ChuckleNet — Chunk-Level Laughter Detector

Headline

ChuckleNet is a chunk-level binary laughter detector that classifies 5-second audio segments from comedic video content using WavLM audio embeddings fused with prosodic features.

Model Description

ChuckleNet combines a 768-dimensional WavLM embedding with 23-dimensional prosodic features (F0, energy, ZCR, spectral characteristics) into a 791-dimensional input vector, processed through a fully-connected classifier. Trained with class-weighted BCE loss on the StandUp4AI EMNLP subset of 221 comedic videos, achieving CV F1 = 0.879 on held-out video groups.

Intended Use

Primary:

• Detecting laughter events in stand-up comedy and similar comedic video content
• Indexing comedic video archives for laughter-rich segments
• Laughter density analysis across video corpora
• Highlight reel generation from high-laugh segments

Out-of-scope:

• Word-level or turn-level laughter timing (±2.5s precision only)
• Non-comedic content with different acoustic profiles
• Real-time streaming (validated on fixed 5-second chunks)

Training Data

Dataset: StandUp4AI EMNLP Subset — 221 comedic videos Audio: 5-second non-overlapping chunks Labels: Binary (laughter / no-laugh) per chunk Class distribution: ~30% positive, ~70% negative Features per chunk:

• WavLM-base embedding (768d): frame-level features averaged across chunk
• Prosodic features (23d): F0, RMS, ZCR, spectral centroid/bandwidth

Evaluation Results

Cross-validation: 5-fold GroupKFold grouped by video ID

Fold	Precision	Recall	F1
1	0.861	0.897	0.878
2	0.882	0.874	0.877
3	0.847	0.931	0.887
4	0.903	0.835	0.867
5	0.899	0.868	0.883
Mean	0.878	0.881	0.879
Std	—	—	0.022

Limitations

• Chunk-level precision: ±2.5s localization — word-level timing unavailable
• Domain bias: English stand-up comedy only; other genres/languages untested
• False positives: May fire on coughs, sighs, audience murmurs
• Short laughs: Brief (<1s) or soft laughs may be missed

Architecture

Input(791)
  → Linear(791 → 512) + BatchNorm + ReLU + Dropout(0.3)
  → Linear(512 → 256)  + BatchNorm + ReLU + Dropout(0.3)
  → Linear(256 → 64)   + BatchNorm + ReLU + Dropout(0.3)
  → Linear(64 → 1)     → raw logit

Training: BCEWithLogitsLoss(pos_weight=2.33), Adam(lr=1e-3), early stopping.

How to Use

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("subhajitdas/chucklenet")
# Extract 791-dim features per 5s chunk → model → laughter probability

Citation

@article{chucklenet,
  title={ChuckleNet: Chunk-Level Laughter Detection on Stand-Up Comedy Video},
  author={Das, Subhajit},
  year={2026}
}