TIC-TALK — Timing In stand-up Comedy: Text, Audio, Laughter, Kinesics

Time-aligned text, audio, and vision features from 90 stand-up specials (2015–2024). Show identities are anonymized (SHOW_XXXX); no source audio, video, or transcripts are distributed.

Dataset contains:

Shows	90
Total runtime	≈ 94 h (mean 63 min / show)
Rows (60 s blocks)	5 416
Topics (BERTopic)	24
Video frames (1 fps)	322 973 (22 % full-body)
Laughter coverage (mean)	17.8 % of block duration, ≈ 1.2 events / 10 s
Embedding dim	384
File	single Parquet, 44 MB (zstd)

Loading

Single split named train

from datasets import load_dataset
ds = load_dataset("ENC-PSL/TIC-TALK", split="train")

Schema (1 row = 1 block)

field	type	description
show_id	string	SHOW_0001…SHOW_0090
block_id	int32	block index within the show
start, end	float32	block window, seconds from show start
topic_id	int32	BERTopic id (24 topics, -1 = outlier)
embedding	float32[384]	sentence-BERT embedding of block text
laugh_events	list<struct>	start, end, type, confidence
pose_keypoints	list<struct>	time, has_detection, bbox_*, keypoints[17][2] (COCO-17 order, [x,y] in pixels)
shot_events	list<struct>	time, label, class_id, score
show_duration_sec, show_n_blocks	float32 / int32	show-level
show_has_laugh / _pose / _shots	bool	modality flags
topic_show_id	string	per-show topic model id

Shot labels: full_shot, medium_close-up, medium_long_shot, medium_shot, other_angles, other (6 classes, average F1 = 0.91).

All timestamps are in seconds on the same absolute axis (from show start). Block assignment is by event start: pose_keypoints[i].time, shot_events[i].time, laugh_events[i].start ∈ [start, end); a laugh_events[i].end may extend ≤ 0.8 s past end (Whisper-AT stride).

Example

from datasets import load_dataset
ds = load_dataset("ENC-PSL/TIC-TALK", split="train")

def pose_near(row, t, window=1.0):
    return [p for p in row["pose_keypoints"]
            if p["has_detection"] and abs(p["time"] - t) <= window]

for row in ds:
    for ev in row["laugh_events"]:
        poses_at_onset = pose_near(row, ev["start"])
        # ... features at laughter onset

Provenance

Derived features only (no source media is redistributed).

• Text / topics: subtitle-derived 60 s blocks, sentence-BERT sentence-transformers/all-MiniLM-L6-v2 (384-dim), BERTopic with UMAP (n_neighbors=15, n_components=5, min_dist=0.1) + HDBSCAN (min_cluster_size=15, min_samples=5), 24 topics.
• Laughter: Whisper-AT (AudioSet audio tagging), 0.8 s stride, high-recall configuration; contiguous positive windows merged into events.
• Pose: YOLOv8s-pose (pretrained), 17-keypoint output in canonical COCO order, 1 fps, restricted to full-body / frontal frames.
• Shot: YOLOv8-cls fine-tuned on 594 manually annotated frames (val on 128 held-out frames, average F1 = 0.91), 1 fps.

License

CC-BY-NC-4.0 for research use. Underlying audiovisual material remains the property of its rights holders; only derived features are distributed.

How to cite

Paper accepted for the 2nd Workshop on Computational Humor (CHum 2026) @ACL 2026.

Preprint :

@misc{zribi2026timingstandupcomedytext,
      title={Timing In stand-up Comedy: Text, Audio, Laughter, Kinesics (TIC-TALK): Pipeline and Database for the Multimodal Study of Comedic Timing}, 
      author={Yaelle Zribi and Florian Cafiero and Vincent Lépinay and Chahan Vidal-Gorène},
      year={2026},
      eprint={2603.21803},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.21803}, 
}

Acknowledgements

This study was conducted as part of the DH master’s program at École nationale des chartes–PSL (Yaelle Zribi master thesis) and with the support of the PSL Research University’s Major Research Program CultureLab, implemented by the ANR (reference ANR-10-IDEX-0001).