You have a 90-page PDF. You need to understand it by tomorrow morning. You could read it – highlighting, note-taking, re-reading the dense sections. Or you could hand it to Audio Overview and listen to two AI hosts walk through the key ideas on your commute, at the gym, or while cooking dinner.

Audio Overview is NotebookLM’s breakout feature. It takes your documents – PDFs, markdown notes, web pages, YouTube transcripts, audio recordings – and generates a podcast-style conversation between two AI hosts. They discuss your material with natural banter, cite specific passages, debate implications, and surface connections you might have missed. The result sounds like a well-produced podcast episode, not a robotic text-to-speech readout.

This feature drove a significant share of NotebookLM’s 12-month sustained growth on Google Trends. It is the reason people tell their colleagues about NotebookLM. And with GeminiDesktop, Audio Overviews are now generated and saved locally on your Mac, Windows, or Linux machine – no browser tab required.

TL;DR / Key takeaways

Two-host AI podcast from your sources – Gemini 3 Pro plans the conversation arc, gemini-3.1-flash-tts-preview with MultiSpeakerVoiceConfig renders natural two-voice audio.
80+ languages with native-sounding discussion, not awkward translation.
Four styles: conversational, debate, interview, lecture. Three length targets: 2-4 min, 5-10 min, 15-20 min.
Interactive Mode lets you interrupt the hosts mid-sentence, ask a question, and resume – powered by Gemini Live’s WebSocket streaming with native barge-in.
NotebookLM lives in a browser; GeminiDesktop saves the MP3 locally so Apple Music, CarPlay, iPhone sync, and Spotify local files all see it natively.
No true competitor: ElevenLabs, Descript, and Spotify AI DJ each solve adjacent problems, but no one else generates two-host podcast conversations from documents you own.
Windows gap – Google has no native Gemini Windows app. GeminiDesktop is the only native Windows option. See Native Gemini Windows app.

What Audio Overview actually produces

An Audio Overview is a generated audio conversation between two AI hosts. Here is what distinguishes it from standard text-to-speech or summarization:

Two distinct voices. The hosts have different speaking styles, different cadences, and different roles in the conversation. One tends to introduce topics and provide context. The other asks clarifying questions, pushes back on assumptions, and draws connections to related ideas. The dynamic creates a natural listening experience that holds attention far better than a single narrator reading a summary.

Natural conversational flow. The hosts interrupt each other, build on each other’s points, pause to think, and occasionally disagree. The audio does not sound scripted. It sounds like two people who have both read the same material and are genuinely discussing it.

Source-grounded content. Every claim the hosts make is traceable to your uploaded documents. They reference specific sections, quote key passages, and flag when the source material is ambiguous or contradicts itself. This is not a hallucinated podcast – it is a faithful synthesis of what your documents actually say.

80+ languages. Audio Overview supports generation in over 80 languages. The hosts speak naturally in the target language, not as if they are reading a translation. Chinese, Japanese, Korean, Spanish, French, German, Arabic, Hindi – the quality is consistent across all supported languages.

Customization options

Audio Overview is not a one-size-fits-all output. You control three dimensions:

Length

Short (2-4 minutes). A tight summary covering the top-level takeaways. Good for email briefs, meeting prep, or quick triage of a document you are not sure is worth your time.
Medium (5-10 minutes). The default. Covers the main arguments, key evidence, and notable implications. This is the format most people use for commute listening.
Long (15-20 minutes). A deep dive that walks through the source material section by section. Useful for academic papers, legal documents, or any material where detail matters.

Style

Conversational. The default two-host discussion format. Informal, accessible, good for general audiences.
Debate. The hosts take opposing positions on the material’s central claims and argue their cases. Useful for evaluating the strength of an argument or identifying weaknesses in a research paper.
Interview. One host plays the role of an interviewer asking probing questions; the other plays the role of the author or an expert on the topic. Good for material written by a specific person whose perspective you want to interrogate.
Lecture. A single-voice, structured presentation that follows the document’s organization. More formal, more linear, better for materials that need to be absorbed in order.

Language

Any of the 80+ supported languages. You can generate an Audio Overview of an English-language paper in Japanese, or a Chinese-language report in Spanish. The hosts will discuss in the target language while faithfully representing the source content.

The technical stack behind it

Audio Overview’s quality comes from a two-stage pipeline:

Stage 1: Script generation with Gemini 3 Pro

The source documents are sent to Gemini 3 Pro with a specialized prompt that produces a full conversation script. This is not a simple summarization prompt. The model plans the arc of the conversation – what to open with, where to go deep, when to step back and provide context, how to close. It assigns dialogue to two distinct speakers and includes natural interjections, questions, and transitions.

Gemini 3 Pro’s long-context capabilities (up to 2 million tokens) mean that even massive source collections are processed in full. The model does not truncate your documents or rely on extractive snippets. It reads everything and reasons over the complete content.

Stage 2: Voice synthesis with gemini-3.1-flash-tts-preview

The generated script is rendered into audio using Google’s gemini-3.1-flash-tts-preview model with MultiSpeakerVoiceConfig. Two voice profiles are used:

Kore – a measured, authoritative voice that typically handles context-setting and explanatory segments.
Puck – a more conversational, energetic voice that handles questions, reactions, and commentary.

