PodcastGPT

Overview

PodcastGPT is a pipeline that takes an article URL and produces a ready-to-publish podcast episode. GPT-3.5 rewrites the article as a conversational podcast script, and Amazon Polly renders it as audio using distinct synthetic voices for different speakers. I built the initial version as a side project, then expanded it at a hackathon — adding a web UI for pasting article links and selecting voice configurations, and extending the multi-voice support so conversations feel natural rather than monolithic.

Key Features

• •GPT-3.5 prompt pipeline that rewrites articles as conversational podcast scripts
• •Amazon Polly TTS rendering with multiple distinct synthetic voices
• •Web UI for pasting article links and configuring voice/speaker assignments
• •Multi-voice conversation flow making episodes feel natural and engaging

Challenges

Prompt engineering the GPT-3.5 rewrite to produce scripts that sound genuinely conversational — not like a narrated article — required significant iteration. Mapping speaker turns to Polly voices and keeping the audio stitching seamless was the other core challenge.

Outcomes

Shipped a working demo at the hackathon. The project attracted interest from other participants as a practical demonstration of LLM + TTS pipelines.