← Back to home

How the Podcasts Are Generated

The podcasts here are AI-generated from episode stubs, source material, and prompt templates. The result is not a recording of real hosts in a studio. It is a pipeline that plans an episode, writes a transcript, turns each speaker line into audio, and publishes the final MP3 and RSS feed.

Each show has a folder with its own metadata and episode JSON files. A new episode usually starts as a small JSON stub with a topic, language, and episode type. From there, the transcript generator picks the matching prompt template, builds a sectioned plan, writes the intro, content sections, and conclusion, then saves the transcript and metadata back into the episode file.

Once the transcript exists, the audio step assigns synthetic voices to the speakers and generates the lines with text-to-speech. Speaker lines are processed in parallel where possible, but each speaker keeps their own request history so the voice remains more consistent across the episode. The generated clips are joined with pauses and music, normalized, encoded as an MP3, and given podcast metadata.

Finished episodes are uploaded to Cloudflare R2 and the RSS feed is regenerated so podcast apps can pick them up. The website syncs those episode files into the pages you see here, including the show pages, episode pages, transcripts, descriptions, durations, and feed links.

The whole thing is deliberately explicit about being synthetic. The episodes are useful experiments in summarizing, explaining, and discussing material, but they should be treated like AI output: interesting, sometimes helpful, and always worth fact-checking when the details matter.

Updates

Any complaints please let me know

url: https://vellori.cc/posts/podcast-generation/

← Back to home