MultiPartyTurnBench is a controlled benchmark for multi-party conversational floor control. It contains 240 synthetic, channel-separated conversations with 2–5 speakers and six interaction types, each with an exact, role-relative event ground truth labels. The log records who holds the floor, when a speaker yields, pauses or is cut off, and which listener speech successfully takes the floor or leaves it unchanged. Two causal tasks are defined on this log:

End-of-Turn (EOT): detect when the current floor holder has yielded the floor (positives), and do not fire during mid-turn hold pauses after which the same speaker resumes (hard negatives).
Interruption (INT): detect when a listener successfully takes the floor while another speaker holds it (positives), and do not fire on backchannels or laughter that leave the floor unchanged (hard negatives). Failed floor claims are excluded from scoring.

This repository contains the complete benchmark data: per-speaker audio (noisy evaluation input and clean reference), ground-truth labels, derived scoring anchors, the dev/test split and all metadata needed for evaluation. The evaluation code (scorer and operating-point selection) is released separately on GitHub.

1. Quick start

Download.

from huggingface_hub import snapshot_download
root = snapshot_download("puneetUMD/MultiPartyTurnBench", repo_type="dataset")

or huggingface-cli download puneetUMD/MultiPartyTurnBench --repo-type dataset --local-dir MultiPartyTurnBench.

Load a conversation for evaluation.

import json, os, soundfile as sf
dev = [json.loads(l) for l in open(os.path.join(root, "dev/manifest.jsonl"))]
c = dev[0]
events  = json.load(open(os.path.join(root, c["events"])))               # ground-truth event log
anchors = json.load(open(os.path.join(root, c["anchors"])))["anchors"]   # scoring anchors per speaker
audio   = {spk: sf.read(os.path.join(root, p))[0] for spk, p in c["audio_noisy"].items()}  # 24 kHz mono per speaker

Browse. The dataset viewer on this page shows every conversation with a mixdown preview and its metadata (load_dataset("puneetUMD/MultiPartyTurnBench") loads the same preview table; it is not the evaluation input). For a richer offline view, download the repository and open viewer/index.html in a recent browser. It shows a clickable per-speaker timeline coloured by ground-truth label, yield and interruption markers, hold pauses, the floor-holder track, per-speaker clean and noisy channels and the transcript synchronized to playback.

Score a system. With the GitHub evaluation code (Python ≥ 3.8, standard library only):

python eval/score.py score  --root MultiPartyTurnBench --split test --submission my_test_predictions.json
python eval/score.py select --root MultiPartyTurnBench --split dev  --submission 0.5=dev_0.5.json 0.7=dev_0.7.json

select applies the benchmark's operating-point rule: the highest recall with dev FPR ≤ 0.10, chosen per track. Freeze the selected threshold and run score on test. A system is within budget if dev FPR ≤ 0.10 and test FPR ≤ 0.15.

2. Contents

MultiPartyTurnBench/
├── README.md                     this dataset card
├── DATASHEET.md                  datasheet for datasets
├── LICENSE.md                    licenses and third-party notices
├── SHA256SUMS                    checksums of every file
├── dev/                          development split (48 conversations; tune thresholds here only)
│   ├── manifest.jsonl            one line per conversation with root-relative paths to all its files
│   ├── conversations.csv         per-conversation metadata for this split
│   ├── audio/noisy/<id>_speaker_<k>.flac   EVALUATION INPUT: per-speaker channels with the shared acoustic scene
│   ├── audio/clean/<id>_speaker_<k>.flac   same channels before the noise stage, for inspection
│   └── annotations/
│       ├── events/<id>.json      ground-truth event log + conversation metadata (canonical gold)
│       ├── anchors/<id>.json     EOT/INT positives, hard-negative spans and excluded spans per speaker (derived)
│       ├── scripts/<id>.json     tagged dialogue script (text, event type, timing hint) produced in Stage 1
│       └── render/<id>.json      placement of every utterance on the timeline (start/end, truncation) + noise manifest
├── test/                         test split (192 conversations), same structure as dev/
├── metadata/
│   ├── split.json                48 dev / 192 test conversation ids
│   ├── conversations.csv         all 240 conversations: split, N, type, domain, tone, duration, voices, noise, event counts, QC
│   ├── voices.csv                24 synthetic voices (Kokoro voice, gender, style, dev/test pool)
│   ├── dataset_stats.json        corpus statistics overall, by N and by interaction type
│   ├── qc_report.json            automatic acoustic quality-control results per conversation and channel
│   └── noise_seeds.json          noise-stage seed per conversation (reproduces audio/noisy from audio/clean)
├── generation/
│   ├── stage1_system_prompt.txt          verbatim LLM system prompt
│   ├── stage1_user_prompt_template.txt   user prompt template and repair message
│   └── generation_config.json            all generation parameters (Stages 1–5)
├── data/{dev,test}.parquet       dataset-viewer preview: metadata, mixdown preview audio and the event log per conversation
└── viewer/
    ├── index.html                offline dataset viewer
    ├── data.js                   viewer data (derived from the annotations)
    └── mix/<id>.mp3              64 kbps mono preview mixdowns of the noisy channels (not an evaluation input)

