TEDify

TEDify is an NLP model for real-time audience feedback (laughter and applause), trained on audience reactions from thousands of TED Talks.

Demo (YouTube)

Installation & Usage

After cloning the repository, optionally create a new virtual environment and install the required dependencies.

pip install -r requirements.txt

You'll need Google Cloud credentials in the form of a JSON file for the speech-to-text to function. Then, export the location of that file as an environment variable.

export GOOGLE_APPLICATION_CREDENTIALS="path/to/credentials.json"

At this point, you should be able to run TEDify:

python tedify.py

To decrease the verbosity, add the argument:

python tedify.py --noverbose

TEDify's NLP encodes tokens to integers to be passed into the Transformer. These encodings are found in word_to_index.json. To use a different set of Transformer and encodings, pass in the following parameters:

python tedify.py --encoding="path/to/encoding.json" --transformer="path/to/transformer"

Good luck with your TED Talks!

Project structure

• sounds/ is a directory containing audience sound effects
• transformer/ contains the saved Transformer for text processing
• README.md is what you're looking at right now
• play_audience_sounds.py is a utility file for playing crowd audio
• requirements.txt is a list of required Python libraries
• ted_scraped.txt is a file containing 200 scraped TED Talks– note that the final Transformer is actually trained on the TED Ultimate Dataset
• ted_scraper.py is a Python script for scraping TED Talks using Selenium– see above note regarding training data
• tedify.py is the main file for running TEDify
• tedify_training_pipeline.ipynb is a Jupyter notebook for cleaning the dataset and training a Transformer
• tedify_util.py is a utility file for tedify.py that performs text cleaning and classification
• word_to_index.json is a JSON file containing mappings of tokens to indexes [1, 20001]