Voice Cloning vs Voice Actors: What a Course Costs

Prices and rates checked on 30 September 2026.

Can a clone of a teacher's voice teach as well as the teacher? In a randomised trial at a French medical school, published in 2026, it made no measurable difference: students who heard the teacher's own recording scored 51.8 out of 100, and students who heard a clone of the teacher's voice scored 51.2 (JMIR Medical Education, July 2026). On cost the gap is wide. A professional US e-learning narrator charges about USD 30 to 55 per finished minute (GVAA rate guide, 2024 edition). An AI voice costs cents per minute, and the gap grows every time the script changes or needs another language.

But price should rarely decide it. Who is listening decides it. If your learners know your voice, a clone of it is the one option they will notice. If they need another language, they need a native voice in the accent they will meet, not your clone. If the performance is what you are selling, hire an actor. For most lessons and training modules, the best choice is neither your clone nor an actor. It is a ready-made library voice.

voice recording vs text to speech software

We make a voice cloning tool. Keep that in mind as you read: this article often advises against using it. A few figures come from archived pages or from a 2024 rate guide; How we checked these numbers says which.

Four options, side by side

Short answer: if you teach or train, you can record the material yourself, hire a narrator, use a ready-made library voice, or clone a voice. For a 20-minute module in one language, a narrator costs roughly USD 500 to 1,100. On Kukarella, a library voice uses about $1.26 of credits, at the pace we assume below, and the Prime plan includes one clone of a voice.

A library voice is a ready-made AI voice: paste your script, pick a voice, generate. Kukarella has 2,000+ voices, many of them native speakers of one language in a regional accent you choose. A clone is an AI copy of one real person's voice, made from a recording: yours, or that of someone who has agreed.

Record it yourself

• Money for 20 minutes: your own equipment.
• Your working time: about 70 minutes of active work, including script drafting and editing, at the Lyon trial's 35 minutes per 10-minute video.
• Wait for the first audio: none.
• A 10-sentence change later: re-record, in the same room and on the same mic.
• Other languages: only the ones you speak natively.
• Best when learners know you and your presence matters. Avoid when the script changes often.

Hire a narrator

• Money for 20 minutes: USD 500–1,100 (sources in Step 1 below).
• Your working time: a brief, then checking the files.
• Wait for the first audio: 24 hours on average on Voices.com (its own figure); 6 days on average on Fiverr.
• A 10-sentence change later: about $100–550, plus a new booking (Step 2 below).
• Other languages: one narrator per language.
• Best when the performance matters, or you want one particular person. Avoid when the script changes every term.

Use a library voice

• Money for 20 minutes: about $1.26 of credits*.
• Your working time: generating, then listening to all 20 minutes once.
• Wait for the first audio: usually 1 to 3 minutes for 20 minutes of audio, depending on the voice**.
• A 10-sentence change later: about $0.35*.
• Other languages: a native voice for each language.
• Best when you make lessons, procedures, dialogues and language work. Avoid when you need one particular real person.

Clone a voice

• Money for 20 minutes: credits for the speech, plus the clone (1 included in Prime, extra $5).
• Your working time: a clean sample, then as for a library voice.
• Wait for the first audio: under a minute to make the clone from a recording, about 30 seconds from a description**, then as for a library voice.
• A 10-sentence change later: only the changed paragraphs are charged again.
• Other languages: it tends to carry the speaker's accent.
• Best when the voice is the point and listeners do not know it. Avoid when your listeners know the voice.

* Kukarella Prime costs $21 a month for 300,000 credits: $0.07 per 1,000. A library voice uses 1 credit per character (home page FAQ). We assume about 900 characters per finished minute (150 words a minute, about 6 characters a word including the space): about $0.06 a minute. We count the whole plan price as the price of credits, which slightly overstates the cost.

** Our own timings on Kukarella, September 2026. Generation time depends on the voice and the length of the text.

AI voices save money, not all of your time. ElevenLabs says of its own models: "Our models are non-deterministic, meaning outputs can vary based on inputs" (ElevenLabs troubleshooting). So listen to everything once, fix names and terms, and regenerate the lines that came out wrong.

