How-to seriesNovember 18, 2026
How to turn a PDF into a two-voice podcast
Have an agent read the PDF, write a two-host script you can edit, voice each line with two voices and join them into one MP3. Free options, steps and cost.
To turn a PDF into a two-voice podcast, have an agent read the PDF and write a short two-host script, edit the script yourself, voice each host's lines with a different voice, and join the lines into one MP3. With a coding agent and a Vaaya wallet, an 8-minute episode costs well under a dollar in voice calls.
What people usually try, and where it falls short
- NotebookLM's Audio Overview. Free, fast and good. You can steer it with instructions, but you cannot edit the script line by line or pick the voices.
- A text-to-speech reader. Apps that read the PDF aloud give you one voice reading every word, including tables and footnotes.
- Podcast tools in voice apps. Some, like ElevenLabs, have a PDF-to-podcast feature inside their own app, tied to that app's plans and voices.
- Recording it yourself. The best result if you have the time and a quiet room.
The gap is control: a script you can correct before anyone hears it, voices you chose, and a length you set.
How to do it
- Get the text. A coding agent such as Claude Code can read a local PDF on its own. For a PDF at a public URL, firecrawl/scrape with
parsePDFreturns the text for 1¢. - Write the script. Ask for two hosts with clear roles: one explains, one asks the questions a listener would ask. Set a word count: about 150 words per minute.
- Edit the script. Check every number and claim against the PDF. Write numbers and abbreviations the way you would say them, such as "twelve percent" and "Q3, the third quarter".
- Voice each line. deepgram/speak turns up to 2,000 characters into an MP3 per call, at 1¢ per 250 characters with a 2¢ minimum. Use one voice per host, for example
aura-2-thalia-enandaura-2-orion-en. - Join the lines. The agent downloads each MP3 and joins them in order with ffmpeg.
- Listen to the whole episode before you share it.
For warmer voices, fal/generate with the elevenlabs--tts--turbo-v2-5 model costs 5¢ per started 1,000 characters, with a 5¢ minimum per call. ElevenLabs' Text to Dialogue endpoint, reached through monid/run-metered, makes several speakers in one call and is billed on usage. Ask consult for a quote on your script first.
Make it sound like a conversation
Short turns sound natural. Mix one-line questions with longer answers of three or four sentences. Let the second host sum up a point in their own words now and then, since that is where listeners catch up. End with one clear takeaway and where to read the full PDF.
The prompts to copy
Run this in a coding agent with Vaaya connected, such as Claude Code, Cursor or Codex, in a folder with the PDF:
Read report.pdf. Write a two-host podcast script about it, about 1,200 words.
Host A (Maya) explains. Host B (Sam) asks short, curious questions a
listener would ask. Open with the single most useful finding. Use only facts
from the PDF, and after each line add the page number in brackets so I can
check it. Write numbers the way they are spoken. No "Exactly!" or "Great
point!" filler. Save it as script.md and stop so I can edit it.
After you edit the script:
Use Vaaya to voice script.md. Remove the page-number brackets first.
For each line, call deepgram/speak: Maya uses voice "aura-2-thalia-en",
Sam uses "aura-2-orion-en". Keep each call under 2,000 characters.
Download each MP3 as line-001.mp3, line-002.mp3 and so on.
Then join them in order with ffmpeg into episode.mp3.
Before you start, tell me the total price. Spend at most $1.
One line on its own:
{
"service": "deepgram",
"action": "speak",
"params": {
"text": "So the headline is that costs fell twelve percent. What drove that?",
"voice": "aura-2-orion-en"
},
"max_cost_cents": 2
}
The join step runs on your machine and costs nothing:
for f in line-*.mp3; do echo "file '$f'"; done > list.txt
ffmpeg -f concat -safe 0 -i list.txt -c copy episode.mp3
What it costs
Prices as listed in the Vaaya catalog on 18 November 2026:
| Step | Tool | Price |
|---|---|---|
| Read a PDF from a URL | firecrawl/scrape |
1¢ |
| Write the script | your agent's own model | no Vaaya charge |
| Voice a line | deepgram/speak |
1¢ per 250 characters, 2¢ minimum |
| Premium voice | fal/generate, ElevenLabs Turbo v2.5 |
5¢ per started 1,000 characters |
| Join the audio | ffmpeg on your machine | free |
A 1,200-word script is about 7,000 characters. Split into 40 short lines, each line costs the 2¢ minimum, so about 80¢. Split into 20 longer turns, about 40¢. The same 40 lines with ElevenLabs cost about $2. Failed calls are not charged, and the agent's key limit caps a runaway loop.
Common mistakes
- Letting the script run long. A 20-page report does not need 20 minutes. Pick the three findings that matter.
- Two hosts who only agree. Give the second host real questions and one point of doubt.
- Reading tables aloud. Turn a table into one spoken sentence about what it shows.
- Skipping the fact check. A model can blur numbers. The page numbers make the check quick.
- Putting both hosts in one call. Each call is one voice. Split by speaker.
- Using content you cannot share. Check the PDF's rights before you publish the episode.
To narrate a single script in one voice, read how to create a voiceover for a script. To make a voiceover in your own voice, see how to make a voiceover in your own voice.
Questions
What is the easiest free way to turn a PDF into a podcast?
NotebookLM's Audio Overview. Upload the PDF and it makes a two-host conversation. Use the method in this post when you want to edit the script, choose the voices, or keep the episode short and on topic.
How long should the podcast be?
Five to ten minutes suits most reports. People speak about 150 words a minute, so an 8-minute episode is a script of about 1,200 words.
How much does an 8-minute two-voice episode cost?
With Deepgram voices on Vaaya, about 40¢ to 80¢ for the audio, depending on how many lines the script has. Reading a PDF from a URL adds 1¢. ElevenLabs voices cost more per line.
Can I publish the podcast?
Only if you have the rights to the PDF's content, and label the voices as AI-generated where the platform asks. Do not make the hosts sound like real people.