Planet Money Podcast Generator - AWS AI Challenge 2018

Planet Money Podcast Generator - AWS AI Challenge 2018

Hello, my name's Nathan Glover, and my partner is Steven Maude. We're excited to showcase our Planet Money podcast generator, developed for the AWS Artificial Intelligence Challenge. This project, while having an interesting focus on elephants, involves a unique and complex architecture. Let me explain how we built and executed it, step by step.

Architecture Setup

The first part of our architecture involved creating a stage and making an instance. Here's a detailed breakdown of the process:

1. Transcripts Scraping: We set up a script to scrape every single transcript of Planet Money episodes. This was because, while we used AWS Transcribe for some episodes, we couldn't afford to rely entirely on it for all the analysis. Hence, we scraped transcripts directly from the Planet Money website, allowing us to perform in-depth analysis across numerous episodes.

2. Text Analysis: We then fed the scraped transcripts into a text analysis setup. Using textgenrnn, a neural network-based text generator, we input a corpus of words and output a trained model. Below is a glimpse of our training process:

   

   Every ten epochs, we output some variety of generated sentences to examine whether our training model was underfitting or overfitting. The goal was to strike a balance where generated content remained funny, interesting, and witty, without merely replicating the original text corpus.

Training Process

We ran the training across a small corpus of words, closely monitoring if the model achieved the sweet spot in text generation. Below is a snippet of our text generation during training:

Sample output text from training...

Podcast Generation

Post-training, we utilized AWS Polly to synthesize speech for our generated scripts. By passing the text and the desired voice ID, we concatenated the speech files into one cohesive podcast. Here's a bit of the generated podcast for demonstration:

"... a classic example of the oil curse is a fascinating tale..."

Our infrastructure is platform-agnostic, meaning you can re-run our cloud-based scripts to set it up yourself.

Demo Audio

Enjoy more demo audio from our generated podcast, which dives into complexities of topics like the oil curse:

"... you can't have what you want then you have this commodity which the whole world wants to buy oil from Norway ..."

You can experience firsthand the power and potential of our podcast generator, which crafts engaging narratives from scratch.

Keywords

• AWS Artificial Intelligence Challenge
• Planet Money
• Podcast Generator
• Textgenrnn
• AWS Polly
• Transcripts Scraping
• Text Analysis
• Synthesize Speech

FAQ

Q1: What is the main focus of the podcast generator? A1: The main focus is on creating engaging and witty podcasts using transcript data from Planet Money, enhanced with AI technologies like textgenrnn for text generation and AWS Polly for speech synthesis.

Q2: How do you handle the transcripts for analysis? A2: We scrape transcripts of Planet Money episodes directly from their website and also use AWS Transcribe for some episodes to ensure comprehensive text analysis.

Q3: What AI technologies are used in this project? A3: We use textgenrnn for text generation and AWS Polly for converting text into speech to create the podcast audio.

Q4: Can I replicate this project on my platform? A4: Yes, our solution is platform-agnostic, and you can run the cloud-based scripts on your own infrastructure to generate similar results.