Does a cloned voice teach as well? The Lyon trial

Short answer: in the one randomised trial we found on teaching, the researchers detected no difference in learning between students who heard a teacher's clone and those who heard the teacher's real recording, and the clone saved production time. The study is small, and it did not measure money.

At Université Claude Bernard Lyon 1, medical students watched four 10-minute recorded lectures (Gavoille, Subtil and Nourredine, 2 July 2026). Only the voice changed: the teacher's recording, or a clone trained on 30 minutes of the teacher's audio with ElevenLabs' professional voice cloning. Researchers randomised 88 students and analysed 64.

The authors: "We did not detect a significant difference in learning outcomes or satisfaction". It is one small group in one country. The production time leaves out the minutes the software took to generate each track. The paper does not make clear whether the recording that trained the clone is counted. The authors also warn that "some students may perceive synthetic voices as less authentic". And the study stopped after the first recording, while a real course gets updated and translated. That is where most of the cost is.

One course, followed over its life

Short answer: in our illustration, voicing one 20-minute module through a first recording, a policy update, three more languages and a second update costs about USD 2,500 to 7,150 with human narrators. With AI voices on Kukarella, the same work uses about $7 to $20 worth of credits.

An illustration built from published rates. It is not a customer.

Imagine a company that trains new staff with a 20-minute module on handling customer data: about 3,000 words at 150 words a minute, the pace Voice123 uses for its rates, or about 18,000 characters. It will be used for years.

Step 1: the first recording, in English

What a human narrator charges for 20 finished minutes, by source:

So a professional narrator for this module costs roughly USD 500 to 1,100. Budget marketplaces go lower, with no quality floor: Fiverr's average voice-over job, of any length, was $73.

With a library voice on Kukarella, the same 18,000 characters use 18,000 credits: about $1.26. That is at $0.07 per 1,000 credits and 900 characters a minute. If you use a clone of your own voice instead, add the clone: one is included in Prime, and an extra one is $5.

Step 2: a policy changes

Seven weeks later, the company changes its rules. Ten sentences in the script must change.

So a human update costs roughly $100 to $550, per language, per change. The larger costs are the booking, the wait, and the risk that your narrator has moved on: a new narrator may mean re-recording the whole module.

Flowchart comparing the complex, multi-stakeholder traditional voiceover workflow to the simple, linear AI voiceover workflow

On Kukarella, making the same text again with the same voice costs nothing. You pay again only for the paragraphs you changed. If the ten sentences sit in ten paragraphs of about 80 words, that is about 5,000 characters: about $0.35 in a library voice.

Hear it: the same procedure before and after a change to its third sentence, made in Kukarella with the same library voice.

Step 3: three more languages

Now the company hires staff in France, Spain and Germany, and the module needs French, Spanish and German versions.

With humans, each language needs its own narrator, booking and wait. At US rates, that is 3 × USD 500–1,100 = USD 1,500 to 3,300. That is our assumption: the sources above give no rates for French, Spanish or German narrators, and local rates differ. Translation costs the same whichever voice reads it, so we leave it out of every route.

With AI voices, you pick a native library voice for each language: 54,000 credits, about $3.78 (character counts differ a little between languages). If your clone read the three new versions instead, it would tend to carry your accent into French, Spanish and German. We explain below why that matters for learners.

Step 4: the next update, now in four languages

The next policy change hits all four versions. With humans: 4 × $100–550 = $400 to $2,200, and four bookings. With library voices: 4 × 5,000 = 20,000 credits, about $1.40.

The total

The whole library-voice route fits inside the 300,000 credits of one month of Prime. If you narrate the English version with a clone of your own voice, the clone included in Prime covers it. But the steps happen months apart, and you pay $21 for each month you keep the plan, so the plan fee is your real cost. Even for a whole year of Prime (12 × $21 = $252), the human route costs about 10 to 28 times as much. The saving is not on day one. It is on every update and every language after it.

A second example: a 2-hour compliance course

Two hours is 120 finished minutes, about 108,000 characters, in one language.

Work out your own course

Your course is not our illustration. Put in your own numbers; if a narrator quoted you a different rate, type it in. The calculator shows every number it uses, with its source.

