# What is NotebookLM, and how do Audio Overviews work?

Date: 2024-09-22 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/notebooklm-audio-overviews-work/

**Answer:** NotebookLM is Google's research notebook that answers questions using only the documents you add to it, with citations back to the exact passages. Its new Audio Overview feature turns those documents into a downloadable, podcast-style conversation between two AI hosts.

NotebookLM works differently from a general chatbot. Instead of drawing on everything a model absorbed in training, it works from a small library you build: the documents, slides and notes you upload into a notebook. When you ask a question, the answer is grounded in that material, and each claim links back to the passage it came from, often with a quote. Google says the personal data you put in a notebook is not used to train NotebookLM.

On September 11, Google added a new way to take in that material: Audio Overview. With one click, NotebookLM writes and voices a lively discussion between two AI hosts who walk through your sources, summarize the main points, connect related ideas and joke with each other. You can play it in the browser or download it and listen on a walk.

## How an Audio Overview is made

Google first showed the idea at its I/O conference in May, explaining that it uses Gemini 1.5 Pro, its large model with a very long context window, to turn your source material into a personalized spoken conversation. Google hasn't published the inner workings, but it helps to picture three stages, like a small radio team at work:

1. **Reading.** The model takes in every source in the notebook at once.
2. **Scripting.** It writes a dialogue in which one host explains and the other asks questions, reacts and summarizes. That format makes dense material easier to follow by ear.
3. **Voicing.** Two synthetic voices perform the script, with enough back-and-forth banter that it sounds unscripted.

Google is clear about the limits. The feature is experimental. Large notebooks can take several minutes to process. The hosts speak only English for now, they sometimes introduce mistakes, and you can't yet interrupt them to ask a question. Most important, Google notes that the conversation is not a complete or neutral view of a subject. It reflects only what you uploaded. If your sources are one-sided, the hosts will be too.

## A worked example: a study notebook with three sources

Imagine a nursing student preparing for an exam on blood pressure. She creates a notebook and adds three sources:

| Source | What it contains |
|---|---|
| Lecture notes (Google Doc) | Her own notes from two classes |
| Textbook chapter (PDF) | The assigned chapter on hypertension |
| Slide deck (Google Slides) | Her instructor's review slides |

The steps:

1. Open NotebookLM, create a new notebook and add the three sources. At least one source is required.
2. Open the Notebook guide and press **Generate** under Audio Overview.
3. Wait a few minutes, then play the result or download it.

The episode she gets back might open with the hosts describing hypertension as a "silent" condition, move through how blood pressure readings work, then compare treatment approaches from the textbook with the summary on the slides. That outline is illustrative; every Audio Overview is different, because it is written fresh from your material.

What makes it useful for studying is the change of format. Reading the chapter a third time feels like work. Hearing two voices explain it, with one of them asking the questions a classmate might ask, lets her review the material on the bus and notice which parts she can't yet explain herself.

> **Try it:** Make a notebook with two or three documents you know well, such as a work report and its appendix, and generate an Audio Overview. As you listen, jot down anything that surprises you; those are the points to check first.

## When to use it, and when not to

Audio Overviews work well for getting the shape of a topic before you study it closely, for reviewing material you have already read, and for sharing a quick briefing with a colleague who will listen but won't read thirty pages. They are a poor fit when every detail matters, such as a contract, a dosage table or a legal deadline. For those, use NotebookLM's written answers, which point to the exact passage, or go straight to the document.

## How do you check what the podcast hosts say?

Treat the audio as a summary written by a fast, enthusiastic colleague: helpful, but not the final word. While you listen, note the specific claims that matter to you, especially numbers, names, dates and anything stated as a rule. Write down roughly where in the episode each one comes up so you can find it again.

Then go back to the notebook's chat and ask about each claim directly, for example: "What do my sources say about the target blood pressure for adults over 65? Quote the passage." The written answer comes with citations you can click to see the original text. If the hosts said something the sources don't support, or rounded a number, you will see the difference right away.

Watch for two quieter problems as well. If you can't find a statement anywhere in your sources, treat it as unverified, however reasonable it sounds. And because they always sound confident, a weak point in a source can come across as settled fact. If a claim matters, the source document, not the audio, is what you rely on.

## Sources

- NotebookLM now lets you listen to a conversation about your sources: https://blog.google/technology/ai/notebooklm-audio-overviews/
- Google I/O 2024: An I/O for a new generation: https://blog.google/inside-google/message-ceo/google-io-2024-keynote-sundar-pichai/
