How to turn your PDFs into a podcast with AI (and listen on the way)

Turning a set of PDFs into a podcast is the fastest way to go through material you have no time to read sitting down: a 200-page tender on the way to the office, the week's three papers on a run, the vendor manual before the meeting. The condition is that the episode comes from the documents; a model without the sources fills in with whatever it thinks it remembers about the topic.

What does an AI tool need to make a good podcast from your documents?

First, the script has to be written from the real content of the files. BrainBox retrieves passages from every selected document, labels each with the file it came from, and asks the model to have the conversation cover several sources, with an explicit instruction not to focus on one. That avoids the episode that spends ten minutes on the first PDF and mentions the others in passing.

Second, control over the form: which language, how many voices, how long and with what focus. A two-minute summary to decide whether something is worth reading and a twenty-minute episode to study for an exam are different products.

Third, the audio has to stay where the sources are and be queryable. In BrainBox the podcast is saved as a file in the Box, transcribed and indexed. You can ask "what did they say about the penalty clause?" and the answer cites the minute in the episode, with the original PDFs one click away to check.

BrainBox is an AI workspace for documents that answers with citations to the exact page. The podcast is one more output of that same Box: it comes from the same sources and leads back to them.

How do you generate the podcast, step by step?

1. Select the files in the Box

   Tick the PDFs, Word documents, slide decks or transcripts you want in the episode. One or several; the script is written across everything selected.

2. Open the podcast generator

   From the chat's quick actions choose Create Podcast, or just type "I want a podcast from these documents". In the second case the assistant asks how many speakers and which language before generating.

3. Choose speakers, language and length

   Number of speakers from 1 to 30 (2 is the default and sounds best), one of the 7 languages, and short, medium or long.

4. Add instructions if you want to steer the episode

   The custom instructions field takes things like "focus on the contract risks", "use the authors' names" or "explain it for someone who is not a lawyer".

5. Wait for it to finish

   The progress bar shows the stages: preparing content, generating script, generating audio, merging segments, uploading, indexing. A medium episode takes a few minutes.

6. Listen, move it or ask it questions

   The WAV shows up in your Private folder under _generated. It opens in the player with the synchronized transcript; you can move it to a shared folder in the Box so your team can listen.

What does asking from the chat look like?

If you prefer the chat to the panel, the assistant asks the two necessary questions and generates. It goes like this:

Once indexed, the podcast behaves like any audio file in the Box. Asking "at what minute do they discuss the disagreement?" returns the segment with its timestamp, just as with a recorded meeting.

00:05:12Voice 1

This is where the three studies stop agreeing. The second reports a large effect and the third, with a bigger sample, almost none.

00:05:24Voice 2

And is that explained by the design or by the population?

00:05:28Voice 1

By the population, according to the third study's own authors: they included patients the other two excluded.

What options and limits does the generator have?

What is a podcast of your own documents good for?

For studying, it is a way to review a topic without a screen: a long episode on a course's notes covers the whole material, and study guides can be turned into audio the same way.

For tenders and contracts, a short episode on the bid documents gets the whole team to the kickoff meeting with the same context; the mandatory requirements still need reading, but the conversation beforehand makes it faster.

For teams, a podcast of the monthly report in the shared folder is an alternative to the email nobody opens.

How is it different from NotebookLM?

NotebookLM also generates a two-voice audio summary from uploaded sources, and does it well. The practical differences are in control and in what happens afterwards.

The point that weighs most in professional work is the last row: the podcast stays as one more source in the project, with the same citations and permissions as the rest.

A long episode on five documents costs around 20 intelligence units and fits in a 20-minute commute.