Is it cheaper to clone a voice actor than hire them repeatedly?

Short answer: per minute, yes, by far: cents against USD 30 to 55 per finished minute (GVAA, 2024 edition). But you cannot clone an actor without their consent and a licence, and the licence is the expensive part. For a course with a few updates a year, paying the narrator for pickups costs less.

Gravy For The Brain's UK rate guide lists a company's licence to use a narrator's voice as an AI voice at £8,000 (typical lower) and £15,000 (typical average), with 100k+ by negotiation at the top. Its Canadian guide lists CAD 12,000 and 20,000.

An illustration, not a customer: a training team books 20 finished minutes a month at the $35 average in GVAA's 2024 rate guide. That is $700 a month, or USD 8,400 a year. A Canadian licence at CAD 12,000–20,000 (about USD 8,460–14,100) could pay for itself in one to two years (our arithmetic). Below that volume, keep hiring the narrator. But Gravy For The Brain notes that the fee usually buys a set number of words, and any extra dialogue is paid for on top.

The actors who say their voices were taken

Paul Lehrman and Linnea Sage are New York voice actors. According to their lawsuit, each was hired through Fiverr to record scripts said to be for research or internal use only. Lehrman was told they "will not be used for anything else". He was paid $1,200, she $400. They allege their clones were later sold by Lovo, an AI voice company, as "Kyle Snow" and "Sally Coleman". On 10 July 2025 a federal judge ruled that copyright "does not extend to something as abstract and intangible as a 'voice'". He let their New York publicity, consumer-protection and contract claims go ahead. He also let Sage's copyright claim over the use of her original recording go ahead. We found no final outcome as of September 2026.

The lesson for anyone who hires a narrator: a recording fee is not a licence to clone. If you want to clone the narrator you hired, ask them, agree what the clone will be used for, and pay for it.

What a licensed voice looks like

ElevenLabs makes its professional clone only of your own voice: "Even with their consent, you cannot clone someone else's voice" (ElevenLabs documentation). People who want to earn from their voice can list it on its Voice Marketplace. ElevenLabs says 10,400+ creators have earned over $22 million there (updated September 2026). That is about $2,100 each by our division; ElevenLabs publishes no median.

The e-learning recording that ended up on trains

In 2021 the Scottish voice artist Gayanne Potter recorded for ReadSpeaker, a speech company, work she understood was for accessibility and e-learning software (Wikipedia, citing the BBC). In May 2025 ScotRail launched an AI announcer, "Iona", which she said was built from those recordings: "I did not know. I was not asked. I did not consent." (STV News, 5 June 2025). From December 2025 ScotRail used a new voice based on a member of its own staff, who recorded 18 studio sessions for it. That is the right pattern for an organisation: a clone of someone who said yes, for this use.

Consent and the law, in brief

The speed and price that help a teacher also help a fraudster. The man who made the fake Joe Biden robocall sent to New Hampshire voters in January 2024 told NBC News it took under 20 minutes and cost $1. Weeks later the FCC ruled that AI-generated voices count as "artificial" under the US law on robocalls. The consultant behind the call was fined USD 6 million, which he has refused to pay. Tennessee's ELVIS Act, signed on 21 March 2024, added voice to the state's personal-rights law. The federal NO FAKES Act passed a Senate committee on 18 June 2026 but is not law.

Kukarella's Terms of Use require "all necessary rights, licenses, consents, and permissions for any audio samples you upload" and forbid impersonation. More in our guide to voice cloning ethics and legal rights.

Before you clone: use your own voice, or get written permission that names the use, the period and the pay, and keep it.

AI voice cloning vs text-to-speech: which do you need?

Short answer: text-to-speech reads your text in a ready-made voice from a library. Voice cloning first copies a real voice from a recording, then reads your text in it. Text-to-speech no longer has to sound robotic; the difference today is whose voice it is. Use a clone only when that person's voice is the point.

