ExplainersBeginner9 min readJun 2026Checked Oct 2026

How Gemini Notebook (formerly NotebookLM) Turns Your Docs Into a Podcast

Grounded answers, a two-host script, and voices that learned to sound real: what Google has published about Audio Overviews, and what it has not.

Gemini Notebook is Google's research notebook. You add your own sources, such as PDFs, Google Docs, web pages and YouTube videos, and then ask questions about them or turn them into study material. Until 16 July 2026 it was called NotebookLM. Google renamed it that day (NotebookLM is now Gemini Notebook) and says old links redirect automatically. Its best known trick is the Audio Overview: two AI hosts talk through your sources like a podcast.

How does a pile of documents become a conversation that sounds this natural? Google has published part of the answer, mainly about the voices, and kept much of the rest to itself. This article sticks to what Google has said and marks everything else as an educated guess.

What goes into a notebook?

Gemini Notebook accepts many kinds of sources, according to Google's help page on adding sources: PDF, Word, PowerPoint, text, Markdown and CSV files, Google Docs, Slides and Sheets, images, audio files, eligible Play Books, web pages, and public YouTube videos that have captions. From a web page it imports only the text, and from a YouTube video only the captions.

As of October 2026, each source can hold up to 500,000 words, or 200 MB for an uploaded file. A free notebook holds up to 50 sources. Google's plans page raises that to 100 with Google AI Plus, 300 with Pro, and 500 or 600 with Ultra, depending on the plan.

Multiply that out and a free notebook can store about 25 million words. That is storage, not a reading limit. It is many times more text than a model can take in at once (its context window, which Context Windows: The Model's Working Memory explains), so for each question or Audio Overview the system has to pick the parts of your sources that matter. Google has not published how it does that.

How are answers tied to your sources?

When you chat with a notebook, Google says you get "grounded information based on your sources with clear in-line citations" (Learn about Gemini Notebook). Grounded means the answer is built from your documents rather than from the model's general knowledge, and each citation points to the passage it came from. Grounding and Citations: Answers You Can Check explains the general technique.

Citations make checking easy, but they are not a guarantee. A model can still attach a real passage to a claim that the passage does not quite support, so click through on anything that matters.

How does an Audio Overview get made?

You open the Studio panel, choose Audio Overview, and pick a format. Google's help page on generating an Audio Overview lists four:

Deep Dive (the default): two hosts unpack and connect topics from your sources.
The Brief: a single speaker gives the key takeaways in under two minutes.
The Critique: two hosts give a constructive review of material like an essay or a design document.
The Debate: two hosts argue different sides of the topic.

You can also choose Shorter, Default or Longer (English only), type a prompt that asks the hosts to focus on certain topics or pitch it at a certain level of expertise, and generate in more than 80 languages. In interactive mode, which is English only for now, you press Join while you listen and ask the hosts questions with your voice.

Step 1: a script

Google has not published how the script is written. What we know comes from the other end of the pipeline. The speech model Google DeepMind described for Audio Overviews produces dialogue "when given a script of dialogue and speaker turn markers" (Pushing the frontiers of audio generation). So a written script, marked with who says what, comes first. The product runs on Google's Gemini models, so a Gemini model most likely writes that script from the passages it picked, following the format and any focus you typed.

Step 2: the voices

DeepMind's post explains why the hosts sound so human. The speech model was first trained on hundreds of thousands of hours of speech. Then it was fine-tuned (given extra, focused training) on a much smaller set of unscripted conversations between voice actors, complete with real "umm"s and "aah"s. That step, DeepMind says, taught the model to switch between speakers reliably and to produce realistic pauses, tone and timing.

So the hesitations are not sprinkled in by a rule. The model learned them from real conversation, the same way it learned everything else about speaking.

The same post explains the speed. The model writes the audio as a long string of tokens (small numbered pieces of sound, over 5,000 for two minutes of dialogue), which are then decoded back into a sound wave. DeepMind says it can produce two minutes of dialogue in under three seconds on one of Google's TPU (tensor processing unit) chips. These figures come from a 2024 post, and Google does not say which version of the model the product runs today.

Step 3: a watermark

Google's page on SynthID says the technology embeds a watermark in audio made by the podcast feature of NotebookLM. You cannot hear it, but tools can later identify the audio as AI generated.

