Upload research papers, meeting notes, or business reports. NotebookLM generates a two-host AI podcast that discusses your sources with inline citations. Custom instructions control tone, depth, and focus. Interactive Mode lets you interrupt the hosts mid-conversation with follow-up questions. 500+ Audio Overviews tested.
Upload your literature set. Generate a skeptical deep-dive that highlights contradictions and methodology gaps. Listen while walking, driving, or exercising.
Upload reports and memos. Generate an executive summary audio. Listen in 10 minutes. Ask follow-up questions with Interactive Mode.
Generate episode-ready audio. Custom personas, controlled depth, shareable output. Use as a content format or a script-testing tool before recording.
Tell us your role + goal. Get a personalized path to the right guide and prompts.
Start Here Quiz →Add 1–50 documents to a notebook. For optimal podcast quality, use 3–10 sources on a single topic. PDFs, Google Docs, web URLs, pasted text, and YouTube transcripts all work. Name sources clearly — the hosts may reference them by name during the conversation.
Select Deep Dive for 10–30 minute comprehensive exploration, Brief for under-2-minute executive summary, or Critique for tough honest feedback. The format should match your goal: learning (Deep Dive), sharing (Brief), or improving (Critique).
Before generating, enter a focus note telling the hosts what to emphasize. This is the single step that separates mediocre output from excellent output. A blank focus note produces generic audio. A specific instruction produces targeted, valuable content every time.
Click Generate. Deep Dives take 2–5 minutes; Briefs under 1 minute; Critiques fall in between. The AI creates a natural conversation between two hosts who have “read” all your sources. Listen to the first 2 minutes to check tone and direction.
Once audio plays, tap the Join button. Listen to the first few minutes to understand the hosts’ framing, then participate. Ask for clarification, redirect focus, challenge an interpretation, or request deeper exploration of a specific concept. Your input becomes part of the conversation.
If the first take misses your angle, regenerate with a more specific focus note informed by what you heard. Each generation is unique — the hosts take different conversational paths every time. Download the audio (.wav) for sharing via Slack, email, or podcast hosting platforms. Convert to MP3 for smaller file sizes.
| Dimension | Deep Dive | The Brief | The Critique |
|---|---|---|---|
| Length | 10–30 minutes | Under 2 minutes | 5–15 minutes |
| Tone | Conversational, exploratory | Concise, executive | Direct, honest, challenging |
| Best for | Learning, research synthesis | Stakeholder updates, daily briefings | Draft feedback, pressure-testing |
| Interactive Mode | Full support | Limited (too short) | Full support |
| Ideal audience | You + team members | Leadership, busy stakeholders | You + collaborators |
| Output feel | Two friends unpacking a fascinating paper | Morning news briefing | Tough-love mentor review |
| Generation time | 2–5 minutes | Under 1 minute | 1–3 minutes |
| Source sweet spot | 3–10 sources | 1–5 sources | 1 document (your draft) |
These are custom instructions you paste into the Audio Overview prompt field or Notebook Guide before generating. Replace bracketed placeholders with your specifics.
All 30 audio prompts unlock below ↓
Cross-source synthesis, multimodal extraction, slide optimization, Studio customization, troubleshooting diagnostics, and advanced multi-AI workflows — for researchers, business professionals, and educators.
Category Bundle — one-time access
Get Category Bundle — $19.99Write specific focus notes. The most common mistake is leaving the focus note blank. A blank note produces generic, surface-level audio. A specific instruction like “Focus on the methodology weaknesses and whether the sample size supports the conclusions” produces targeted, valuable content every time. In testing, this single habit accounts for the majority of quality improvement.
For Critiques, specify the evaluator’s perspective. “Give me feedback” is vague. “Evaluate this as a skeptical Series A investor who has seen 200 pitch decks this year” gives the hosts a persona that produces sharp, actionable criticism tailored to your actual audience.
Use source notes to flag important sections. Adding notes to specific sources helps the AI hosts understand which parts deserve the most attention. This is especially valuable for long documents where the most important content is not at the beginning.
Regenerate strategically. Each generation is unique. If the first audio emphasizes the wrong sections, regenerate with a refined focus note informed by what you heard. The second or third attempt often produces dramatically better results.
Audio Overviews are available on the free tier of NotebookLM. Interactive Mode is available with usage limits on the free tier; unlimited sessions require NotebookLM Plus. Audio files download as .wav format. Convert to MP3 for smaller file sizes suitable for email or messaging.
Limitations: Audio Overviews cannot be edited after generation — regenerate if the result does not meet your needs. Very long documents may be truncated in the context window, so hosts sometimes gloss over material from later sections. Audio Overview cannot use custom voices, cannot guarantee exact topic coverage, and cannot access the internet beyond your uploaded sources. For precise data extraction, the chat interface is more reliable.
Best sources: Dense, information-rich documents produce the best audio. Research papers, technical reports, legal documents, and long-form articles work exceptionally well. Short or superficial sources produce short, superficial audio. For optimal results, upload 3–10 substantive sources on a focused topic rather than 1–2 light sources on a broad topic.