Does a clone sound more human than a library voice? In one 2025 study, listeners labelled clones "human" 58% and 70% of the time in two experiments, close to the 62% and 72% for real recordings. In a separate experiment of the same study, a vendor's stock AI voices were called human only 41% of the time, against 81% for real voices (Lavan and colleagues, PLOS One, 24 September 2025; 50 people in each experiment). Those stock voices were not Kukarella's.

For teaching, sounding human matters less than you might expect. In an older randomised study with an online sample, learners rated a modern text-to-speech voice as credible and as helpful for learning as a human voice (Craig and Schroeder, 2017). And many lessons need more than one voice: a listening task or a role-play needs two or more speakers, which one clone cannot give.

Screenshot of Kukarella showing a hybrid project with a cloned narrator voice and multiple different stock AI voices for other characters

Who is listening decides it

Short answer: if your listeners know the voice well, do not clone it. People who know a speaker can pick out that speaker's clone well above chance; strangers cannot. A clone earns its place when nobody listening knows how the real person sounds, or when the real voice can no longer be recorded.

Researchers at University College London cloned their participants' voices. Then they asked 47 participants to tell clones from real recordings of their own voice and a friend's voice, and asked 47 strangers to do the same (Rosi, Soopramanien and McGettigan, 2025). People who knew the voice told the clone from the recording well above chance. Strangers did no better than guessing. The familiar listeners also found the clones "less (or similarly) trustworthy, attractive, and competent" than the recordings.

Strangers are easy to fool: in another 2025 study, listeners judged an AI voice to be the same speaker as the real person about 80% of the time (Barrington, Cooper and Farid, Scientific Reports). It matches what we tell users on our home page: "learners who know your voice hear the difference, and a voice that is plainly not you often works better in a lesson."

We see the same thing in how people use Kukarella. Many of our users clone their own voice, and some keep using the clone. But over time many of them move to library voices. An exact copy of your own voice is still hard to make, the people who know you hear the difference, and that takes away the reason to clone yourself in the first place.

So if you teach your own class, cloning yourself is usually the weakest option: your students are the one audience that can tell. Record it yourself when your presence is the point. Use a library voice when the material is the point. And if you do clone yourself, play the clone to a colleague who knows your voice before your students hear it.

Hear the difference yourself. The same paragraph three ways: Nazim Ragimov reading it, his clone made in Kukarella, and a native English voice from the library.

Where a clone earns its place

A clone is the right tool when the real voice can no longer be recorded, or when the listeners do not know it: a voice you have lost, a character you invented, or a colleague who has agreed but cannot record every update.

The first case is the clearest. Ed Riefenstahl directs MBA experiential learning at the Neeley School of Business at Texas Christian University. After a fall in July 2021 and a brain injury, he developed bulbar palsy, which attacks speech. In August 2024 he received a clone of his voice built from old recordings (AACSB Insights, 9 April 2025). Now he types, and the device speaks in his voice. He went back to coaching his students, and told them: "It felt like a part of me had been restored."

If your learners need another language or accent

Short answer: if your learners must hear a native accent, use a native voice for each language. A clone of you tends to carry your accent into other languages, and cloning tools are not reliable with accents.

Language teachers care about this most. In a British Council survey, 80% of 1,112 English teachers in 115 countries and regions agreed that AI should support different varieties of English, such as local pronunciation; only 5% disagreed (July 2024, page 38). One teacher wrote: "students need to be exposed to global Englishes"; another, "BBC English is enough".

A clone is built to sound like one person, accent included, and cloning services say plainly that accents are a weak point:

• Descript says its clones use a model based on US English pronunciation, so "accents may not always be preserved accurately" (Descript help).
• ElevenLabs says its instant clone "may struggle with uncommon accents or unique voices" (ElevenLabs documentation).
• The UCL researchers found their cloning tool sometimes gave a speaker the wrong accent, for example a British speaker cloned with an American one, and regenerating did not always fix it (Rosi and colleagues, 2025, page 3).

For example, a new hire who will work with colleagues in Monterrey needs to hear Mexican Spanish, not your English accent in Spanish. A native library voice gives them that. This is what Kukarella's voice library is for: voices in the accent your learners will meet.

Hear it in Spanish: the same sentence read by Nazim's clone, then by a native Mexican Spanish voice from the library.