Conversation ids are dev_00001–dev_00048 and test_00001–test_00192. Speakers are speaker_1 … speaker_N. All times are in seconds on the shared conversation timeline. All channels of a conversation are sample-aligned and have the same length. Audio is 24 kHz, mono, 16-bit FLAC (1,680 files: 840 noisy + 840 clean). The viewer and the paper figures label speakers S1…SN. Paths in manifest.jsonl and in the preview table are relative to the dataset root.

3. Composition

	Conv.	Hours	EOT + / −	INT + / −	Failed claims (excl.)
dev	48	1.42	706 / 55	86 / 309	—
test	192	5.47	2,716 / 186	281 / 1,172	—
total	240	6.89	3,422 / 241	367 / 1,481	605

N	Conv.	Hours	Mean turn (s)	Events/min	BC/min	INT/min	Overlap / dialogue (s)
2	60	1.90	7.07	13.8	3.12	0.70	10.2
3	60	1.66	6.38	16.3	3.61	0.96	10.8
4	60	1.67	6.10	16.3	3.60	0.94	10.9
5	60	1.66	6.26	16.6	4.06	0.97	11.0

Every (N, interaction type) cell has exactly 10 conversations: 2 dev and 8 test.
The interaction types are casual, task-oriented, instructional, collaborative, argumentative and narrative.
The six annotated event classes total 6,116 events: 3,422 yields, 241 holds, 367 interruptions, 605 failed claims, 1,176 backchannels and 305 laughter vocalizations.
In addition, 367 holder turns are cut off by interruptions.
Conversations last 48–208 s (median 100 s).
Development and test use disjoint voice pools (8 and 16 voices, gender-balanced), and no voice repeats within a conversation.

4. Annotation format

4.1 Event log (`<split>/annotations/events/<id>.json`)

{"conversation_id": "test_00013", "conversation_type": "narrative", "num_speakers": 5,
 "speakers": ["speaker_1", ..., "speaker_5"], "duration_s": 157.529,
 "audio": {"mode": "channel_separated", "files": {"speaker_1": "test_00013_speaker_1.flac", ...}},
 "metadata": {"domain": ..., "voices": {...}, "voice_styles": {...}, "noise_profile": {...},
              "sample_rate": 24000, "tone": ..., "generator_model": "gpt-4.1", "engine": "kokoro"},
 "events": [
   {"id": "e004", "speaker": "speaker_1", "label": "TURN", "onset": 0.0, "offset": 19.19,
    "current_holder": "speaker_1", "floor_outcome": "yield", "next_holder": "speaker_2"},
   {"id": "e001", "speaker": "speaker_2", "label": "VOCALIZE_NO_CLAIM", "sublabel": "BACKCHANNEL",
    "onset": 7.965, "offset": 9.175, "current_holder": "speaker_1"},
   {"id": "e003", "speaker": "speaker_4", "label": "CLAIM_FLOOR", "resolved_as": "NON_FLOOR_TAKING_INTERRUPTION",
    "onset": 14.88, "offset": 16.21, "current_holder": "speaker_1"},
   ...]}

