How to Use Google AI Podcast: A Practical Walkthrough

Why Turning Content Into Audio Is Hard (and Why AI Podcasts Fix That)

Ask anyone who has launched a podcast and they'll tell you the same thing: the recording is the easy part. Post-production can consume over 70% of a creator's total effort, according to industry analysis — and that's after you've sourced equipment, booked voice talent, and carved out editing hours you don't have.

That traditional pipeline is exactly what AI podcast generation dismantles. The starting point is no longer a human voice but a text prompt, a research document, or a URL. For business owners, consultants, and content teams sitting on whitepapers, reports, and long-form articles, this means existing material can be repurposed into audio without a studio or a microphone. Google estimates its approach saves creators 5–10 hours per episode, and the economics are hard to ignore: Inception Point AI has produced 200,000 episodes at roughly $1 each, building an audience of 400,000 subscribers across Apple and Spotify.

The scale of this shift is visible across the industry. Hume AI raised $50 million while tens of thousands of creators adopted its software, and voice quality has crossed a threshold where AI voices are, per one executive, "pretty much indistinguishable from human."

Google's answer comes in two distinct flavors, and knowing which one fits your situation is half the battle:

• NotebookLM Audio Overview — a no-code, one-click feature that turns uploaded sources into a two-host AI conversation, ideal for non-technical users and content teams.
• Gemini 1.5 Pro + Text-to-Speech API — a developer-grade workflow with 380+ voices across 50+ languages, custom voice creation, and a free tier, for those who want control over script, voice, and deployment.
• Gemini app Audio Overview — a quick middle option for converting PDFs, text files, and Word documents on the go.

The right choice depends less on ambition than on workflow. If you're a business owner who wants a whitepaper turned into a listenable briefing, NotebookLM handles it in minutes with no coding. If you're a developer building audio into a product or pipeline, the Text-to-Speech route gives you the knobs to turn. Either way, the barrier that used to require equipment, talent, and hours of editing has collapsed into a prompt and a source document.

One caveat worth internalizing before you start: Google itself notes these generated discussions "may contain inaccuracies or glitches" and reflect only the sources you upload — so treat the output as a draft, not a finished broadcast. And if you want to push automation further, agent skills can chain these steps into a repeatable content-to-audio workflow; SkillzBench is building a free directory where you can browse, verify, and benchmark skills before trusting them with your pipeline.

Choosing Your Path: NotebookLM vs. Gemini App vs. Developer Workflow

Google gives you three different ways to turn documents into AI-hosted podcast audio, and picking the right one up front saves you real time. The best choice depends on how much control you need and how comfortable you are with code.

NotebookLM is the recommended starting point for most people. It requires no coding, and it accepts the widest range of sources: PDFs, websites, YouTube links, images, copied text, and even audio files. According to PCMag's hands-on tester, NotebookLM "offers the most natural voices, ones that sound like actual people" among the four podcast generators tested. Google's official walkthrough is just four steps: open NotebookLM, create a notebook, add at least one source, and click Generate.

The Gemini app is a faster but narrower alternative. It works on web and mobile with a free Google account, and PCMag calls it "a quick and convenient alternative to NotebookLM." The catch: it only accepts PDFs, text files, and Word documents — no images, video, or audio sources.

The developer path is different in kind, not just degree. Combining Gemini 1.5 Pro with Google's Text-to-Speech API, it suits teams that need fine-grained control over script, voice, and deployment:

• 380+ voices across 50+ languages, plus custom voice creation
• A free tier of 1 million characters per month on WaveNet and Neural2 voices
• Full script control via prompts, with output stored in Google Cloud Storage

That flexibility comes at a cost: this route requires Python coding, Google Cloud Storage setup, and chunking longer scripts around the API's 5,000-character-per-request limit. Google's own cloud tutorial frames it as the option for people who "want more control over generating unique audio experiences."

One caveat applies to every path: Google states the generated audio "may contain inaccuracies or glitches" and is "simply a reflection of the sources that you've uploaded" — not an objective summary. If you're building a repeatable content pipeline around these tools, agent skills can automate much of the surrounding workflow, from source gathering to publishing. SkillzBench is building a free directory and MCP connector for exactly that — browse and test skills at no cost as the platform rolls out.

Step-by-Step: Creating Your First AI Podcast with NotebookLM

Creating your first AI podcast takes less effort than writing an email. Google's official process is just four steps, and independent testing confirms it works exactly as advertised.

First, open NotebookLM — either the dedicated website or the mobile app on iOS or Android. Select "Create New" to start a fresh notebook. This becomes the container for your sources and generated audio.

Next, add at least one source. According to PCMag's hands-on testing, your options include:

• PDF files, uploaded via "Choose file"
• Website links, added by URL
• YouTube video links
• Images
• Copied text — enter a topic description and click "Insert"

For copied text, select the "Copied text" option, describe your topic, and click Insert. Need more material? Use the "Add" button to layer in additional sources. All sources appear on the left side of the page, where you can review or remove them before generating.

Now click "Generate." Expect to wait at least several minutes, depending on how many sources you've included. When it finishes, you get playback controls for Play, pause, speed adjustment, sharing, and download as a WAV file — handy for editing or distributing your episode elsewhere.

NotebookLM has grown well beyond its original English-only format. It now supports 50+ languages and multiple audio formats, including Deep Dive, Brief, and Debate, according to recent feature guides. Output even includes citations, so listeners can trace claims back to your uploaded sources.

