AI Podcast Generator: Turn Text Into a Two-Host Show

AI podcast generators turn source material or a script into a spoken show, often with two hosts. Some tools draft the conversation from documents you upload. Others voice a script you write. As of October 2026, Google's NotebookLM makes audio overviews from your sources, ElevenLabs and Google offer multi-speaker voice tools, and Apple Podcasts requires AI-generated hosts to be disclosed.

Two ways to make an AI podcast

The first decision is who writes the talk. Auto-generated tools draft the conversation for you. Scripted dialogue puts you in charge of every line. The table compares the two approaches.

| Question | Auto-generated from sources | Scripted dialogue | |---|---|---| | Who writes the conversation? | The tool drafts it from your sources | You write it | | How much control do you have? | Format, length, language and a custom prompt | Every line, pause and delivery cue | | Best for | Briefings, study guides and quick explainers | Branded shows, interviews and fiction | | Main risk | Factual errors you must catch | More writing time |

Auto-generated tools are quick, but Google warns that its audio overviews "may contain inaccuracies or audio glitches." Scripted dialogue takes longer, and it means you check every fact before the audio is made.

Generators compared, as of October 2026

The first table compares what each tool accepts and how it works. Each row is based on the provider's documentation as read on October 17, 2026.

| Tool | Input | Hosts and format | Languages | Control | |---|---|---|---|---| | NotebookLM Audio Overviews | Sources you upload | Four formats: Deep Dive and The Critique (two hosts), The Debate (two hosts), The Brief (one speaker) | More than 80 | Length, custom prompt, interactive mode in English | | ElevenLabs GenFM on ElevenReader | PDFs, articles, ebooks, docs, text and links | Two AI co-hosts | 32 | App-based; controls not detailed in the source we read | | ElevenLabs Text to Dialogue | Your script | No limit on speakers | 90+ on Eleven v4, 70+ on Eleven v3 | Audio tags on each turn, dashes and ellipses | | Google Gemini text to speech | Your script | Multi-speaker on three of the four listed models | 87 (24 generally available, 63 in preview) | Style prompts and bracketed tags | | Microsoft VibeVoice | Open-source research code | Up to four speakers in the research release | Not stated | Not for production |

Microsoft's VibeVoice repository describes a research release and says the text-to-speech code was removed in September 2025 after misuse. Do not build a production show on it.

Pricing varies widely, so compare plans by credits and minutes, not by the monthly headline price.

| Service | Plan | Monthly price (USD) | Credits or minutes | |---|---|---|---| | Wondercraft | Free | $0 | 150 credits, one time | | Wondercraft | Creator | $25 ($21 billed yearly) | 1,000 credits, full commercial rights, no watermark | | Wondercraft | Pro | $45 | 2,000, 4,000 or 6,000 credit options | | ElevenLabs | Pro | $99 | 600,000 credits (about 600 minutes) | | ElevenLabs | Scale | $299 | 1.8 million credits (about 1,800 minutes) | | ElevenLabs | Business | $990 | 6 million credits (about 6,000 minutes) | | OndaLabs | Starter | $9 | 170,000 credits (about 189 minutes) | | OndaLabs | Creator | $22 | 400,000 credits |

These are the prices shown on each provider's pricing page as read on October 17, 2026. ElevenLabs Starter and Creator prices were under first-month promotions, so they are left out. ElevenLabs annual billing is 10 times the monthly price, and OndaLabs annual billing is the same.

How NotebookLM audio overviews work

NotebookLM is the most automatic option in this list. Google's help page describes these features:

Formats. Deep Dive is the default, with two hosts. The Brief has one speaker and covers key takeaways in under two minutes. The Critique has two hosts giving constructive feedback. The Debate has two hosts on opposing sides.
Length and languages. You can choose Shorter, Default or Longer. Longer is English only. Overviews are available in more than 80 languages.
Customization. A custom prompt can focus the topics or adjust the expertise level.
Speed. Generation runs in the background and can take a couple of minutes.
Interactive mode. You can join the hosts by voice, but only in English and only for newly generated overviews.
Sharing. Link sharing needs the notebook to be shared, and public link sharing is for consumer accounts. Deleting an overview breaks its share links.
Limits. Google lists notebook and source limits by plan, subject to change. Standard allows 100 notebooks and 50 sources per notebook, Plus allows 200 and 100, and Pro allows 500 and 300. The limits page shows no prices and no daily audio limit.

Sources: NotebookLM Audio Overviews and NotebookLM plan limits.

