ChorusDuplex-SideTalk

Most speech datasets teach a model what to say. This one also teaches when to stay silent.

大多数语音数据在教模型怎么说；这份数据补上另一半——什么时候不该说。（中文版 →）

100 Chinese multi-party duplex dialogues (2–3 users + 1 AI assistant), each fully annotated with a speech score: who speaks when, to whom, whether it is a normal turn, an interruption or a backchannel (with real audio-level overlap), and why the AI assistant speaks — or why it stays silent (e.g. when someone turns to a third party).

Each dialogue is provided in two audio variants — 200 audio files in total:

File	Content
audio/{id}.flac	clean dialogue — 100 files
audio/{id}_babble.flac	same dialogue + background speakers (recorded human babble) mixed into the non-assistant channel — 100 files

Audio is stereo: left = non-assistant (users / third party), right = assistant.

---

Three rare things (why this dataset is worth downloading)

1. Third-party side-talk: a speaker turns to someone outside the conversation (delivery courier, phone call, neighbor) — annotated with its own voice and turns. Such "AI should NOT respond" negatives are almost absent from public Chinese speech data.
2. Audio-level overlap, not text-level annotation: interruptions and backchannels are actually mixed on the timeline with negative gaps. Each turn carries gap_before_ms (negative = speak early, i.e. overlap) and insert_ratio (where inside the previous turn a backchannel lands).
3. Controllable AI participation, with labels: every dialogue has a participation profile (intensity low/mid/high + trigger type), directly usable to train/evaluate when an assistant should chime in (cue / volunteer / invited / coordinate).

What's new in this release (v2, 2026-09)

Area	v1 (first release)	v2 (this one)
Audio layout	48 kHz mono	48 kHz stereo: L = non-assistant, R = assistant
Background speakers	none	_babble variants: recorded human babble mixed into the non-assistant channel (10 dB SNR)
Turn edges	hard cuts	8 ms fade-in/out per turn (no clicks)
AI replies	—	directionality rules: the assistant addresses specific people (name / title / pronoun) instead of always "talking to everyone"; multi-target ratio is measured and bounded
Identity perception	—	in-dialogue grounding: names must be spoken in the dialogue before the assistant may use them; gender / age references must match the assigned voice
Diversity	few recurring topics	15 topic domains, per-dialogue persona subsets, rotating voice cast
Side-talk placement	could land at the end	placed mid-dialogue (≥ 3 turns before the end), with the AI required to re-anchor to the main topic afterwards
Timeline file	timing only	timing + turn text (text field)

Contents (v2)

Item	Value
Dialogues	100
Audio files	200 (100 clean + 100 with background)
Turns	1,667
Audio duration	≈ 3.4 hours per variant (48 kHz stereo)
Scenario	side_talk (café / home / office / study room …)
Speakers	2–3 users + 1 AI assistant per dialogue (+ optional third party)
Turn kinds	speech 1443 / side_talk 134 / backchannel 55 / interruption 35
AI participation	low 34 / mid 37 / high 29
AI triggers	cue 91 / volunteer 72 / invited 46 / coordinate 17
With third-party speech	98 / 100 dialogues

File layout

audio/{dialogue_id}.flac             # clean dialogue audio (48 kHz stereo; overlaps preserved)
audio/{dialogue_id}_babble.flac      # dialogue + background speakers (mixed into the left/non-assistant channel)
audio/{dialogue_id}.timeline.json    # realized timeline: per-turn absolute start/end + overlaps_prev + turn text
annotations.jsonl                    # one dialogue per line (full annotation + participation profile)
README.md / README_zh.md

Annotation schema (annotations.jsonl)

Dialogue: dialogue_id, scenario, setting, topic, participants[], turns[], profile
  Participant: id, persona, style, speech_rate, state
  Turn:
    turn_id        turn index (0-based)
    speaker        user_01… / assistant / third_party_01… (third party = non-member)
    text           what was actually said (interrupted turns keep the cut-off half-sentence)
    addressee      who it is addressed to: user_01 / assistant / third_party / [multiple]
    gap_before_ms  gap to the end of the previous turn (negative = earlier start → overlap)
    kind           speech | interruption | backchannel | side_talk
    insert_ratio   backchannel only: position inside the previous turn (0–1)
    responds_to    which turns this turn responds to / summarizes
    events         non-speech sound events (doorbell, phone ring …)
    trigger        assistant only: cue | volunteer | invited | coordinate
  profile          participation profile: intensity + plan + realized distribution

timeline.json gives the realized timing (actual placement in the synthesized audio), complementing the intent encoded by gap_before_ms.

Quick start

import json, soundfile as sf

dia = json.loads(open("annotations.jsonl", encoding="utf-8").readline())

clean, sr  = sf.read(f"audio/{dia['dialogue_id']}.flac")           # 100% dialogue
babble, _  = sf.read(f"audio/{dia['dialogue_id']}_babble.flac")    # + background speakers

for t in dia["turns"]:
    print(t["turn_id"], t["speaker"], t["kind"], t["gap_before_ms"], t["text"][:20])

Typical uses: full-duplex speech model training · turn-taking & interruption prediction · backchannel prediction · "should the assistant speak?" classification (labels = addressee + kind + trigger) · speech activity & overlap detection · learning to ignore background speakers (clean-vs-babble pairs give a controlled noise/speaker-distractor setup).

How it was produced (transparency)

• Dialogue text: designed by an LLM per scenario and participation profile (includes interrupted half-sentences, backchannels, third-party talk, sound events).
• Speech: MOSS-TTS-Nano (Apache 2.0), synthesized turn by turn → VAD silence trimming/compaction → loudness normalization (per-turn ±1.5 dB jitter; backchannels 3 dB below the turn they follow) → 8 ms fade-in/out per turn → placed on the annotated timeline with stereo split (left = non-assistant, right = assistant); negative gaps create real overlap.
• Background-speaker variants: recorded human babble (see attribution below) is stitched (15 s clips, 50 ms crossfades, per-clip loudness normalized), randomly matched per dialogue (room + clips), and mixed only into the non-assistant channel at a target SNR of 10 dB.
• This dataset is synthetic speech (not human recordings); each speaker uses a fixed voice.

Background speakers — attribution (required)

The _babble variants contain audio from the VOiCES corpus (Voices Obscured in Complex Environmental Settings, SRI International / Lab41, In-Q-Tel; Interspeech 2018), babble subset, distributed under CC BY 4.0. Our modifications: 15 s clips stitched with crossfades, per-clip loudness normalization, random room/clip matching per dialogue, mixed into the non-assistant channel at 10 dB SNR. The background speakers are English-speaking (VOiCES recordings); the dialogue is Chinese. Please keep this attribution when redistributing the _babble files.

Known limitations

• Speech comes from a relatively small TTS model; audio quality is below human recordings.
• Background babble is English-speaking while the dialogue is Chinese — models trained on the _babble variants could learn a "language = background" shortcut (kept deliberately as a controlled distractor; use the clean variants if this is a concern).
• The assistant is never interrupted in this release (no barge-in against the AI yet).
• Scenario and phrasing diversity are bounded by the generating model.

License

CC BY 4.0 — free for commercial and non-commercial use with attribution. (Synthesized audio generated with Apache-2.0-licensed MOSS-TTS-Nano; background-speaker variants include VOiCES babble audio, CC BY 4.0 — see attribution above.)

Citation

@misc{chorusduplex_sidetalk,
  title  = {ChorusDuplex-SideTalk: A Multi-Party Duplex Dialogue Speech Corpus},
  author = {Dawei Yang},
  year   = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/BiggestWave/ChorusDuplex-SideTalk}}
}