Two practical routes:

1. Narration. Write the lesson once, ask the built-in assistant to translate it, have a native speaker check it, then choose a native voice for that language and region in Text to Voice. It is a manual route, not one-click dubbing.
2. Role-plays. In DialoguesAI, describe a conversation; the assistant writes it and gives each speaker a voice. The first cast is not always in the accent you need, so switch speakers to native voices by hand.

When human narrators are still the right call: a flagship course that rarely changes, or a regional variety that voice libraries hold thinly. Check the library for your variety first.

If you train staff in Quebec: the Charter of the French Language, as amended by Bill 96, requires employers to make the training documents they produce for staff available in French, on terms at least as favourable as any other version (BLG, 2022). Whether recorded narration counts is not settled in anything we read; ask your counsel.

AI voices or voice actors for commercials, tutorials and films?

Short answer: tutorials and training suit AI voices; films suit actors; commercials depend on how much the performance sells the product. In three cases a human wins whatever the price: your listeners know the voice, the performance is the product, or you want one particular person.

When the performance is the product, the difference is direction. Ask an actor for a line warmer or slower, and you hear it in the next take. An AI voice gives you a new take each time you generate, sometimes better, sometimes not. And cloning is not a way to hire an actor without hiring them.

A performance is affordable: an explainer video of up to 90 seconds for your website and social channels is $300–525 (GVAA, 2024 edition, not including paid advertising), small next to what the video must do. And an actor's AI replica is not cheap either: in US union commercials it triggers a 1.5 times session fee, plus holding and use fees.

Some voice actors welcome paid, agreed AI use. Cissy Jones won a BAFTA Games Award in 2017 for Firewatch. In a 2023 interview she called paid uses helpful, such as paying her for placeholder "scratch" lines in a game and bringing her in live "when we actually need emotion". Her condition was consent: "They need to get active consent" (Screen Rant, 4 August 2023). By October 2023 she was a co-founder of a start-up making paid, approved digital copies of actors' voices (Bloomberg, via Taipei Times, 15 October 2023). It is a good rule: AI for the scratch work, the actor for the emotion.

What to do next, by situation

• Your course changes every year. Pick one library voice and keep it for every module and update. Start in Text to Voice.
• You localise. Write once, translate, have a native speaker check it, then choose a native voice for each language and region.
• Your learners know your voice. Record the personal parts yourself; use a library voice for readings and instructions. If you clone yourself, use a long, clean recording and test it on a colleague first.
• The performance must move people. Hire an actor and agree pickup terms before the session. Use an AI voice for the draft, to time the video.

How much voice cloning costs per minute, tool by tool

Short answer: in the self-serve apps below, speech in a cloned voice costs about 7 to 18 cents per finished minute, and paid cloning starts at $6 a month. Developer APIs cost less per minute. Custom voices from Google, Microsoft and OpenAI are only for approved customers.

Checked on 30 September 2026. List prices, no promotions, no retakes. Per-minute costs are our arithmetic at 900 characters per finished minute.

ElevenLabs, instant clone

• Plan we priced: Starter, $6 a month, 30,000 credits.
• Audio it needs: at least 1 minute. Ready in: not stated.
• Consent check: you confirm you have the right and consent.
• Cost per finished minute: about $0.18.

ElevenLabs, professional clone

• Plan we priced: Creator, $22 a month, 121,000 credits.
• Audio it needs: at least 30 minutes, ideally 2–3 hours. Ready in: "generally" 3–6 hours, sometimes longer.
• Consent check: spoken verification; your own voice only.
• Cost per finished minute: about $0.16.

Speechify Studio

• Plan we priced: Studio Starter, $100 a year, 86,400 credits (1 credit per second of audio).
• Audio it needs: a 20-second recording. Ready in: not stated.
• Consent check: its API made a spoken consent phrase mandatory on 23 September 2026.
• Cost per finished minute: about $0.07, if you use the whole year's credits.

Descript

• Plan we priced: Hobbyist, $16 a month billed yearly or $24 monthly, for full use; the Free plan includes limited clones.
• Audio it needs: the consent statement you read aloud (63 words, about 25 seconds). Ready in: typically within minutes.
• Consent check: your recorded consent statement.
• Cost per finished minute: not published.

