Built by Nimo Rotem and Willy Arisky to demonstrate automatic content creation using artificial intelligence and neural networks.
How it works:
TL;DR: With one click of a button, PodBot.ai generates and records a podcast episode about your topic of interest. Completed with a generated host, researched but made up content, opening music summary and cover photo.
PodBot runs a modified (fine tuned) GPT-3.5 model (A neural network published by Open AI)
When given an "episode title" as input, Podbot first searches the web and finds relevant content from 4 existing articles.
These sentences are used as seed input for the GPT-3.5 neural network for inference of the full podcast script.
A fake name is generated for the host from a random database, and a picture of a person who doesn't exist is generated using styleGAN - (The Tensorflow implementation of a generative adversarial network which uses unsupervised learning to generate realistic looking photos of fake people who do not exist).
At the same time, PodBot searches google via API to find a relevant cover photo for the episode. It then uses Google's High-fidelity text to speech neural networks to convert the text into realistic sounding human voice.
To finish it up, PodBot selects a royalty free tune from an MP3 library and attempts to create a pleasant realistic sounding opening to the generated podcast.
* Note: In times of high server load or internal error the app may use fall backs such as previously pre-generated host photos, simpler GPT models and other 3rd party API based solutions for content creation via GPT-3.5