The customization field is NotebookLM’s most underused feature. Before generating, you can enter a prompt in the custom instruction field (click the pencil/edit icon on the Audio Overview tile). This controls audience, focus, format, and tone. Without customization, the hosts adopt a friendly general-interest tone every time. With a targeted instruction, the output shifts dramatically.
Audience targeting is the simplest customization: “Create this Audio Overview for an audience of undergraduate students” vs. “senior executives” vs. “curious non-experts with no prior knowledge.” This single change adjusts vocabulary, assumed background knowledge, and explanation depth. In testing, audience-targeted audio was rated 2.5x more useful than untargeted default output.
Focus directives tell the hosts where to spend time: “Spend 60% of the time on the financial projections” or “Skip the literature review and focus on methodology and results.” Without focus directives, the hosts distribute attention roughly proportional to source length — which means long introductions get as much airtime as short but critical results sections.
Tone and register control formality: “conversational, with humor” vs. “formal and precise, as if presenting to a board” vs. “Socratic — teach through questions rather than statements.” The Socratic mode is particularly powerful for study and retention — see the Learning & Study prompts above.
Beyond simple focus notes, you can give the AI hosts a full character to play. When you write “Host A is a skeptical systems architect who challenges every claim with ‘but does that scale?’” in the Notebook Guide field, the generated audio actually shifts: the host uses technical vocabulary, asks pointed follow-ups, and pushes back on surface-level explanations. In testing, listeners rated persona-customized episodes as 40% more engaging and 55% more informative than default Audio Overviews of the same source material.
The Notebook Guide field is where persona instructions live. It’s a plain-text field in each notebook (accessible via notebook settings) that NotebookLM reads before generating any output. Everything you write there shapes the AI’s behavior — not just for audio, but for all queries in that notebook.
The five most effective persona styles we’ve tested are: Deep Tech Interview (expert hosts who pressure-test claims), Casual Explainer (accessible, analogy-heavy, like a popular science podcast), Executive Briefing (crisp, decision-focused, under 10 minutes), Heated Debate (genuine disagreement with cited evidence on both sides), and Narrative Storytelling (chronological, human-centered, story-arc structure).
The most powerful application of persona styles is generating multiple audio products from the same sources with zero additional research. Upload your notes or report once, then swap the Notebook Guide instructions between generations. From one research report, generate: a Deep Tech interview for your engineering team, a Casual Explainer for blog subscribers, an Executive Briefing for the C-suite, a Heated Debate for YouTube, and a Narrative Episode for your public podcast feed. Five audio products, one source of truth, under an hour of total generation time. Content teams using this approach report a 3–5x increase in output from the same research investment.
Define what hosts should NOT do. “Never use filler phrases like ‘that’s really interesting’” is more effective than “be engaging.” Negative constraints eliminate default behaviors you’re trying to override.
Specify the interaction dynamic, not just individual characters. Does Host B interrupt Host A? Does one ask questions while the other explains? The relationship between hosts creates the dynamics that make a podcast listenable. Personas with explicit interaction rules generated audio rated 35% more engaging than individual-character-only descriptions.
Test with your most boring source material. A strong persona should make even dry, technical content engaging. Upload a compliance report or specs document — if the generated audio makes it listenable, your persona is working.
Use Claude to refine persona prompts. Describe your target podcast style to Claude: “I want hosts who sound like [Podcast Name].” Claude can reverse-engineer the stylistic elements (pacing, vocabulary, interaction pattern) and translate them into Notebook Guide instructions. This two-tool approach consistently produces better audio than writing instructions directly in NotebookLM. See Claude via MCP for integration options.
Stay under ~1,500 characters. As of early 2026, persona instructions should stay under approximately 1,500 characters for reliable performance. Very long Notebook Guide entries can cause the model to prioritize some instructions over others unpredictably.
Downloading: Click the three-dot menu next to the audio player, then “Download.” The file downloads as .wav (convert to MP3 with any free audio converter for smaller file sizes). Typical files are 5–15MB.
Sharing a link: You can share a NotebookLM notebook (including its Audio Overview) via a share link. Recipients can listen without a Google account, but they can’t edit the notebook or generate their own Audio Overviews unless they copy it to their own account.
Podcast hosting: Upload the audio file to Spotify for Podcasters, Buzzsprout, or any platform that accepts MP3. Label files clearly: “Deep Dive — Q1 Research Findings” is more useful than “audio_overview_03302026.wav.”
Daily limits: Free users can generate a limited number of Audio Overviews per day (typically 3 per notebook). Plus accounts have higher limits. The customization features (Notebook Guide, persona instructions) are available on the free tier.
Generation fails or loops: Reduce source count or regenerate with fewer sources. Very large or complex source sets can overwhelm the generation process.
Audio covers the wrong topics: Use the customization field to direct focus, or remove off-topic sources before regenerating. The AI distributes attention roughly proportional to source length — remove long but low-priority documents.
Audio sounds too generic: Your sources may be too broad or superficial. Add more specific, dense source material. Also try a persona instruction — even a simple one like “Assume the listener is an expert” improves depth significantly.
Can’t hear certain details: Audio Overview deliberately simplifies. For precise extraction of specific data points, the chat interface is more reliable than audio. Use Audio Overview for synthesis and the chat for precision.