Kukarella

• Plan we priced: Prime, $21 a month, 300,000 credits, 1 clone included; extra clone $5.
• Audio it needs: we recommend at least 15 seconds; the app accepts up to 5 minutes. Ready in: under a minute from a recording; about 30 seconds from a description.
• Consent check: our Terms require the rights and consents; there is no spoken verification.
• Cost per finished minute: see Pricing (a library voice is about $0.06).

Kukarella: clones come with every paid plan; Prime is the plan priced here.

ElevenLabs counts about 1,000 characters as a minute, so the minutes on its own price table are about 10% fewer than ours for the same credits.

Outside this table: Google's Instant Custom Voice is for allow-listed users; Microsoft's Personal Voice for approved use cases (a 5–90-second sample plus a spoken statement) and its professional voice for customers managed by a Microsoft account team; OpenAI's custom voices for "eligible customers". Murf clones only on its Enterprise plan, in 1–4 weeks. Resemble's cloning API needs a Business plan or higher; its pricing page lists a plan named Business at $1,000 a month. WellSaid does not let users clone.

Two notes. If you do not need a particular voice, do not pay for one: a library voice is the cheaper and safer default. And a clone lives in the vendor's account. In July 2025 Meta acquired Play AI, the company behind PlayHT (TechCrunch). Inworld, a competitor, says in a migration guide that PlayHT closed on 31 December 2025 and that its voice clones were deleted with no way to export them (Inworld). Keep your scripts and original recordings. They are the part you own.

How much does voice cloning cost on Kukarella?

Short answer: checked on 30 September 2026, voice cloning comes with Kukarella's paid plans. Prime costs USD 21 a month, includes 300,000 credits and one cloned voice, and an extra clone costs USD 5. Speech is paid in credits; a library voice uses 1 credit per character.

* At about 900 characters per finished minute (150 words a minute), treating the whole plan price as the price of the credits.

How much audio does Prime cover? It depends on the language and on how fast the voice speaks. At 130 to 170 words a minute, 300,000 credits make about 5 to 6½ hours of finished audio in a library voice (our arithmetic, at about 6 characters a word).

How you make a clone

Open Clone-a-Voice from the Applications menu and press Clone a Voice. There are two ways in:

• Clone with Audio. Record in the browser (the screen suggests 10 to 60 seconds per take), or upload up to 10 files, 20 MB and 5 minutes in total (the app's limits). We recommend at least 15 seconds of clean speech; longer is better. The app warns you about background noise, because the clone copies the room with the voice.
• Generate with Text. Describe the voice you want in up to 500 characters and press Create Preview. The preview is free; you spend your clone only when you press Create voice. This is the route for a voice nobody knows, such as a character.

On a phone, the menu marks Clone Voice as "Desktop only", so plan to record at a computer.

Three pieces of advice from us:

1. Make your first clone from your best recording, not a quick test. A clone uses your clone credit, so treat it as permanent.
2. Record in the room you will always use. Background noise becomes part of the voice.
3. Clone only a voice you have the right to use. Our home page says it plainly: someone else's voice only "with that person's clear permission".

Where your voice is kept

Our Terms of Use say that voice samples and voice models are stored by our third-party voice-cloning provider under its retention policy, and that Kukarella does not directly control deletion. To ask for deletion, write to support@kukarella.com and we forward the request.

Questions people ask

How much do voice cloning platforms charge per minute of audio?

In the self-serve apps we checked on 30 September 2026, about 7 cents (Speechify, and only if you use a whole year's credits) to 18 cents per finished minute, at 900 characters a minute (see the list of tools). On Kukarella, a library voice costs about 6 cents a minute on Prime. Developer APIs are cheaper: Speechify's API Starter, $10 a month for 1.9 million characters, works out under 1 cent a minute. That is a route for programmers, not for a course author.

Are AI voiceovers cheaper than studio recording?