How to write a script for two AI hosts

Scripted dialogue gives the best control. A few habits make the output sound more like a conversation:

Write each turn with a speaker name, and keep turns short enough to sound like speech.
Use dashes for interruptions and ellipses for trailing speech. ElevenLabs' documentation describes both.
Add delivery cues in square brackets where the engine supports them. ElevenLabs Text to Dialogue uses audio tags in brackets, and Gemini text to speech uses tags such as [sigh] and [whispering].
Keep each request within the tool's limit. ElevenLabs documents 2,000 characters per Text to Dialogue request, so join the chunks yourself.
Generate more than one take. ElevenLabs says output is nondeterministic, so you may need several generations to get one you like.

Write the script yourself. The US Copyright Office says that prompts alone do not make a user the author of AI output, while human-written material is protected as human expression (US Copyright Office, Part 2 report).

Turn your text into a two-host show in OndaLabs

OndaLabs' two-speaker dialogue tool voices a script you provide. It is included on the Free plan with Onda Lite, and Onda Flash, the most expressive model, is on paid plans. Here is what else the studio covers for a show:

Voices. There are 30 prebuilt voices, usable in every supported language.
Delivery. Describe the delivery in plain words and add supported expressive tags to the script.
Show notes. Speech to text costs 50 credits per minute, so you can transcribe the finished episode for show notes.
Music. Music generation is available on Creator and up. It costs 2,500 credits per clip or 5,000 credits per song, which covers a bumper or a theme.
Commercial use. It is included from the Starter plan at $9 per month.

Be clear about the limit. The dialogue tool voices a script you provide. It does not turn documents into a finished show on its own, so write or draft the script first with any writing tool. Compare the OndaLabs plans before you commit to a monthly show.

Publishing rules: Apple Podcasts and Spotify

Disclosure rules matter as much as the audio. The two platforms with the clearest policies are:

Apple Podcasts, guideline 1.11. Creators who use AI to generate audio or video, including synthetic voices, AI-generated hosts or personas, or AI replicas of real people, must disclose this prominently to listeners in the content and in the metadata for each episode and show (Apple Podcasts guidelines).
Apple Podcasts, guideline 1.12. AI must not be used to mislead or to portray real-life events deceptively, such as fabricated news.
Spotify, announced May 19, 2026. Spotify extends its Verified by Spotify badge to podcasts, and it says it will remove shows and content that impersonate another creator or host's likeness without permission, including through AI voice cloning (Music Business Worldwide).

The sources we checked show no Spotify rule that requires AI disclosure for podcasts, so this post does not claim one. Apple's requirement is the clear one.

Audio checks before you upload

Spotify's creator help page says it accepts MP3, M4A and WAV files, mono or stereo, with no file size limit. It says nothing about AI disclosure (Spotify: publishing audio episodes). For loudness and bit rate targets, read the podcast intro voice guide, which covers Apple's specifications in detail. If you want a comparison of voices for the host lines, see the best AI voices for narration.

Frequently asked questions

What is an AI podcast generator?

An AI podcast generator produces a spoken show from text, documents or a script. Some tools write the conversation from your sources, while others voice a script you provide. Check how much control you have, because automatic generators can include inaccuracies that you must catch before publishing.

How does NotebookLM make a podcast from documents?

You add sources to a notebook, then choose an audio overview format, a length and an optional custom prompt. NotebookLM generates the audio in the background, which can take a couple of minutes. The formats include Deep Dive with two hosts, The Brief with one speaker, The Critique and The Debate.

Can I publish an AI-generated podcast on Apple Podcasts and Spotify?

Apple Podcasts allows AI-generated shows when they are disclosed prominently, in the content and in the metadata for each episode and show. Spotify removes shows that impersonate someone's likeness without permission, including through AI voice cloning. Check both platforms' current rules before you publish.

Do I have to disclose that my podcast is AI-generated?

On Apple Podcasts, yes. Apple's guidelines require prominent disclosure of synthetic voices, AI hosts and AI replicas of real people. The sources we checked show no Spotify disclosure rule. This is general information, not legal advice.

Is NotebookLM free, and are there limits?

NotebookLM has plan-based limits on notebooks and sources, and Google says these are subject to change. Google's limits page does not show prices or a daily audio limit. Confirm the current plan details in your Google account before you depend on it for regular production.

Can AI hosts clone a real person's voice?

Spotify says it will remove content that impersonates someone's likeness without permission, including through AI voice cloning. Clone a voice only from a recording you have permission to use. OndaLabs' custom voice option works on that basis.