unstoPEPPAble
Jan Pokorný
About
All good things have an end. Thankfully, computer-generated Peppa Pig cartoons are not a good thing, and also have no end. They make just enough sense to be uncanny, and just enough nonsense to be on the edge between 'dark' and 'hilarious'. And what is more, you cannot stop them — they're unstoPEPPAble. A text-generating neural network (trained on the whole of the Internet), a state-of-the-art set of TTS voices (meant to help the visually impaired) and a bunch of video and image editing tools are misused with one goal in mind: destroying your childhood.
Video
Technology
Python: Programming language that does nothing perfectly, but it does everything.
GPT-2: Transformer-type text-generating neural network from OpenAI, known for powering the game AI Dungeon. This project uses the 355M size, which is 1/4 of the full network. If you are in posession of a GPU that can train the larger sizes, definitely hit me up.
gpt_2_simple: Python library with a simple interface to train and infer from the GPT-2
Google Colaboratory: Cloud Jupyter instance with access to a powerful machine-learning GPU (powerful enough for the 355M model at least, which requires over 8 GB of VRAM)
requests: Python library for HTTP requests, used to scrape episode transcripts and character images from the Peppa Pig wiki
Microsoft TTS Voices: With a little encouragement (in the form of registry editing), Microsoft provides us with a very usable set of TTS voices
pyttsx3: Python library that uses system voices to read text
MoviePy: Python library for composition of videos, pictures, audio and text -- basically Windows Movie Maker but for Python
ImageMagick: Command-line utility for generating and modifying images
ffmpeg: Command-line utility for all types of video operations, used for voice pitch editing and the post-process effects