Key takeaways
Start with the format. A solo narrator works in any text-to-speech tool, while a two-host conversation needs a dialogue tool like ElevenLabs Text to Dialogue, Wondercraft, or Gemini Notebook's Audio Overviews.
Write for the ear. Eleven v4 and v3 don't support SSML break tags, so pauses come from punctuation, and tags like [laughs] or [whispers] set the delivery.
Only clone a voice with the owner's consent. ElevenLabs and Descript both require it, and Descript makes you record a consent statement.
Mix to about -16 LUFS with true peaks at or below -1 dBFS, the loudness Apple Podcasts asks for.
Apple Podcasts now requires you to disclose AI voices in the audio and the metadata, and listeners are wary. In a 2025 Sounds Profitable survey, 47% said they'd be less likely to keep listening to a favorite show that featured AI-generated voices.
What you need
You need a script or source material, a text-to-speech or dialogue tool, an audio editor, and a podcast host. Most of the tools below have free plans that are enough to make a test episode, but commercial rights usually start on a paid plan. ElevenLabs grants them from its $6 Starter plan, Speechify Studio from its $100-a-year Starter plan, and Wondercraft from its $25 Creator plan.
Step 1: Pick a format
There are three common shapes for an AI-voiced show.
A solo narrator reads a script, like a newsletter turned into audio. Any text-to-speech tool works, including ElevenLabs Studio, Descript, Speechify Studio, and Cartesia.
A two-host conversation has two synthetic voices discussing a topic or a document. ElevenLabs Text to Dialogue and GenFM, Wondercraft, Gemini Notebook's Audio Overviews, and the open-weights Dia2 model all handle this.
A cloned-host show uses a clone of your own voice, either reading scripts or paired with a real guest recording. Voice cloning is built into ElevenLabs, Descript, Wondercraft, and Cartesia.
If you're turning a document into a quick show, Gemini Notebook is the fastest route. Its Audio Overviews come in four formats, Deep Dive, Brief, Critique, and Debate, in more than 80 languages, though you can't pick or clone the hosts.
Step 2: Write a script for text-to-speech
AI voices read exactly what you give them, so write the way people talk. Keep sentences short, put names in full on first mention, and spell out numbers that a model might misread. ElevenLabs' own guide notes that its Flash v2.5 model can read "$1,000,000" as "one thousand thousand dollars."
Control pacing with punctuation. ElevenLabs says its newest models, Eleven v4 and v3, don't support SSML break tags, so use commas, ellipses, and line breaks for pauses. Both models read audio tags in square brackets, like [laughs], [whispers], or [sighs], and ElevenLabs says tags work best when the delivery fits the voice you picked. For a conversation, give each line its own speaker. In ElevenLabs Text to Dialogue, each turn has its own text and voice, and a line that ends in a dash lets the next speaker cut in. Open-weights Dia and Dia2 use [S1] and [S2] to switch speakers.
Fix tricky words before you generate. Descript lets you respell names and acronyms, and Eleven v4 reads phonetic spellings in IPA.
Step 3: Choose or clone your voices
Stock voices are the easiest start, and every tool here has a library. Pick two voices that sound clearly different for a two-host show, so listeners can tell who's talking.
To use your own voice, clone it. ElevenLabs makes an instant clone from about 10 seconds of audio, and its higher-quality Professional Voice Clone needs the $22 Creator plan and a voice check to confirm the samples are yours. Descript makes you record a consent statement, in English, before it will clone a voice, and it won't clone someone who's died or hasn't agreed.
Never clone someone else without their written consent. ElevenLabs' use policy bans cloning a living person's voice without explicit, informed consent, and Tennessee's ELVIS Act, in effect since July 1, 2024, protects a person's voice, real or simulated, against unauthorized use and distribution.
Step 4: Generate the episode
For a narrated show, paste your script into a project, assign voices, and generate a section at a time so you can redo a line without regenerating everything. ElevenLabs Studio is free to use and keeps long scripts organized in chapters.
