NotebookLM audio overviews — podcast-style summaries

When Google shipped the audio overview button in September 2024, the reaction was immediate and loud. Researchers, students, and journalists shared clips on social media within days. The feature felt genuinely new: not text-to-speech, not a voice reading a summary, but something closer to a radio programme about your uploaded documents. This page explains exactly how it works, what you can change, and where the limits are.

The mechanism behind the feature is more involved than it appears. The tool does not simply convert a written summary to audio. It composes an entirely separate script — a dialogue — that is structured around the natural rhythms of a conversation: one host introduces a claim, the other asks a clarifying question, the first provides an example, the second raises a counterpoint. That structure is what makes the output feel listenable rather than mechanical.

The four output modes

Brief

The Brief mode targets busy listeners who want an orientation in under four minutes. It covers the top-level themes and skips examples or counterarguments. Use it before a meeting when you have uploaded the briefing pack and want the thirty-second version of each section, or as a quick reminder of a notebook you have not opened in a while.

Standard

Standard is the default. The tool reads the full source set, identifies major threads, and produces an eight-to-eighteen-minute conversation that works through each thread in sequence. The pacing is designed for commute or gym use — enough context to follow without enough detail to require a notepad. Most users find this the right length for a reading list of three to eight sources.

Deep Dive

Deep Dive targets notebooks with dense or voluminous sources. The hosts go further into examples, data, methodology, and implications. Runtimes often land between twenty-five and forty-five minutes, occasionally longer for very large notebooks. Researchers processing a literature review or a full conference proceedings file have found this mode most useful, because the extended length allows the hosts to surface distinctions that the Standard mode glosses over.

Critique

Added during 2025, the Critique mode flips the tone of the conversation. Instead of explaining and exploring, the hosts probe and question. They look for internal contradictions in the corpus, claims that are asserted without evidence, and places where the sources themselves disagree. Critique mode is particularly valuable for legal, scientific, or policy research where the quality of the argument matters as much as the conclusion. The NIST Cybersecurity and Privacy Resource Center has noted that adversarial review approaches improve the rigour of AI-assisted analysis, which is precisely what this mode attempts.

Customisation options

Before generating an overview you can type a focus prompt — a plain-English instruction to the tool about what to emphasise or ignore. Examples: "focus on the methodology sections and skip the background review," or "concentrate on the 2024 data and ignore the historical context." The assistant uses this as a steering signal, not a rigid constraint, so it will still surface important points that fall outside the focus if they are prominent enough in the sources.

You can also hide individual sources before generating. If your notebook contains ten papers but you only want the overview to cover three of them, hide the other seven temporarily. They remain indexed and available for chat; they just do not influence the audio script.

During playback, the interjection feature lets you interrupt the hosts by typing a question or comment. The hosts respond in character — "good question, let's unpack that" — and then continue. This is useful when a host skims past a point you want clarified, or when the overview raises a question that your prompt did not anticipate.

MP3 export and offline use

Once an overview has been generated, a download button appears in the player. The file is a standard MP3. You can transfer it to a phone, load it into a podcast app, attach it to an email, or archive it alongside the notebook. Many research teams generate an overview as a shareable artefact — a colleague who does not have access to the notebook can still listen to the audio summary. The download is available on both free and Plus tiers.

Language and voice hosts

English hosts have been available since launch. The tool added localised voice pairs in Japanese, Spanish (Latin American), Portuguese (Brazilian), German, French, Italian, Korean, and Hindi during 2025. Additional languages are being added on a rolling basis. The host language is independent of the source language: you can upload German documents and produce a Spanish-language overview, for instance. Source content is indexed in the original language; the generation step then produces the dialogue in the selected output language.

Mode comparison table

Mode	Typical length	Tier	Best for
Brief	Under 4 min	Free & Plus	Pre-meeting orientation, quick refresh
Standard	8–18 min	Free & Plus	Commute listening, reading-list digest
Deep Dive	25–45+ min	Free & Plus (higher cap)	Dense literature, long reports
Critique	Varies	Free & Plus	Legal, policy, scientific review

Quentin F. Salvatierra, Knowledge Ops Manager at Sycamore Cove Publishing in Portland, ME, described the effect plainly: "We dropped our weekly digest call from ninety minutes to forty after staff started generating audio overviews of the week's submissions. The Standard mode covers everything a junior editor needs to know before the meeting; the Critique mode catches the structural problems before anyone has to say them out loud."