The combination produces audio that sounds like two real people having a conversation, not like two TTS engines reading alternating paragraphs. The prosody, pacing, and emphasis are all generated contextually based on the content of the script.

The TTS model is the same one NotebookLM uses for its web-based Audio Overviews. The quality ceiling is identical – GeminiDesktop’s differentiator is the pipeline around it (local files in, MP3 out, filesystem integration), not a different underlying model.

The desktop advantage: local files in, local files out

In NotebookLM’s web interface, Audio Overview plays inside the browser. You can listen to it. You can share a link. But you cannot download the MP3 to your local filesystem through a simple, native file-save operation. The audio lives inside the product.

GeminiDesktop changes this in two ways.

Local source files, no upload required

Your documents are already on your machine. GeminiDesktop reads them directly from disk. There is no upload step. A folder of 50 PDFs becomes a notebook in the time it takes to scan the filesystem. When you generate an Audio Overview, the source content is sent to the Gemini API from your machine, processed, and the audio is returned directly.

MP3 saved locally

The generated Audio Overview saves as an MP3 file to ~/Music/GeminiDesktop/ on macOS, Music\GeminiDesktop\ on Windows, or any directory you configure. Once saved:

Apple Music indexes it automatically. It appears in your library alongside your regular music and podcasts.
iPhone syncs it via iCloud Music Library. Listen on the go without any manual transfer.
CarPlay plays it. Your morning commute becomes a briefing session on yesterday’s research.
Spotify local files picks it up if you point Spotify at the folder.
Overcast / Pocket Casts lets you import via local file on iOS.
Any media player reads it. Drag the MP3 into VLC, Winamp, foobar2000, or a web player.

The file is yours. It is a standard MP3 on your filesystem. You can rename it, move it, email it to a colleague, upload it to a CMS, or archive it. It does not depend on a browser session, a cloud service, or a product subscription to remain accessible.

Interactive Mode: interrupt the podcast and ask questions

Audio Overview’s newest capability is Interactive Mode. Instead of passively listening to a pre-generated conversation, you join it live.

You activate Interactive Mode and the two AI hosts begin their discussion of your material. At any point, you can barge in – literally interrupt them mid-sentence – and ask a question. “Wait, go back to that methodology section. What was the sample size?” The hosts pause, address your question with grounded references to your source documents, and then resume the conversation from where it makes sense.

This runs on Gemini Live’s WebSocket-based real-time API with native barge-in support. The latency is low enough that the interaction feels like a real conversation, not a turn-based exchange.

In GeminiDesktop, Interactive Mode is triggered via a global hotkey (Cmd+Shift+L on Mac, Ctrl+Shift+L on Windows/Linux). You do not need to navigate to a specific tab or open a specific window. Press the hotkey from anywhere on your machine, and the conversation starts.

Use cases that work well

Research paper triage. You have 20 new papers from arXiv. Generate a 3-minute Audio Overview for each. Listen through them during your morning walk. The ones that sound relevant, read in full. The rest, skip. Total time: one walk instead of one day.

Meeting prep. You have a 40-page strategy document to review before a 2:00 PM meeting. Generate a 10-minute Audio Overview. Listen while eating lunch. Walk into the meeting with a solid understanding of the key proposals, the risk factors, and the open questions.

Language learning. You have an English-language textbook and you are learning Japanese. Generate an Audio Overview in Japanese. The hosts discuss the English source material in Japanese, giving you both subject matter understanding and listening practice simultaneously.

Content repurposing. You have a long blog post or whitepaper. Generate an Audio Overview and publish it as a podcast episode on your feed. Your audience gets the same content in a format they can consume while multitasking.

Commute briefings. Configure a watched folder. New documents that arrive during the day trigger Audio Overview generation overnight. By morning, your Music folder has fresh briefings ready for your commute. Zero manual steps.

Legal document review. A stack of contracts in a folder. Generate a 10-minute “debate” style Audio Overview where the hosts argue competing interpretations. Great for spotting ambiguities you’d miss in silent reading.

Journalism investigations. Deposition transcripts, leaked documents, source interviews. Generate an Audio Overview in “interview” style – one host plays the investigator asking probing questions. Exposes narrative gaps and inconsistencies.

How it compares to adjacent audio tools

Capability	NotebookLM	ElevenLabs Studio	Descript Overdub	Spotify AI DJ	GeminiDesktop
Two-host podcast generation	Yes	No (you write script)	No (you write script)	No (music-focused)	Yes
From your documents	Yes	No	No	No	Yes
Grounded source citations	Yes	N/A	N/A	N/A	Yes
Natural conversation (interruptions, reactions)	Yes	Limited	Limited	No	Yes
Interactive / barge-in mode	Yes	No	No	No	Yes
80+ languages	Yes	29 languages	Limited	N/A	Yes
Local MP3 save	No	Yes (per generation)	Yes	No	Yes
Voice cloning	No	Yes (strong)	Yes	No	No
Native desktop app	No	Web only	Yes (Mac/Win)	Mobile-first	Yes (Tauri)
Price	Free + Pro	$5+/mo	$19+/mo	With Spotify Premium	Free + API

ElevenLabs Studio is the best TTS quality on the market, and its multi-speaker mode can produce two-voice audio – but you write the script. It’s infrastructure for audio production, not a document-to-podcast generator.