Interactive Mode is the standout feature. Create a new notebook, add an audio file as a source (PCMag's tester used an M4A file), click Generate, then select "Interactive mode." Press Play, then hit the "Join" button to speak — you'll need to grant browser microphone access. You can pause the hosts and ask them questions aloud in real time.

One critical caveat: Interactive Mode works in Chrome and Microsoft Edge but not Firefox, and it's still in beta — PCMag notes it's "not always reliable." Plan your browser accordingly.

Google also cautions that generated audio "may contain inaccuracies or glitches" and reflects only the sources you uploaded, not an objective view of the topic. Verify before publishing.

If you want to extend this workflow — say, automating source gathering or episode scripting — agent skills can handle the repetitive parts. SkillzBench is building a free directory of tested, security-scanned agent skills plus a free MCP connector, so you can browse and equip the ones that fit your pipeline.

The Developer Route: Scripting Podcasts with Gemini and the Text-to-Speech API

When NotebookLM's one-click Audio Overview isn't flexible enough, developers can assemble a fully scripted, voice-controlled pipeline using Gemini 1.5 Pro and the Text-to-Speech API. The tradeoff is real coding work, but you gain control over every word, voice, and storage location.

Start by preparing your source material. Clean, well-structured content produces cleaner scripts, so edit for clarity and decide whether to ship one episode or split the material into a multi-episode series.

Next, hand the content to Gemini 1.5 Pro and prompt it to write a conversational script. Google's own example prompt targets technical listeners: "Generate an engaging audio overview script from this podcast, including an introduction, transitions, and a call to action. Target audience is technical developers, engineers, and cloud architects." Adjust the prompt to steer tone, pacing, and episode length.

For longer content, ask Gemini to extract sections as JSON first, then generate a script per section. That structure makes it easier to assemble multi-episode feeds or apply different voices to different segments.

Feed each finished script to the Text-to-Speech API, which offers 380+ voices across 50+ languages, plus custom voice creation. Both the Text-to-Speech API and Gemini models ship with a free tier, so prototyping costs nothing.

Two constraints shape the implementation:

• The API enforces a 5,000-character limit per request, so scripts must be chunked and stitched together for anything longer than a short segment.
• The WaveNet and Neural2 voices are recognized as AI roughly 60–70% of the time in blind listening tests, per an Oakgen comparison — fine for utilitarian briefings, less convincing for entertainment.
• Output arrives as base64-encoded MP3, which you decode and upload to Google Cloud Storage using the `google-cloud-storage` Python library, specifying content type `audio/mp3`.

Once the MP3 lives in a bucket, you can layer on background music, sound effects, or final edits before distribution. The same workflow scales from a single internal briefing to a recurring automated show.

For developers who want to automate more of this pipeline — script generation, chunking, storage orchestration — agent skills can package the whole sequence into reusable instructions. SkillzBench is building a free directory and MCP connector where you can browse and test those skills once the platform launches.

Avoiding the Pitfalls: Accuracy, Disclosure, and Automating the Rest of the Workflow

Before you hit publish on your first AI-generated episode, it's worth pausing on what Google itself says about the output. The company describes NotebookLM's Audio Overview as "still experimental," warning that the audio may contain inaccuracies or glitches — and that the generated discussion is "not a comprehensive or objective view of a topic, but simply a reflection of the sources that you've uploaded."

That second caveat matters more than most people realize. Your AI podcast is only as good as the sources you feed it. Upload a biased report or an incomplete dataset, and the two AI hosts will discuss it with the same confident, natural-sounding authority — PCMag's tester called NotebookLM's voices "the most natural" of the tools they tested, which is precisely what makes errors harder to spot. Independent Mode adds another wrinkle: it's still in beta and "not always reliable," and it doesn't work in Firefox at all.

The industry context reinforces why verification and disclosure aren't optional. AI voices have crossed a threshold where they're "pretty much indistinguishable from human," according to Hume AI's CEO, and listener trust suffers when audiences feel AI is replacing the human connection between hosts and listeners. When podcaster Erica Mandy substituted an AI voice for her own during laryngitis, she disclosed it to her audience — a model worth following.

A practical checklist for responsible publishing:

• Verify every factual claim in the generated audio against your original sources before publishing.
• Review your source material for bias and gaps, since the output mirrors whatever you uploaded.
• Disclose AI-generated audio clearly in show notes or episode intros.
• Listen to the full episode — glitches and mispronunciations won't surface in a spot check.

There's also the workflow itself. Post-production can consume over 70% of a creator's total effort, and AI podcasting saves an estimated 5–10 hours per episode — but only if the surrounding tasks don't eat the savings. This is where agent skills come in: instruction packages that let AI agents handle source preparation, script drafting, uploading, and publishing automatically. The catch is quality control — published research found that roughly 26% of agent skills carry at least one vulnerability, so grabbing skills blindly from the internet is its own accuracy risk.

SkillzBench is building a free, independent directory that addresses this directly: every skill is pinned to a specific version, security-scanned, and benchmarked in a sandbox so you can see its measured Lift before trusting it with your workflow. A free MCP connector lets you connect any agent to only the skills you've chosen. You can join the waitlist now to be first in line when it launches.

Verify the audio, disclose the AI, and automate the rest — carefully.