Should your podcast use an AI voice?

Your podcast does not need an AI voice. If you already record well and have the time, keep recording, and let a clone of your own voice fix small spots and read voiceovers. If you have something to say and no show yet, a synthetic version of your own voice is a great option: it takes the recording sessions off your week. Keep interviews recorded. If you use one, clone your own voice, say so at the start and have a person listen to every episode.

Narrated by a synthetic version of Lam Hoang's own voice.

Summary

• You don't need a synthetic voice. If you record well and have the hours, keep recording; a clone of your own voice can still patch a word or read a voiceover.
• With something to say and no show, a synthetic version of your own voice can get it made. Interviews and co-hosted chats stay recorded.
• In a 2026 blind test of short radio spots, most listeners who heard an AI voice took it for human. Finding out later is what costs trust.
• Disclose at the start, in one plain line that says what the machine does and who is in charge. Apple's rules require it in the audio and in the episode and show details.
• Clone only your own voice, record the training audio well, keep the raw files, and have a person listen to every episode before it goes out.

Which case is yours?

It depends on two things: whether listeners already know your voice, and whether you have the hours to record.

Two in three weekly podcast consumers say they would react badly if their favorite host switched to an AI version of themselves. That figure comes from a question about a host people already love. Edison Research at SSRS added it to Edison Podcast Metrics in the first quarter of 2026 and published the answers as Hands Off My Host, AI ↗︎. It asked US weekly podcast consumers how they would feel if their favorite host started using an AI-generated version of themselves to host the show. 66% picked at least one negative response, 27% picked any positive one, and 18% said they would be indifferent. Respondents could pick more than one answer. Deceived and upset tied at the top, at 33% each, then skeptical at 32% and betrayed at 28%, according to Radio Ink's report on the survey ↗︎, which gives the base as listeners aged 13 and older.

Three limits sit beside that figure. It is a hypothetical switch by a known host, not a show that says it uses a synthetic voice from its first episode. It measures stated feelings, not behavior: nobody was asked whether they would stop listening. And Edison's page gives no sample size or margin of error. Edison sells podcast measurement.

An earlier survey points the same way, with a middle. Sounds Profitable's Podcast Landscape 2025 ↗︎ was fielded in June 2025 with 5,034 Americans aged 18 and older on one online panel, weighted to the Census. In Tom Webster's write-up of the AI question ↗︎, 47% of people who had ever consumed a podcast said they would be less likely to keep listening if a favorite show featured AI-generated voices in its content or its ads, and 21% said more likely. Webster's own reading, in a column two days later ↗︎: there is "a sizable middle ground of people who just... don't care that much". The study's sponsors and research partner are audio companies.

The cost is real. I think it falls hardest on shows people already love. Nobody has measured a new show that is open about its synthetic voice from day one: the research for this episode looked for that study and did not find one.

Without a show, the question is time. The episode on how to start a podcast for your business gave Dave Jackson's test: time one episode from start to finish, then ask honestly whether you have those hours every week. If you don't, a synthetic voice changes the answer. It takes recording sessions and retakes off your week. Writing for the ear and listening to every render stay on your list.

No study has counted the hours a synthetic voice saves. The nearest figure is self-reported. In The Podcast Host's 2025 survey of independent podcasters ↗︎, most of the 500-plus respondents said an episode takes 1 to 3 or 4 to 5 hours, and 45% of audio-only creators said more than six. The sample is self-selected, and it does not separate recording from writing and editing.

A clone fits best where you write every word: a solo show, a scripted segment, a voiceover. In an interview or a chat with a co-host, the live exchange is the show, so record those. In the listener responses behind Edison's Evolving Ear research ↗︎, presented in January 2026, hosting, narration, banter and improvisation were described again and again as the "human core" of podcasting. Richard Berry of the University of Sunderland studied Virtually Parkinson ↗︎, a 2025 series in which a clone of the late Michael Parkinson interviewed living guests. He found it "struggles to reproduce key aspects of human interaction, such as turn-taking, humour, and natural empathy". That is one series, a paper in press read from its abstract, and a result of its time.