Descript Overdub offers voice cloning and podcast editing. Again, script-first. You can import an Audio Overview MP3 into Descript and edit it, which is a nice complementary workflow.

Spotify AI DJ narrates music recommendations with an AI personality. Completely different category – it doesn’t read your documents.

NotebookLM is the direct peer. Same generation quality. Architecture differs on surface: browser-only vs. native-desktop-with-local-files.

Why Google hasn’t shipped desktop Audio Overview

NotebookLM’s Audio Overview is excellent, and Google has invested in making it a headline feature. But the desktop gap persists because Google Labs ships NotebookLM as a web-first product with no public API, tiny team, and no platform coverage mandate. A desktop Audio Overview surface from Google would require either a full native NotebookLM app (which Google hasn’t built) or a NotebookLM API (which Google hasn’t shipped). GeminiDesktop closes the gap by calling the public Gemini API’s TTS model with a MultiSpeakerVoiceConfig that matches NotebookLM’s pattern, then handles the file I/O natively on your machine. See Why NotebookLM Has No Native App for the full structural analysis.

Windows and Intel Mac context

On Windows, the story is even starker – Google ships no native Gemini app at all, and the Google App for Windows (2026-04-14) is a thin search shell with no NotebookLM surface. On Intel Mac, Google’s Gemini for Mac (2026-04-15) is Apple Silicon only. GeminiDesktop’s Audio Overview works on both – Tauri 2.x produces a universal binary for Intel + Apple Silicon, plus a native Windows build. Context: Why Google didn’t make a Gemini Windows app and Intel Mac alternatives.

How it compares

Capability	NotebookLM Web	GeminiDesktop
Two-host podcast generation	Yes	Yes
80+ language support	Yes	Yes
Length customization (2-20 min)	Yes	Yes
Style options (conversational/debate/interview/lecture)	Yes	Yes
Interactive Mode (barge-in)	Yes	Yes
Save MP3 to local filesystem	No	Yes
Generate from local files without upload	No	Yes
Apple Music / CarPlay integration	No	Yes (via local file)
Watch-folder auto-generation	No	Yes
Global hotkey for Interactive Mode	No	Yes
Works on Intel Mac	Browser	Yes (native)
Works on Windows	Browser	Yes (native)
Offline playback of generated MP3	No	Yes

Limitations

Audio Overview generation uses Gemini API quota. Long-form (15-20 minute) audio at high quality consumes more tokens than a short summary. Voice cloning is not supported – if you want your own voice narrating, use Descript Overdub or ElevenLabs instead. The “debate” and “interview” styles work better on opinionated source material (essays, op-eds, case studies) than on neutral reference docs. Interactive Mode requires a stable internet connection for the Gemini Live WebSocket stream. Source documents beyond 2 million tokens may be sampled rather than fully included. And while 80+ languages are supported, audio quality varies – English, Mandarin, Spanish, and Japanese are the strongest; some less-resourced languages are still improving.

FAQ

Does NotebookLM have a desktop Audio Overview? No. NotebookLM’s Audio Overview plays only inside the browser at notebooklm.google.com. There is no MP3 download, no native desktop player, and no filesystem save. GeminiDesktop.app is the only native desktop option today.

NotebookLM vs Gemini – what’s the difference for Audio Overview? Audio Overview is a NotebookLM feature, not a general Gemini feature. The standard Gemini app doesn’t generate two-host podcasts from your documents. GeminiDesktop reproduces the NotebookLM workflow against the public Gemini TTS API.

Can I use Audio Overview offline? Generation requires the Gemini API (internet). Once the MP3 is saved locally, playback is fully offline.

Is Audio Overview free? NotebookLM offers Audio Overview with a daily generation quota on the free tier; NotebookLM Plus (Google AI Pro) raises the quota. GeminiDesktop’s generation counts against your Gemini API quota – the free tier supports typical daily use.

What file formats can Audio Overview read? PDFs, plain text, Markdown, YouTube URLs, web pages, audio recordings (transcribed first), images (multimodal). GeminiDesktop reads the same list plus anything your Gemini API key supports.

Can I use my own voice for Audio Overview? No. Audio Overview uses Google’s Kore and Puck voices for the two hosts. For voice cloning, look at ElevenLabs or Descript Overdub and combine workflows.

How long does generation take? Script generation with Gemini 3 Pro: 30-90 seconds depending on source length. TTS rendering: 10-30 seconds for a medium-length overview. Total: typically under two minutes.

Try it

Download GeminiDesktop, drop a folder of documents, and generate your first Audio Overview. The MP3 saves to your Mac, Windows, or Linux machine. Listen however you want, wherever you want.

geminidesktop.app/app