Figuring Out How Audio Overviews Are Structured

July 1, 2025 • Wietse Venema

I love the Audio Overviews in the Gemini app. These are podcast-style conversations based on documents. To start, upload a file in the Gemini app, and a button to generate the overview will appear.

I wondered how they were made and how to start creating my own using the audio support in the Gemini 2.5 models.

So, I took a few of my favorite overviews, transcribed them, and fed the texts to Gemini to see if it could uncover the underlying structure. The results revealed the archetype that all these audio overviews share. It's a deconstruction process anyone can follow; here's what I did.

Transcribing

I started with getting a written version of a few Audio Overviews. I downloaded them as MP3 files. Since Gemini doesn't accept that format, I used ffmpeg to convert them into MP4s (with a black screen).

As the next step, I prompted Gemini to "Convert this audio to a written transcript with speaker identification" and the results looked like this:

Speaker 1: It's everywhere, right? And uh, it can be pretty hard to cut through all the noise, you know. Today, we're doing a deep dive into something, well, genuinely groundbreaking, I think. Google Deep Mind's latest AI models, the Gemini 2.X family. Our mission here, really, is to unpack what's actually new, what these big advancements mean for AI's future, and, you know, why it matters to you.

Speaker 2: Yeah, exactly. And we've uh, we've really dug into Google's own report on Gemini 2.X here. It's, well, it's packed with some fascinating stuff about how they're pushing things forward. We're talking about understanding really complex data, interacting with the world, just new ways of doing things.

Finding the Underlying Structure

Once I had a few transcripts, I asked Gemini in AI Studio to analyze the text and identify the structure.

What's the archetype of these conversations? I want to generate a framework to create any deep dive. There's a lot of structure to this conversation.

After that analysis, I asked it to distill its findings into a framework. The result is a five-phase framework called “The Guided Deep Dive.” Since it is quite long, I pasted the full description of The Guided Deep Dive on a separate page. You can use it as a prompt to generate a conversation based on your documents. Here's the summary of it:

The Guided Deep Dive is a framework for creating in-depth, conversational content using a two-person dialogue. It consists of:

  • Two Roles:

    • The Anchor: Acts as the audience's guide, asking questions and simplifying information.
    • The Expert: Provides the detailed knowledge, data, and explanations.

  • A Five-Phase Structure:

    1. Hook: Grab attention and state the purpose.
    2. Overview: Provide foundational concepts.
    3. Core Analysis: Dive into details through a question-and-answer cycle.
    4. Implications: Discuss practical applications and consequences.
    5. Synthesis: Summarize the key takeaway and pose a final, thought-provoking question.

* * *