Even a recorded show can use a clone for fixes:

• Patching a word. Descript's Regenerate ↗︎, paired with a custom voice made from your own, lets you type the corrected words and generate them in place "to fix a flubbed delivery, a misspeak, or a wrong spoken word". It needs your own custom voice, works on English audio only and takes up to 250 characters at a time. Descript builds that voice from a consent read and a short passage, per its custom AI speaker guide ↗︎. The research found no other editor that patches a word in place, so check yours.
• Voiceovers. YouTube's disclosure rules ↗︎ list cloning your own voice to create voiceovers or dubs among the uses that need no AI label. Apple's rule, below, has no such exemption.

Will listeners notice?

The audio version of this page is narrated by a synthetic version of my voice, so you can judge for yourself. In short clips, many people can't tell.

Crowd React Media, the research division of the radio consultancy Harker Bos Group, ran a blind test ↗︎ in May and June 2026 with 1,326 US weekly radio listeners aged 18 to 45. Half heard a human voice actor and half an AI voice reading the same scripts: a station spot with a gas-card giveaway and a Mother's Day call to action. Overall appeal was virtually identical, 60% for the human voice and 61% for the AI one on the first clip. Across both clips the two scored comparably on professionalism, authenticity, credibility, energy and likability. "Funny" was the only significant gap, 33% against 26%, in the human voice's favor.

Detection was near chance. Of those who heard the AI voice, 55% took it for human; of those who heard the human voice, 59% called it human, by Radio Ink's account of the study ↗︎.

Told afterwards, a majority of the AI group felt nothing either way, 20% felt worse and 25% felt better. The researchers' verdict: "Disclosure is the real variable." Their respondents' comments, they wrote, "aren't saying 'it sounded fake.' They're saying 'it felt like lying.'" The limits: short radio spots, not podcasts or long narration; nobody was warned first; and the study comes from a firm that sells research and advice.

Finding out later can break trust. In Edison's Evolving Ear study of 2025 data ↗︎, as Radio Ink reported it, listeners were open to AI for background tasks such as scripting help, editing and transcription. Several said they would feel "cheated" or "robbed" if they discovered a host they believed was human was actually AI, "even if the illusion had worked". These were comments shared in a webinar, with no count.

How do you tell them?

Start with what Apple requires. Its content guidelines ↗︎, section 1.11, read on 10 October 2026: "Creators using AI to generate audio or video content - including synthetic voices, AI-generated hosts or on-screen personas, or AI-synthesized replicas of real individuals - must prominently disclose this to audiences in the content and metadata for each episode and show." For an audio show, that means in the audio, and in the episode and show details. Apple may label or remove content that doesn't meet its guidelines.

The rule changed in 2026. Apple's guidelines as archived in May ↗︎ still spoke of "a material portion of the podcast's audio"; that wording was gone by 21 September 2026. Guides that quote the older wording are out of date, among them RSS.com's March 2026 guide to AI disclosure ↗︎, which is still useful on hosts that let you write an AI flag into your feed. A disclosure test site ↗︎ by the same author says a clone of your own voice reading your own script needs no label. Apple's current rule makes no such exception, so follow Apple.

Other rules point the same way:

• Acast. Its community guidelines ↗︎ require disclosure when AI is used to replicate a host's voice for an episode.
• The EU. Article 50 of the AI Act ↗︎ requires anyone deploying AI that generates or manipulates audio amounting to a deep fake to disclose it, from 2 August 2026. Travers Smith's summary of the Commission's guidelines ↗︎ lists voice cloning of a podcast's presenters as needing a label, and Lewis Silkin's summary of the voluntary code ↗︎ says audio-only content opens with an audible disclaimer in plain language. Whether a disclosed clone of your own voice counts as a deep fake is untested. This is not legal advice.

Where in the audio: at the start, in one plain line. Leaving it until later is worse, because anyone still listening has already given the show a chance. Say it, and move on. This show says it in the first sentences of every episode, and this page carries the same line beside the player.

