ChuckleNet: Real-Time Laughter Detection

Simple spectral features (20-dim) outperform massive transformer embeddings (WavLM 768-dim)

• F1=0.952 @ IoU=0.4 on StandUp4AI test set (32 videos, 854 segments)
• vs baseline F1=0.51 from StandUp4AI (EMNLP 2025)
• +87% improvement using RMS, ZCR, spectral centroid/bandwidth/rolloff/flatness + MFCCs

Results

Method	F1 Score	Dataset
Our Spectral (20-dim)	0.952	StandUp4AI test (32 videos)
StandUp4AI baseline	0.51	StandUp4AI val (reported)
Gillick et al. (Interspeech 2021)	0.75	Switchboard
Truong et al.	0.85	TV comedy

Key Insight

"5 dimensions of pitch (F0) beats 768 dimensions of WavLM by 4.3x"

Hand-crafted prosody features capture laughter's acoustic signature (rhythmic excitation, voiced+unvoiced transitions) more effectively than learned embeddings for this task.

Notebooks

Notebook	Purpose	Link
StandUp4AI Eval (IoU)	Our F1=0.952 result	Colab
StandUp4AI Fixed	Segment-level F1=0.935	Colab

Paper

📄 Pitch-Perfect: Hand-Crafted Prosody Features Outperform Deep Audio Embeddings

Dataset

• 32 StandUp4AI test videos with risa/no_risa labels (854 segments, 86% positive)
• Val labels from StandUp4AI authors are held out (not publicly available)

Key Files

• docs/paper_f0_breakthrough.md - Main paper draft
• docs/DEFINITIVE_PLAN.md - Project plan
• docs/LITERATURE_REVIEW_LAUGHTER_DETECTION.md - Literature benchmarks