Yes. One finished minute from a US e-learning narrator costs $30–55 (GVAA, 2024 edition). The same $55 buys about 785,000 Kukarella credits at $0.07 per 1,000: about 7 hours of narration, if you count 2 credits per character and 900 characters a minute. Voices that use 1 credit per character give about twice that.

How long does it take to create a custom AI voice clone?

From seconds to weeks. An instant clone from a short sample is ready in seconds to minutes; Resemble says under a minute. A professional clone needs at least 30 minutes of audio and usually 3–6 hours of training; enterprise voices take weeks. The audio needed keeps falling: Apple's Personal Voice asked for 15 minutes of reading in 2023 and 10 phrases in 2025. On Kukarella, a clone from a recording is ready in under a minute, and one from a text description in about 30 seconds; what takes your time is a clean recording.

What is the difference between instant and professional voice cloning?

An instant clone copies a voice from seconds or minutes of audio, usually within minutes. A professional clone trains for hours on at least 30 minutes of audio.

Can I work directly with the voice actor behind an AI clone?

Sometimes. On Audible's ACX, narrators in a beta programme can offer an AI replica of their voice, choose title by title how they are paid, and Audible "will not separately use a narrator's voice replica for any content without their approval" (ACX, 9 July 2025). If you want that actor's own performance, hire them. Many narrators offer pickups, some free within 48 hours of delivery.

What does an AI audiobook cost compared with a narrator?

A US non-union narrator charges $200 to $500 per finished hour, production included (GVAA, 2024 edition), for about 6.2 hours of work per finished hour (ACX). A 10-hour book is about 540,000 characters: on Kukarella, about $38 of credits in a library voice, at $0.07 per 1,000 credits and 900 characters a minute. For a literary book, where the performance is the product, we would still hire a narrator.

How is AI voice cloning changing the voice-acting industry, and can I use it on a SAG-AFTRA project?

The fight has moved to consent and pay. US video game performers ratified AI protections on 9 July 2025 after a strike that began in July 2024, and the TV and film contract of 4 June 2026 adds terms that "further restrict use of synthetics". On a union project, use AI only on the union's terms, and ask the union first.

Will my learners notice that it is an AI voice?

Strangers often will not; people who know the real voice usually will (see above). Either way, accessibility rules ask for a text version of prerecorded audio (WCAG 2.2, Level A), and with text-to-speech you already have the script.

How we checked these numbers

We read every price and rate on 30 September 2026, on the vendor's, union's or narrator's own page. Two sources are not the rate-setter's own page: the SAG-AFTRA rate sheet is a copy hosted by the payroll company ART Payroll, and Gravy For The Brain is a training company that publishes rate guides. Voices.com and Fiverr block automated reading, so their figures come from the latest Internet Archive copies (July 2026 and November 2025). The GVAA's current rate guide requires a sign-up and forbids using it to build AI tools, so we quote its last public edition, from 2024; current rates may be higher. Studies were read at the source, except Craig and Schroeder (2017), which we read in abstract only. Kukarella facts come from our Pricing page, home page FAQ, Terms of Use and the product. Where a number is our arithmetic, the page says so. We did not test other vendors' clones.

A correction: an earlier version of this page gave cost ranges without sources and carried a quote attributed to the voice actor Cissy Jones that we could not trace to any source. We removed both. Her public position is described above.

We sell the cloning tool this article discusses, and we wrote it to be useful whichever option you choose.

About the author. Nazim Ragimov has been building Kukarella since 2019, first for himself and his team, then for the users whose feedback shaped almost everything in it. Most of them teach, and he wants Kukarella to delight them and their students: a lesson that makes the class say "Now we get it." He started these articles himself, with help from his wife, Anna; today he sometimes uses AI for research and first drafts.

Sources

All pages read on 30 September 2026 unless stated.