`label`	qualifier	meaning	role
`TURN`	`floor_outcome: yield` (+ `next_holder`)	holder releases the floor (or the conversation ends)	EOT positive at `offset`
`TURN`	`floor_outcome: hold`	holder pauses mid-turn and resumes	EOT hard negative over [`offset`, resume onset]
`TURN`	`floor_outcome: interrupted`	holder is cut off by a successful interruption	EOT excluded (±0.25 s)
`CLAIM_FLOOR`	`resolved_as: INTERRUPTION` (+ `floor_transfer_time`)	listener takes the floor	INT positive at `onset`
`CLAIM_FLOOR`	`resolved_as: NON_FLOOR_TAKING_INTERRUPTION`	listener tries to take the floor and fails	INT excluded over [`onset`, `offset`]
`VOCALIZE_NO_CLAIM`	`sublabel: BACKCHANNEL` / `LAUGHTER`	listener vocalization without a floor claim	INT hard negative over [`onset`, `offset`]

Acoustic nuisance (ambient bed, sound effects, distant third-party speech) is described in metadata.noise_profile and the render noise manifest. It never creates anchors.

4.2 Anchors (`<split>/annotations/anchors/<id>.json`)

For every speaker channel and track, the file lists positives (anchor times), negative_spans and excluded_spans. It is derived from the event log by the anchor-derivation module of the evaluation code (mptb_eval/anchors.py), and the scorer uses exactly these anchors.

4.3 Script and render

<split>/annotations/scripts/<id>.json holds the Stage-1 dialogue: each utterance's speaker, text, event_type, timing_hint and, where applicable, target, resolved_hint and sublabel. It also records the generation targets and tone.
<split>/annotations/render/<id>.json lists every placed clip: speaker, start_s, end_s, kind (holder / claim / backchannel / laughter), overlaps, truncated, and src_index (the index into the script). It also holds the realized noise manifest (noise).
Each clip is one script utterance, so script text can be aligned to audio through src_index.

5. Scoring protocol

Submission. For every conversation and speaker channel, a system submits causal decision times for each track. Systems that output scores are converted to decisions by taking rising edges above θ with a 2 s refractory period.
Matching (mptb_eval/scorer.py in the evaluation code, following TurnBench):
- Predictions inside excluded spans are dropped.
- Each gold positive at time t is matched to the earliest unclaimed prediction in [t − 0.25, t + 3.0] s, counting as a TP if found and an FN otherwise.
- Each hard-negative span counts as one FP if any remaining unclaimed prediction falls inside it, and one TN otherwise.
- Signed latency is t_pred − t_gold.
Metrics. Recall = TP/(TP+FN) and FPR = FP/(FP+TN), pooled over all speakers and conversations of a split.
Operating point. θ maximizes dev recall subject to dev FPR ≤ 0.10, chosen per track, and is frozen for test. A test FPR above 0.15 means the system is out of budget.

6. Intended use and limitations

Intended use: diagnostic evaluation of causal end-of-turn and interruption detectors and of full-duplex agents in controlled multi-party conditions, with breakdowns by N and interaction type.
Not intended for: estimating error rates on natural conversation, certifying deployment safety, or speaker identification. All speech is synthetic and all participants are fictional.
Limitations:
- The corpus is English-only.
- The prosody, voice timbre and disfluency of a single TTS model do not cover natural variation.
- Timing comes from parametric distributions.
- Hold pauses are inserted silences of 0.4–1.2 s.
- Dialogue content reflects the generating LLM.

See DATASHEET.md for details.

7. License and citation

Data: CC BY 4.0. Viewer code: MIT. Third-party notices (Kokoro-82M, MUSAN, gpt-4.1, TurnBench) are in LICENSE.md.
If you use MultiPartyTurnBench, please cite:

@misc{multipartyturnbench2026,
  title  = {MultiPartyTurnBench: Benchmarking Floor Control in Multi-Party Spoken Dialogues},
  author = {Puneet Mathur, Dinesh Manocha},
  year   = {2026},
  note   = {Dataset: https://huggingface.co/datasets/puneetUMD/MultiPartyTurnBench}
}

Downloads last month: 5

Total file size:

2.46 GB

You need to agree to share your contact information to access this dataset

MultiPartyTurnBench (MPTB)