For a conversation, use a dialogue tool. ElevenLabs' GenFM, under Studio, turns an uploaded document or URL into a host-and-guest script you can edit before it generates the audio. Wondercraft plans, writes, and edits episodes with music and chapters, and its Creator plan, at $25 a month, removes the watermark. Gemini Notebook generates a finished conversation from your sources, and you can pick the format, language, and length and steer it with a custom prompt, but you can't choose the hosts.
If you'd rather run it yourself, Nari Labs' Dia2 is free under Apache 2.0 and generates up to two minutes of English dialogue at a time.
Step 5: Edit and mix
Bring the voice tracks into an editor, add intro music and transitions, and cut anything that sounds off. Descript is the easiest for this, because you edit the episode by editing its transcript, and its Studio Sound cleans up any real recordings you mix in. Adobe Podcast's Enhance Speech and Async's Magic Dust do the same for a recorded guest.
Then set the loudness. Apple asks for about -16 LKFS, within 1 dB, with true peaks no higher than -1 dBFS. Many producers use -19 LUFS for mono files, and Auphonic, which is free for two hours of audio a month, can set both automatically. Export an MP3 or M4A for your host.
Step 6: Publish and disclose
Upload the episode to a podcast host, which creates the RSS feed that Apple Podcasts, Spotify, and YouTube read. Then check each platform's AI rules.
Apple Podcasts guideline 1.11 requires you to disclose AI-generated voices, AI hosts, and replicas of real people, in both the audio and the metadata of each episode and the show.
Spotify doesn't publish an AI disclosure rule for podcasts, but its policy bans impersonating another creator or host without permission.
YouTube requires creators to label realistic synthetic content, but it exempts cloning your own voice for a voiceover.
The simplest approach is to say it at the top of each episode and in the show notes. Inception Point AI, the top publisher of new podcasts in one recent count, has its AI hosts say they're AI at the top of the episodes.
How big are AI podcasts?
Big by volume. Podcast Index data reported by Gizmodo in May 2026 showed 35.4% of the new feeds added in one day were AI-generated, and Inception Point AI alone published 23.6% of all new podcast output. TheWrap reported in November 2025 that the company was producing about 3,000 episodes a week at about $1 each. Audiences are less keen. Sounds Profitable's April 2026 follow-up found 48% of mostly audio podcast listeners less likely to keep listening to a favorite show with AI voices, against 15% more likely.
Tools to skip
Hume is shutting down its Octave text-to-speech API on November 13, 2026, and PlayHT's platform reportedly closed at the end of 2025 after Meta acquired the company. Microsoft pulled the code for its VibeVoice long-form podcast model in September 2025, and the models that remain are for research only.
Frequently asked questions
Can you make a podcast entirely with AI?
Yes. Tools like ElevenLabs GenFM, Wondercraft, and Gemini Notebook can turn a document into a two-host episode, and text-to-speech tools can read any script. You still need to edit, mix, publish, and disclose the AI voices on Apple Podcasts.
What is the best AI voice tool for podcasts?
ElevenLabs, for most people. Text to Dialogue handles any number of speakers and reads audio tags for delivery, and an instant clone needs only about 10 seconds of your voice. Wondercraft is the easiest all-in-one podcast app.
Can I use Gemini Notebook Audio Overviews as a podcast?
You can make them easily, but Google's terms only say it won't claim ownership of what you generate, and no Gemini Notebook page grants an explicit right to publish them, so check before using one commercially.
Do I have to disclose AI voices in a podcast?
On Apple Podcasts, yes, in the audio and the metadata. YouTube requires labels for realistic synthetic content, and it's good practice everywhere.
Is it legal to clone someone's voice for a podcast?
Only with their consent. Voice tools ban cloning without permission, and laws like Tennessee's ELVIS Act cover simulated voices.
What loudness should a podcast be?
About -16 LUFS for stereo, with nothing peaking above -1 dBFS, which matches Apple's spec. Many producers use -19 LUFS for mono.