Studies

• Gavoille, Subtil, Nourredine. Voice Cloning Using AI vs Traditional Audio Recording for Prerecorded Courses in Medical Pedagogy: Randomized Controlled Trial. JMIR Medical Education, 2 July 2026, 12:e86569. PMC
• Lavan, Irvine, Rosi, McGettigan. Voice clones sound realistic but not (yet) hyperrealistic. PLOS One, 24 September 2025. Article
• Rosi, Soopramanien, McGettigan. Perception and social evaluation of cloned and recorded voices: Effects of familiarity and self-relevance. Computers in Human Behavior: Artificial Humans 4 (2025) 100143. PDF, UCL Discovery
• Barrington, Cooper, Farid. Scientific Reports 15:11004 (2025). PDF
• Craig, Schroeder. Reconsidering the voice effect when learning from a virtual human. Computers & Education 114 (2017) 193–205. Abstract
• Edmett, Ichaporia, Crompton, Crichton. Artificial intelligence and English language teaching, 2nd edition. British Council, July 2024. PDF

Human rates and turnaround

• Global Voice Acting Academy (GVAA) rate guide, 2024 edition, archived 17 September 2024. Wayback Machine; current guide, sign-up required: GVAA
• Voice123, voice-over rates reference budget, updated 2 June 2026. Voice123
• Voices.com, How it works, archived 27 July 2026. Wayback
• Fiverr, hire voice-over, archived 22 November 2025. Wayback
• SAG-AFTRA 2025–28 Corporate/Educational and Non-Broadcast Contract rates, updated 1 November 2025, copy hosted by ART Payroll. PDF
• UDA, XN rates from 1 December 2025. UDA; UDA–Xn Québec agreement 2022–2025. PDF
• Kim Handysides, eLearning Narration Rate Guide, page last modified 18 August 2022. Page
• ACX, Narrate an audiobook, 19 August 2025. ACX help; Narrator voice replicas beta, 9 July 2025. ACX blog
• Bank of Canada, daily exchange rates, 29 September 2026 (USD/CAD 1.4188). Valet API
• Gravy For The Brain rate guides, UK and Canada, undated. UK · Canada
• SAG-AFTRA help centre, AI protections in the 2025 Commercials Contracts, updated 18 August 2025. SAG-AFTRA

AI voice tools

• ElevenLabs: pricing · voice cloning · instant clone · professional clone · troubleshooting · $22 million earned by voice creators, 22 May 2026, updated 22 September 2026
• Speechify: Studio pricing · AI voice generator · API plans · API changelog, updated 23 September 2026
• Descript: pricing · create a custom AI speaker
• Resemble AI: voice creation · pricing · clone overview
• Murf: can I clone my own voice · recording specifications
• WellSaid: ethics
• Google Cloud: Instant Custom Voice
• Microsoft: Personal Voice · Limited Access for text to speech
• OpenAI: custom voices
• Apple Newsroom, 16 May 2023. Apple; 13 May 2025. Apple
• TechCrunch, Meta acquires voice startup Play AI, 13 July 2025. TechCrunch; Inworld, Migrate from PlayHT. Inworld

Kukarella

• Pricing · Home page FAQ · Terms of Use · Privacy Policy · Clone-a-Voice · Text to Voice · DialoguesAI

Stories, law and unions

• Lehrman and Sage v. Lovo, Inc., 24-CV-3770 (JPO), S.D.N.Y., opinion filed 10 July 2025. PDF
• STV News, 5 June 2025. STV; Holyrood, 12 December 2025. Holyrood; Wikipedia, ScotRail. Wikipedia
• Screen Rant, interview with Cissy Jones, 4 August 2023. Screen Rant; Bloomberg via Taipei Times, 15 October 2023. Taipei Times; Wikipedia, Cissy Jones. Wikipedia
• AACSB Insights, How AI helped one professor continue to teach, 9 April 2025. AACSB
• NBC News, 23 February 2024. NBC News; NBC Boston, 25 November 2025. NBC Boston; FCC, 8 February 2024. FCC
• Tennessee Governor's office, ELVIS Act signing, 21 March 2024. tn.gov; Holland & Knight on the NO FAKES Act, 22 June 2026. Holland & Knight
• Wikipedia, 2024–2025 SAG-AFTRA video game strike. Wikipedia; SHOOT, 4 June 2026. SHOOT
• BLG, Bill 96 and employers, 27 May 2022, updated 9 June 2022. BLG
• W3C, Understanding SC 1.2.1, updated 9 March 2026. W3C