Which Gemini model is behind it?

Google says the product is built on its latest Gemini models but does not name one for every plan. In June 2026 it announced an upgrade to Gemini 3.5 for Google AI Ultra subscribers and some Workspace business customers (Do better research with NotebookLM). Google's plans page describes the rest only as "access to Gemini models", with higher access on the paid plans.

Quick reference (as of October 2026)

Property	Value
Name	Gemini Notebook (NotebookLM until 16 July 2026)
Sources per notebook	50 free; 100 Plus; 300 Pro; 500 or 600 Ultra
Size of one source	Up to 500,000 words, or 200 MB for an uploaded file
Audio formats	Deep Dive (default), The Brief, The Critique, The Debate
Length options	Shorter, Default, Longer (English only)
Languages	80+ for Audio Overviews; interactive mode English only
Voices	A speech model trained on real conversations (Google DeepMind)
Watermark	SynthID
Enterprise API	`notebooks.audioOverviews.create`, in Gemini Notebook Enterprise (Google Cloud docs)

The four figures below come from Google's help pages on adding sources and generating an Audio Overview.

Gemini Notebook audio by the numbers

Sources in a free notebook

500,000

Words in one source, at most

80+

Languages for Audio Overviews

2 min

Length of The Brief, at most

Source: Gemini Notebook Help, October 2026.

Frequently Asked Questions

Q: Does Gemini Notebook train on my uploaded documents? Google says your data "is not used to train Gemini Notebook unless you provide feedback". If you do send feedback, people at Google may review that whole interaction, including your uploads, and reviewed feedback is kept for up to three years, disconnected from your account (Privacy and Terms of Use in Gemini Notebook). With a Google Workspace work or school account, Google says your uploads, questions and answers are not reviewed by humans and not used to train AI models, even when you give feedback. Your sources stay stored in the notebook until you delete them. Notebooks you open inside the Gemini app follow the Gemini app's rules instead: those chats are saved under your Keep Activity setting and may be reviewed by humans (Gemini Apps Privacy Notice).

Q: Can I change what the hosts talk about? Yes, within limits. Before you generate, you can pick the format and length and type a prompt that asks the hosts to focus on certain topics or pitch it at a certain level. In interactive mode you can join the conversation and ask them to explain something differently.

Q: What happens with a very large notebook? A notebook can store far more text than a model reads at once, so the system works from the parts that seem relevant to your request. Google has not published how it chooses them. If an Audio Overview skips something you care about, ask for it directly in the focus prompt.

Q: Why does the audio sometimes mispronounce words or get things wrong? Google's own help page says Audio Overviews, "including voices, are AI-generated and may contain inaccuracies or audio glitches." The voices can stumble on names and jargon, and the hosts can misstate or embellish a point from your sources.

Q: Is the generated podcast safe to share? Mostly, as long as you check it first. The hosts work from your sources, but they can still get details wrong, so listen through and compare the key claims with your documents before you share. Remember that sharing the audio shares what is in your documents, and that it carries a SynthID watermark that marks it as AI generated.

Test yourself

Your manager uploads a 300-page annual report and a short strategy memo to a free notebook, generates a Deep Dive, and wants to send the audio to the whole team as "an accurate summary". She adds that the hosts' natural pauses prove a person must have edited it. What would you tell her?

Answer: Three things. First, the pauses do not mean a person edited it: Google DeepMind says its speech model learned the "umm"s, pauses and timing from real unscripted conversations during training. Second, "accurate" needs checking. The hosts work from the notebook's sources, but Google's own help says Audio Overviews may contain inaccuracies, and a 300-page report holds far more than one episode can cover, so the hosts have picked what to discuss. She should listen through, check the key claims against the chat's in-line citations, and use a focus prompt if one section matters most. Third, if the team only needs the headlines, The Brief gives a single speaker summary in under two minutes. And if the documents are confidential and she is on a personal account, she should not send feedback on that notebook, because feedback can be reviewed by people at Google.

If you enjoy unpacking how real AI products work, subscribe to Internals Decoded at internalsdecoded.com.

Sources

#notebooklm#gemini-notebook#audio-overview

More from the library