What the line should say. Sacha Altay and Fabrizio Gilardi of the University of Zurich ran two preregistered experiments ↗︎ with 4,976 people in the US and the UK, published in PNAS Nexus in October 2024. An "AI-generated" label made people rate headlines as a little less accurate and share them less, whether the headlines were true or false and whoever wrote them, though the effect was a third of a "false" label's. The reason: people assumed labeled headlines "have been entirely written by AI with no human supervision". In the second study, the label cost nothing when people were told the machine had only improved clarity and style. It still cost trust when they were told it chose the topic and wrote the whole article.

The limits are wide. It tested headlines, text not audio, and the authors say they cannot tell which part of their definitions caused the effect. Other text studies point other ways: in a Swiss experiment ↗︎, disclosure briefly raised people's willingness to keep reading, and in a 2026 study in China ↗︎, a label lowered the credibility of correct science texts and raised that of misinformation.

So I'd say what the AI does and who is in charge. If your clone reads a script you wrote, say exactly that.

How do you make one?

AI workflows get their own episodes, starting with How do you make a podcast with AI?. This section is about the voice. The providers named here are named for their published guidance and their plans, as read on 10 October 2026; every claim about a product is the vendor's own.

The providers' rule comes first: clone your own voice, or one you have explicit permission to use. Cartesia's voice cloning page ↗︎ puts it plainly: "Only clone your own voice or a voice you have explicit permission to clone and use." ElevenLabs' voice cloning deep dive ↗︎ and Speechify's developer guide ↗︎ say the same. ElevenLabs' professional voice cloning guide ↗︎ goes further for its professional clone: "Even with their consent, you cannot clone someone else's voice." And Spotify's impersonation policy ↗︎ removes podcasts that impersonate another creator's likeness without permission, by AI cloning or any other method.

Then comes the recording you train it on. The same professional cloning guide warns that the clone copies everything it hears, noise and echo included: "the AI will clone everything about the audio". The recording checklist, from that guide, its beginner's guide to recording ↗︎ and other providers' advice:

• The best microphone you have, in your quietest room, with a pop filter, about two fists (7 to 8 inches) from your mouth. No background noise or music. A thick duvet helps an untreated room.
• One set-up throughout. ElevenLabs' tips for a professional-grade clone ↗︎: "Switching mic chains mid-recording confuses the model."
• One speaker, one consistent style, with the delivery you want back. From the same deep dive: "If your performance is flat, your model might bear the same lacking delivery." Cartesia's professional clone guide ↗︎: "Editing can tidy up silences, but it can't fix a flat take."
• Enough audio. For a professional clone, ElevenLabs' minimum is 30 minutes and it recommends "closer to 2-3 hours of audio". A professional clone also needs a live reading to verify the voice is yours. An instant clone ↗︎ needs about one to two minutes, and more than three adds little. It is the less faithful kind.

The cost: start with a monthly subscription. The first paid tier is enough to begin. At ElevenLabs and Cartesia, the professional clone is one tier up.

Source: ElevenLabs' pricing ↗︎, Cartesia's pricing ↗︎, Descript's pricing ↗︎ and Fish Audio's plans ↗︎, read on 10 October 2026. Prices, credits and promotions change, and a re-render spends credits again.

Our calculation: a weekly 15-minute show is about 60 minutes of speech a month before re-renders. That is more than ElevenLabs' first paid tier and within its second, and within Cartesia's and Fish Audio's first.

Free plans generally won't cover a business show, because what you make on them can't be used commercially. ElevenLabs' publishing rules ↗︎ say the free plan "cannot be used for any commercial purpose", and Fish Audio's terms ↗︎ limit free use to "internal, personal, non-commercial" use. A business podcast is commercial use. If you outgrow the first tier, you're probably getting a lot of value from it.

