🎭 ChuckleNet

Prosody-enhanced laughter detection for stand-up comedy audio.

audio-classificationWavLM + ProsodyMIT

📊 Benchmarks (5-fold CV, 221 videos)

VariantF1
Prosody-only0.960
Fusion (this model)0.879
WavLM-only0.220

Key insight: prosody alone beats deep speech embeddings for laughter detection.

🧠 Architecture

791-dim (768 WavLM + 23 prosody) → 512 → 256 → 64 → 1 sigmoid
Binary laughter probability per 5-second audio chunk.

🚀 Get Started

Model: Hayasuki/chuckleNet-v2
Code: github.com/Das-rebel/ChuckleNet

from transformers import AutoModel
model = AutoModel.from_pretrained("Hayasuki/chuckleNet-v2", trust_remote_code=True)

Built by Das-rebel · part of the Standup4AI project