# Turn an article into a two-host podcast

Entry E06. Sector: Online publisher. Status: in production. Delivered: 2026-04.
Source: https://ai.prospicience.in/work/article-to-two-host-podcast

Prospicience built an article to podcast tool for an online publisher. Paste an article and get a two host conversation script. An editor approves it, then two distinct AI voices record it into a finished MP3 in thirty to sixty seconds. The publisher gets audio at volume, with no studio, and every word signed off.

Speed: Built in a day.

## The challenge

The publisher wanted audio versions of its articles at volume, but recording each piece with a studio and hosts was too slow and costly to scale.

## Why it mattered

Audio reaches people who will not sit and read, so every article without an audio version is an audience missed. Fully automatic AI audio was not an option either: the editor had to stay in control of every word that went out under the publisher's name.

## What we built

- Built a pipeline that turns a pasted article into a two host conversation script.
- Made the script step an editorial gate: an editor reads and changes it before any audio exists.
- Recorded each line in two distinct voices and mixed them into a finished, saved MP3.

## The result

- A finished two host episode thirty to sixty seconds after the script is approved.
- Editorial control over every word, with approval required every time.
- No studio and no booked hosts, and a format that holds across subjects.

## Built for trust

Nothing is voiced until an editor has approved the script. The script step is the approval, not a preview, so the editor shapes the words while changes are still quick and cheap.

## AI at work

AI writes a two host conversation script from each article and, once an editor approves it, voices it in two distinct voices. The publisher gets audio versions at volume, with the editor shaping every word.

What it produces: two host conversation scripts, finished podcast MP3s, two distinct host voices, saved episodes ready to download again.

- A finished episode thirty to sixty seconds after the script is approved, with no studio and no booked hosts.
- The editor's time goes into words, where changes are quick, rather than into re-recording audio.
- The same two host format holds across subjects, so each new article starts from a proven shape.

How it works day to day: An editor pastes an article and AI drafts the script. The editor reads it, changes it and approves it, then AI records each line in two voices and mixes the finished MP3.

## Figures from the delivered system

- voices: 2
- time to make audio: 30 to 60 seconds
- editor approval: required every time
- output: MP3, saved

## The technology

- Paste an article, get a two host conversation script
- An editor reads and changes the script before any audio exists
- Two distinct voices record it line by line
- A finished MP3, in thirty to sixty seconds

AI models used: GPT-4o (OpenAI), Gemini 2.5 Pro (Google), Neural text to speech (Google Cloud), ElevenLabs.

### Technical notes

- Script generation, then a mandatory editorial gate, and only then audio synthesis as a separate step
- Two neural voices with distinct speaker identities, synthesised per line and mixed
- Local pipeline with ffmpeg mixing and output persisted to disk

## Questions

### Does a human review AI generated audio before it publishes?

Before it is even generated. The script step is the approval gate, so the editor works on words, where changes are quick, rather than correcting finished audio.

### How long does an episode take?

Thirty to sixty seconds from approved script to finished MP3. The editor's time goes into the script, which is where it counts.

### Can it produce languages other than English?

This build used two Indian English voices. The wider media system we run covers Hindi, Tamil, Telugu, Marathi, Bengali and Gujarati among others.

Capabilities: ai-media-generation, voice-and-multilingual-ai.