🎭 ChuckleNet — Laughter Detection in Stand-Up Comedy

Find the funny moments. ChuckleNet takes any stand-up comedy audio and returns timestamps of laughter events — audience reactions, performer laughs, and the beats between punchlines that get the biggest response.

Built on a frozen microsoft/wavlm-base encoder + a 230K parameter MLP head. Drop-in via Transformers pipeline or use model.detect(audio, sr) for timestamped events.

⚠️ Labels are transcript-markers ([laughter] cues from VTT files), not human-audited gold labels. This is explicitly a weak-label evaluation. On a 118-video human-verified subset the same architecture reaches IoU-F1 0.30–0.40. See the ChuckleNet repo for the full pre-registered experiment log.

⚡ Quick Start

1-line pipeline (easiest)

from transformers import pipeline

pipe = pipeline("audio-classification", model="Hayasuki/chucklenet-laughter-detector", trust_remote_code=True)
# Done — pipe now has a .detect() method

model.detect() — timestamps directly

import librosa
from transformers import AutoModelForAudioClassification

model = AutoModelForAudioClassification.from_pretrained(
    "Hayasuki/chucklenet-laughter-detector",
    trust_remote_code=True
)

audio, sr = librosa.load("comedy_special.mp3", sr=None)   # any sr, mono
events = model.detect(audio, sr=sr, threshold=0.5)        # lower threshold = more recalls

for ev in events:
    print(f"{ev['start']:6.2f}s → {ev['end']:6.2f}s  confidence={ev['p']:.2f}")

Output looks like:

  1.23s →  4.51s  confidence=0.91   ← big audience laugh
 12.07s → 13.82s  confidence=0.87
 18.40s → 19.15s  confidence=0.54   ← quiet chuckle

detect() runs energy VAD → segment padding/splitting → batch scoring → merge adjacent detections (gap ≤ 0.3s). Default threshold=0.85 is precision-oriented; drop to 0.5 for recall-heavy use.

📊 Evaluation Results

Weak-label eval (124 held-out videos, video-disjoint)

Trained on 620 videos · 121,928 utterance segments · 1,965 positives (1.6%)

Human-verified subset (118 videos, pre-registered)

On a subset with word-level human-verified labels, the same architecture reaches IoU-F1 0.30–0.40 and beats order-shuffled acoustic controls by +0.09 F1 — demonstrating the model captures real temporal interaction signal, not just static acoustics.

What does this beat?

• Order-shuffled baseline: +0.09 F1 improvement over shuffled controls (pre-registered, p<0.05)
• Acoustic-only controls: Survives breath/cough/applause/cheer controls (mean FPR=0.189 < 0.2, E01 SUPPORTED)

🔬 Scientific Context

ChuckleNet is a pre-registered research program testing whether interaction behavior carries information beyond what words alone provide.

See the full pre-registered log for experiment details.

🏗️ Architecture

🎯 Intended Use

Great for:

• Finding the funniest moments / laugh lines in comedy specials
• Ranking segments by laughter intensity for highlight reels
• Reaction-event research in performed monologue audio
• Building "best-of" compilations automatically

Not verified for:

• Multi-party conversation (meetings, podcasts)
• Phone/call audio
• Non-English speech
• General emotion classification (see E05 above — behavior ≠ emotion)

📦 Files

🔗 Try It Now

📖 Citation

@misc{das2026chucklenet,
  author       = {Subhajit Das},
  title        = {ChuckleNet: Temporal Interaction Signals for Laughter Detection in Stand-Up Comedy},
  year         = {2026},
  url          = {https://github.com/Das-rebel/ChuckleNet},
  note         = {Preprint forthcoming}
}