Duplex-MPE

Benchmarking Multi-Party Interaction in Full-Duplex Dialogue

Paper (arXiv:2609.31948) · Project page · Hugging Face · ModelScope

Duplex-MPE evaluates when a full-duplex speech assistant should answer, remain silent or stop speaking in a conversation among several people. Each scenario contains three or four human speakers and the assistant Aria. The model receives continuous conversation audio, with no transcript, speaker label or turn boundary, and must decide on its own when to speak.

The dataset contains 2,000 scenario pairs (4,000 scenarios). In each pair, the request to Aria (T) names the assistant in the explicit scenario and identifies it only through conversational context in the implicit scenario. Both versions keep the same task and the same gold answer.

Statistics

	Explicit	Implicit
Scenarios	2,000	2,000
Human turns	26,182	26,182
T: unresolved request to Aria	2,000	2,000
N1: mentions Aria without a request to it	3,041	3,039
N2: addressed to another person or device	14,052	14,054
N3: no designated addressee, including self-talk	3,267	3,267
N4 question to Aria (N4_Q)	1,911	1,911
N4 human resolution (N4_R)	1,911	1,911
Human speech, counted per occurrence	52.58 h	52.97 h

N1–N3 turns are the 20,360 silence-requiring windows per condition. Clips that occur in several scenarios are stored once: the release contains 31,025 WAV files (61.1 hours of distinct audio, 10.56 GB), including the spoken duty preamble. Equal totals do not imply that the two scenarios of a pair have identical events or audio; local wording outside T and the N4 events can differ.

Download

The same files are hosted on Hugging Face and ModelScope.

# Hugging Face
hf download ChengqianMa/Duplex-MPE --repo-type dataset --local-dir Duplex-MPE

# ModelScope
modelscope download ChengqianMa/Duplex-MPE --repo-type dataset --local-dir Duplex-MPE

After downloading, sha256sum -c SHA256SUMS inside the dataset directory verifies every file.

Files

README.md
LICENSE
dataset.json                     summary statistics and preamble metadata
manifests/explicit.json          2,000 scenarios whose T request names Aria
manifests/implicit.json          2,000 scenarios whose T request does not name Aria
audio/aria_duty_preamble.wav     spoken duty instruction played before every scenario
audio/<hash-prefix>/<sha256>.wav one file per distinct human-speech clip
audio_inventory.json             byte size, SHA-256, sample count and sample rate of every audio file
SHA256SUMS                       checksums of all files in the release

All audio is mono, 16-bit PCM WAV at 24 kHz. Audio paths in the manifests are relative to the dataset root. The files are byte-for-byte copies of the clips used in the paper's experiments; they were not re-synthesised, trimmed, resampled or normalised. Each file is named by its SHA-256, so identical clips are stored once.

Manifest fields

Each manifest is a JSON list of scenarios.

Field	Meaning
scenario_id	Scenario identifier, e.g. duplex_mpe_0000_explicit
pair_id	Shared identifier of the explicit and implicit scenarios of one pair
addressing	explicit or implicit; the addressing condition of the scenario's T request
trigger_index	Index of the scenario's single T turn
gold_answer	Intended answer to T, used for scoring only
supporting_indices	Earlier turns that support the answer, used for analysis only
setting, activity, relation, device_context, register	Scenario factors
n_speakers	Number of human speakers (3 or 4)
turns	Ordered list of human turns

Each turn contains turn_index (zero-based), speaker (A–D), text, event_type (T, N1, N2, N3, N4_Q or N4_R), addressee (aria, other_human, device, none or animal), speech_act (statement, request, question, self_talk or backchannel), audio_path, duration_s and sample_rate.

Use addressing to assign a scenario to a condition and pair_id to match the two scenarios of a pair; do not infer either from file names. Speaker letters identify voices in the script; names spoken inside a dialogue need not correspond to them. Gold answers, event labels and annotations are for scoring and analysis and must not be given to the evaluated model.

Loading example

import json
from pathlib import Path

import soundfile as sf

root = Path("Duplex-MPE")  # local copy of this repository
scenarios = json.loads((root / "manifests/explicit.json").read_text())

scenario = scenarios[0]
t_turn = scenario["turns"][scenario["trigger_index"]]
audio, sample_rate = sf.read(root / t_turn["audio_path"])
print(scenario["scenario_id"], t_turn["text"], scenario["gold_answer"], sample_rate)

Evaluation protocol

The audio is the model input; the manifests do not prescribe a fixed timeline because the waiting time after T depends on the evaluated model. In the paper, each scenario is played as follows:

1. Start from a fresh model state and play the 31.92 s duty preamble, then 0.5 s of silence.
2. Play the human turns in order, with 0.5 s of silence between consecutive clips.
3. After T, the model has 5 s to start speaking. Once it speaks, playback waits for 2 s of confirmed silence or a 60 s cap.
4. After N4_Q, wait 3 s and then play N4_R, whether or not the model has stopped speaking.

The duty preamble says:

You are the voice assistant Aria, that is A R I A, Aria. You are in a room with several people. Only respond when someone speaks directly to you, Aria, and asks something you can answer. When people are talking to each other or to another device, or thinking out loud, or when the question they asked you has already been answered by someone else, stay silent and do not speak.

The paper defines four scores on separate denominators: fresh-onset response rate, conditional answer accuracy, silence preservation and answering-window yield, plus two coverage diagnostics. See the paper for the metric definitions, detector settings and scoring rules.

Construction

Scenario scripts, paired rewrites and gold answers were generated with Claude Opus 5. Every human turn was synthesised separately with Qwen3-TTS-CustomVoice (Apache-2.0) using the preset voices ryan, aiden, eric and dylan for speakers A, B, C and D. No recordings of real people were used. Human validation of the scripts, labels and synthesised speech is reported in the paper.

Intended use and limitations

Duplex-MPE is intended for evaluating selective participation of speech assistants in multi-party conversation. The conversations are synthetic and in English, and all speakers use four preset TTS voices. The benchmark does not measure robustness to natural recording conditions, background noise or diverse real voices.

License

The dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Please cite the paper when you use it.

Citation

@article{ma2026duplexmpe,
  title   = {Duplex-MPE: Benchmarking Multi-Party Interaction in Full-Duplex Dialogue},
  author  = {Ma, Chengqian and Feng, Wenhao and Jin, Weixuan and Dai, Gaole and Xie, Tianyu and Ma, Yuexiao and Kang, Zhaolu and Zhao, Xiangyu and Zheng, Xiawu and Chao, Fei},
  journal = {arXiv preprint arXiv:2609.31948},
  year    = {2026}
}