Next, the model. Stay current with the one you use, and listen again after any switch. ElevenLabs released Eleven v4 on 28 September 2026, per its changelog ↗︎. Moving an existing professional clone onto it is a step you take yourself, and its Eleven v4 page ↗︎ warns that "Eleven v4 may sound substantially different from Eleven v3". The same page says v4 copies a voice's accent and delivery more faithfully: the vendor's claim, untested. Cartesia, by its own account in the same professional clone guide, moves professional clones onto new models automatically.

Keep your raw voice recordings. They are how you train the next model. ElevenLabs' professional cloning guide says you cannot download or export a clone as a file, and advises keeping the original audio to recreate it later. Its terms of use ↗︎, updated 9 October 2026, make you responsible for exporting your content before your access to a service ends.

Companies change course, too. PlayHT's home page was still selling voice cloning ↗︎ on 1 December 2025. By 6 January 2026, the same page read ↗︎ "We have shut down the service." In August 2026, Resemble AI's chief executive wrote on the company's site ↗︎: "We're not selling voice AI to new customers."

What stays yours?

The voice is the easy part. In 2025, on the Podnews Weekly Review ↗︎, the Podnews editor James Cridland put the problem on the words rather than the voice: whether a person had checked them. His own example: "I've got a cold this week... maybe I'll use my AI clone... it was still my own voices and my own words." The quotes come from the episode's automatic transcript.

The biggest thing is a person in the loop. Someone listens to every episode and makes sure every citation is real. Build a workflow that makes that check happen; a second AI pass hunting for mistakes helps too. This show is made with the help of AI, and I listen to every episode before it goes out.

The case for the listen. In April 2026, Podnews reported ↗︎ that Inception Point AI, which was publishing hundreds of AI-made shows a day, had put out an episode that "ends with the words, in English, 'I'm sorry, I can't assist with that'". One reported case, not a rate. In 2025 the company's chief executive said there was "no way" it listened to the overwhelming majority of its content before release. A person listening would have caught it.

Make every render a good one. Cartesia's voice cloning page ↗︎ says to "Listen to the generated audio before adding it to your project, especially when the script contains names or specialist terms." ElevenLabs' terms of use ↗︎, as updated in October 2026, say: "You are responsible for reviewing Output before use or distribution."

Some things stay yours however good the tools get: your show's positioning, its direction and its taste, and the subtleties the model doesn't know to pick up on. Without your context and shaping, AI gives you its default style, which is close to what it gives everyone. Jill Walker Rettberg's 2026 study of one AI podcast tool ↗︎ found it rendered other languages in "a perky standardised Mid-Western American accent". That holds whether you make the show yourself or hire a podcast producer. And a cheaper voice doesn't excuse a thin script.

What to do this week

1. Time one episode from start to finish, and decide which case is yours: listeners who know your voice and the hours to record, or something to say and no hours.
2. Pay for one month of a provider's first paid tier, since free-plan output can't be used commercially, and make a quick clone from a couple of minutes of your voice.
3. Have it read a page you wrote, and listen, hardest at names and specialist terms.
4. Write your disclosure line: one plain sentence for the start of each episode that says what the machine does and who is in charge. Put it in the episode and show details too.
5. If you go further, record the training audio once, properly: best microphone, quietest room, one set-up, the delivery you want back. Keep the raw files.
6. Name the person who listens to every episode before it goes out.

FAQ

Do you have to disclose an AI voice on Apple Podcasts?

Yes. As of October 2026, Apple's content guidelines say creators using synthetic voices must prominently disclose it in the content and metadata for each episode and show. For an audio show, say it in the audio, at the start, and in the episode and show details.

Can listeners tell an AI voice from a human one?

In short clips, often not. In Crowd React Media's 2026 blind test of radio spots, 55% of the listeners who heard the AI voice took it for human, and the two voices scored about the same on credibility and likability. That test used short spots, not long podcast episodes.

Can you clone someone else's voice for your podcast?

Only with their explicit permission, and some providers won't do it even then. Cartesia, ElevenLabs and Speechify all say to clone only your own voice or one you have permission to use, and ElevenLabs makes professional clones of your own voice only. Spotify removes podcasts that impersonate another creator's likeness without permission.