Dataset Viewer
Auto-converted to Parquet Duplicate
Search is not available for this dataset
audio
audioduration (s)
6.78
225
label
class label
2 classes
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
0audio
End of preview. Expand in Data Studio

Chorus Duplex (ChorusDuplex)

The synthesis pipeline supports Mandarin and English. The audio currently published here is Mandarin, and the English batch is in production.

GitHub Hugging Face

A synthetic multi-party duplex speech dataset for studying who speaks, who is addressed, and when an assistant should respond, acknowledge, or stay silent. Synthesis code and CLI · 中文说明

Dialogue design

  • Participation and addressee cover exchanges directed to the assistant, exchanges among users or with third parties, and conversational transitions.
  • Backchannels acknowledge a point while the main speaker continues.
  • Simultaneous backchannels · Different listeners start together and may use the same or different words.
  • Sequential backchannels · One or more listeners respond in order, with each acknowledgement starting after the previous one ends.
  • Interruptions change who holds the floor. Two users can interrupt the same turn simultaneously.
  • Backchannel-to-interruption · A user starts with a brief acknowledgement while the assistant keeps speaking. The assistant stops at a pivot such as “but” or “wait.” Each new interrupted turn counts as a separate event.
  • Incomplete speech uses two fragments from the same speaker with a mid-thought pause. The generator samples pauses from 0.5 to 2 seconds and can include natural fillers.
  • Hesitation handoff · The user trails off, and the assistant waits 2.5–3.2 seconds before gently offering a suggestion.
  • Third-party speech forms a brief exchange within the conversation. Speaker, addressee, response references, and realized timing identify its relationship to the main thread.

Comparison with released duplex datasets

Full-Duplex-Bench is a one-to-one benchmark. MultiTalkFT is a multi-party training corpus.

Feature Full-Duplex-Bench MultiTalkFT ChorusDuplex
Multi-party dialogue ✗ ✓ ✓
Third-party exchange ✓ — ✓
Backchannel ✓ ✓ ✓
↳ Simultaneous backchannels — — ✓
↳ Sequential backchannels — — ✓
Interruption ✓ ✓ ✓
↳ Simultaneous interruption — — ✓
↳ Backchannel-to-interruption — — ✓
Incomplete turn ✓ — ✓
↳ Hesitation handoff — — ✓
Paired background speech ✓ — ✓
✓ Supported · — Unclear from public sources · ✗ Explicitly unsupported

Audio previews

Hesitation handoff · complete dialogue · The user trails off while talking about the budget. The assistant offers a suggestion at about 0:18.

Same dialogue with background speech

Sequential backchannels · Two listeners acknowledge the same suggestion in sequence at about 0:15.

Simultaneous backchannels · Two listeners start their acknowledgements together at about 0:07.

Simultaneous interruption · Two users interrupt the same assistant turn at about 0:11.

Backchannel-to-interruption · The user agrees briefly, then switches to a challenge at about 0:09.

Files and annotations

  • annotations.jsonl contains one record per dialogue, with the scenario, participants, turns, and profile.
  • audio/<id>.flac and audio/<id>_babble.flac are the paired clean and background-speech variants at 48 kHz, stereo.
  • audio/<id>.timeline.json gives realized start and end times and turn overlap.
  • speaker, addressee, kind, and responds_to identify the role and context of a turn. gap_before_ms and insert_ratio describe intended timing. An incomplete turn can additionally specify parts and pause_ms.

The left channel contains users and third parties. The right channel contains the assistant. The background version is an alternate mix of the same conversation, with VOiCES babble mixed into the left channel at a target SNR of 10 dB.

Quick start

Run pip install huggingface_hub soundfile. This example loads one annotation, its paired audio, and the turn timeline. Audio arrays are shaped as samples × channels.

import json

import soundfile as sf
from huggingface_hub import hf_hub_download

REPO = "BiggestWave/ChorusDuplex"


def fetch(path):
    return hf_hub_download(REPO, path, repo_type="dataset")


with open(fetch("annotations.jsonl"), encoding="utf-8") as f:
    example = json.loads(next(line for line in f if line.strip()))
dialogue_id = example["dialogue_id"]

with open(fetch(f"audio/{dialogue_id}.timeline.json"), encoding="utf-8") as f:
    timeline = json.load(f)

clean, sr = sf.read(fetch(f"audio/{dialogue_id}.flac"), always_2d=True)
babble, _ = sf.read(fetch(f"audio/{dialogue_id}_babble.flac"), always_2d=True)

print(dialogue_id, sr, clean.shape, babble.shape)
print(timeline[0]["speaker"], timeline[0]["start_sec"], timeline[0]["end_sec"])

Synthesis and attribution

The project generates dialogue scripts, synthesizes each role with MOSS-TTS-Nano, and places the speech on a shared timeline. The assistant uses the built-in Junhao voice. The pipeline also supports reference-based voices for other roles. These recordings are synthetic and can contain TTS or dialogue artifacts.

The dataset is distributed under CC BY 4.0. Background variants include material from VOiCES (Voices Obscured in Complex Environmental Settings), SRI International / Lab41, In-Q-Tel, CC BY 4.0. The background material is selected, level-adjusted, crossfaded and mixed with the dialogue. Preserve this attribution when redistributing the background variants. VOiCES source

Citation

Please cite the dataset if you use it in your work.

@misc{david2026chorusduplexsidetalk,
  author       = {David},
  title        = {ChorusDuplex: Chinese Multi-Party Duplex Dialogue Speech},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/BiggestWave/ChorusDuplex},
  note         = {Dataset}
}
Downloads last month
2,085