's behavior. The resulting three-by-three matrix makes in-turn uptake visible without treating every overlap as an interruption.
Cue intent
Supports the ongoing speech without asking for a content change or claiming the floor.
Supplies an answer, correction, clarification, constraint, or completion without taking the floor.
Attempts to take or retain the floor from the ongoing speaker.
Observed response
The speaker proceeds without observable uptake of the listener's cue.
The speaker acknowledges or incorporates the cue while keeping the turn.
The speaker hands over the floor in response to the cue.
Dataset
Natural conversations, reviewed at the cue level.
The release starts with 80 synchronized, two-channel English conversations and an inventory of 2,591 reviewed overlapping-speech cues. The fixed evaluation cohort contains 300 human-confirmed trials, balanced across three cue classes.
Natural, unscripted two-person English conversations
Synchronized two-channel conversational audio
Stable public aliases across the released corpus
Backchannels, collaborations, and interruptions
100 human-confirmed cues in each cue class
Both responses active at cue onset and scorable
Results
PersonaPlex adapted less often to collaborative cues.
The primary analysis compares 208 identical cues where both the recorded speaker and the generated speaker were active at cue onset and produced a scorable response. On the 66 collaboration pairs, recorded speakers adapted in-turn nearly twice as often as PersonaPlex.
- 01
On identical collaborative cues, recorded speakers adapted within the turn 68.2% of the time; PersonaPlex adapted 34.8% of the time.
- 02
For backchannels, continuation was the dominant response in both conditions: 71.6% for recorded speakers and 70.4% for PersonaPlex.
- 03
For interruptions, PersonaPlex yielded more often than recorded speakers—50.8% versus 39.3%—while recorded speakers adapted more often.
Backchannel
81 paired cuesCollaboration
66 paired cuesInterruption
61 paired cues| Cue | Condition | Continued | Adapted | Yielded | N |
|---|---|---|---|---|---|
| Backchannel | Recorded source | 71.6% | 23.5% | 4.9% | 81 |
| Backchannel | PersonaPlex | 70.4% | 21.0% | 8.6% | 81 |
| Collaboration | Recorded source | 22.7% | 68.2% | 9.1% | 66 |
| Collaboration | PersonaPlex | 42.4% | 34.8% | 22.7% | 66 |
| Interruption | Recorded source | 14.8% | 45.9% | 39.3% | 61 |
| Interruption | PersonaPlex | 16.4% | 32.8% | 50.8% | 61 |
Coverage
The headline result is conditional on activity and scorability.
The benchmark begins with 300 selected trials. Pairwise activity and response-quality checks retain 208 for the primary comparison. Reporting that funnel matters: it keeps generation failures separate from the behavior distribution among comparable responses.
| Stage | Backchannel | Collaboration | Interruption | Total |
|---|---|---|---|---|
| Selected paired trials | 100 | 100 | 100 | 300 |
| Activity-ineligible pairs | 16 | 33 | 38 | 87 |
| Both responses active | 84 | 67 | 62 | 213 |
| Unusable response among active pairs | 3 | 1 | 1 | 5 |
| Primary paired trials | 81 | 66 | 61 | 208 |
Protocol
How the paired comparison works
Find and confirm cues
Overlapping regions are screened and labeled by local intent, then a balanced set of 300 cues is confirmed through human audio review.
Generate the paired response
PersonaPlex continues Speaker A from the same conversational history while Speaker B's recorded cue remains on its original timeline.
Classify observed behavior
Recorded and generated responses are independently labeled Continued, Adapted, or Yielded with timed evidence and coverage checks.
Limitations
What this release can—and cannot—show
- The current case study evaluates one PersonaPlex checkpoint, one continuation setup, and one generation per trial.
- The corpus is English-only, includes repeated participants, and is not demographically representative.
- The recorded listener cannot react to the generated speaker, so the evaluation is conditional rather than fully interactive.
- Automatic transcription, activity detection, and response review can be wrong; human validation is not fully independent for every item.
- The primary result describes the 208 active, scorable pairs, not all 300 selected trials or the natural prevalence of cue types.
- Response labels capture local interaction behavior, not factual correctness or overall conversation quality.
Privacy and use
Access controls do not make human speech anonymous.
The release contains human voices and conversational transcripts. Stable aliases reduce direct linkage to source systems but do not de-identify the audio or its content. Users must not identify speakers, build biometric profiles, clone or impersonate voices, or contact people mentioned in recordings.
The repository does not include a standalone open-data license. Public availability would not grant permission for unrestricted redistribution, commercial use, biometric processing, voice cloning, or identity inference.
Next
Dataset, explorer, and paper
Hugging Face contains the benchmark audio and annotations, canonical trial artifacts, recorded and PersonaPlex responses, scoring results, a local dataset explorer, reproducibility scripts, and the accompanying paper.