When Stand-up Comedy Meets Machine Learning
summary
Experimental pipeline downloads sets of stand-up videos with YouTube-dl, extracts audio and manually labels a couple of 6‑minute clips (15 videos collected total), chops audio into 1‑second windows shifted 100 ms, computes ~31 features (mean/std per feature), trains a multilayer perceptron (31 inputs, one hidden layer) to tag laughs vs clapping vs other, and exports subtitle-like markers, vertical bars and per‑minute laugh rates. Prototype finds 4–6 laughs per minute in popular clips with visual alignment but shows timing lag and needs more data.
AI-written summary · about
chapters (22)
- 0:00Introduction
- 1:40The goal of standup comedy
- 2:38The guideline to follow
- 3:48Agents
- 4:54Video Recording
- 5:15Intuition
- 5:39Poetry
- 5:51Segments
- 6:21Data Collection
- 8:03Download Video
- 9:30Label Data
- 11:18Audio Editor Library
- 13:03Labeling
- 14:04Feature Engineering
- 16:36Feature Extraction
- 18:15Modeling
- 18:27Multiclass problem
- 18:51Multiclass algorithm
- 20:22Visualizations
- 23:44Delay
- 28:23Unpopular
- 29:19Conclusion
related talks
33mWhat if a Robot Could Laugh at Your Jokes? (Laughter Prediction with Natural Language Processing)
20mCapturing, Representing, and Interacting with Laughter
01h 51mNatural Language Processing in Python
59mTechnical Workshop: Sarcastically Speaking: Unlocking Multi-modal Sentiment Analysis With NLP and Facial Expressions
13mTransfer Learning: Leveraging Pretrained Models with Limited Data
20mNo Need to Laugh Out Loud: Predicting Humor Appraisal of Comic Strips Based on Physiological ...
22mMachine learning in Education
21mFrom 0 to a crowd sourced neural network in a jiffy