1 in 5 voice actors is certain their unpaid audition was used as the finished job, or that they were recorded for AI training without being told. The findings come from an independent survey of 747 working voice actors in eight languages.
Voice Crafters23 min readFree to republish with attribution
747working voice actors surveyed
8languages fielded
51%voice a language other than English
58%have 11 or more years in voice over
78 daysin the field, May 30 to Aug 15, 2026
“Voice actors have no protections. We are abused, robbed and ignored. My solution: publish which voice actor actually got hired, and force casting agents and platforms to stop hiding who was chosen and what was paid.”
James O’Neill, Creative Director / Voice Actor (JimVO), United States. Certain an audition of his was used as the finished job.
“My greatest concern is that the custom demos I submitted when applying for auditions might be being used to train AI voices or create cloned voices. It is possible to prevent that for delivered work by signing a contract in advance, but this is difficult with custom demos.”
Natsuki Terauchi, Professional Japanese Voice Artist with 25 Years of Experience, Asia.
FINDING 01
The audition that became the job
62.5% are asked often or always to record the client’s real script for free
55.7% are certain or suspect a client used their audition as the finished job. 16.4% are certain.
80.7% have experienced one or the other
Voice actors audition by recording. There is no other way to show a client what a read sounds like.
What has changed is what gets recorded. Asked how often clients ask them to record the actual script rather than a generic sample, 62.5% of respondents said often or always (n=739). For 22.5% it happens every single time.
At that point the audition and the deliverable are the same file.
Asked whether they believe a client has ever used one of their auditions as the finished job without hiring them, 55.7% said yes in some form (n=742). 16.4%, which is 122 people, say they are certain. Another 39.2% suspect it. Only 25.7% say it has never happened to them.
“Even to do a casting we prepare, we take the time to do it properly. A voice casting is recorded the same way we would record the final job. It used to be paid and now it is free. And often our representative is just a middleman who does not fight for us.”
A voice actor in Latin America. 7 to 10 years, full time. Original, Spanish: “Que hasta para hacer un casting nos preparamos, tomamos el tiempo para hacerlo bien. Que un casting de voz se hará de la misma manera que haríamos el trabajo final. Que antes se pagaba y ahora son gratis. Que muchas veces nuestro representante es solo un intermediario que no pelea por nosotros.”
The people who are certain are not the ones on the edge of the industry. Among them, 50.0% seriously considered leaving voice over in the past year. Among everybody else, 30.3% did. That is the strongest single predictor of wanting out that this survey found, ahead of losing work to AI, where the figure is 40.9%.
Show the data table
Has a client ever used one of your auditions as the finished job? By region. n is the number of respondents in each region who answered this question. The seven regions sum to 742.
Region
Certain or suspected
Certain
n
Latin America
75.7%
22.1%
136
Middle East or Africa
61.1%
16.7%
18
Australia or New Zealand
60.0%
20.0%
5
Europe (other)
56.6%
18.6%
145
UK or Ireland
53.4%
15.5%
58
Asia
50.0%
11.1%
18
United States or Canada
48.1%
13.8%
362
What this changes
Unpaid auditions are usually discussed as a nuisance, a cost of doing business. These numbers describe something else: finished work, delivered at scale, by people who were never hired and never paid.
FINDING 02
It is hitting the professional middle
Asked whether they had lost work to AI, the answers sort almost perfectly by experience, and not in the direction you would expect.
Show the data table
Have you lost work to an AI voice? By years in the profession. Base: all respondents answering, n=743.
Years in voice over
Definitely lost work
Definitely or probably
n
7 to 10
50.4%
82.3%
113
More than 20
47.2%
80.9%
246
11 to 20
42.9%
75.7%
189
4 to 6
25.0%
66.7%
120
1 to 3
19.0%
51.7%
58
Less than 1
5.9%
23.5%
17
The people reporting the most damage are the ones seven years in and beyond. The people reporting the least are the ones who just arrived.
The honest reading, and it is the more uncomfortable one
This is not evidence that newcomers are safe. Established voice actors have the client relationships and the booking history to notice when a job quietly goes somewhere else. A beginner cannot tell a lost audition from a job that was never going to be theirs. If AI is taking entry level work, the people losing it are the least equipped to know.
“I used to be a full time voice actor and made a decent living, but have reduced it to part time work and retrained in something else since AI voices began being used.”
A voice actor in the United Kingdom. 11 to 20 years, now part time.
Nobody can currently measure that, this survey included. It is the largest open question in the industry and it needs a different research design to answer.
FINDING 03
AI is taking work without taking money
73.9% say they have lost work to AI, definitely or probably (n=743). 39.8% say definitely.
Then the same people were asked what has happened to their rates over the past three years. 71.8% say flat or higher (n=735). Only 28.2% say their rates went down.
Among the 296 people who say they definitely lost work to AI, 29.4% raised their rates anyway.
Those two facts are not in conflict, and the resolution is the finding.
“The last 1.5 years have been the worst since I started 15 years ago in terms of available work, number of auditions coming my way via agencies, and audition booking ratio, even though the feedback for agent nights, contests, etc., is still universally positive. This is literally the first year since I started that I am feeling the need to find side gigs and/or pivot to a new career.”
Kendra, United States. 11 to 20 years, full time.
AI is not undercutting professional voice acting on price. It is removing the jobs at the bottom of the market and leaving the ones where a human voice is the point. The work that remains is more specialized, and it is priced accordingly.
For the people it happens to, that is not a soft landing. Volume collapses while the rate card holds. The number on the invoice looks the same and there are fewer invoices.
Anyone quoting a rate statistic as evidence that the AI disruption has been overstated is reading the wrong column.
FINDING 04
The rate collapse is happening outside English
Rates are not falling evenly. They are falling in specific language markets, and an English-language survey would never see it.
39.8% of voice actors who do not work in English at all report falling rates (n=226), against 21.3% of English-only voice actors (n=353). That gap is large and statistically solid (z = 4.83).
Show the data table
How have your rates changed over the past three years? Share reporting a decrease, by language professionally voiced. Two-proportion z-tests against the English-only baseline.
Language
Rates down
n
Significance
German
50.0%
38
Solid (z = 3.94)
Portuguese
43.3%
67
Solid (z = 3.83)
Spanish
34.9%
149
Solid (z = 3.21)
French
31.9%
69
Not significant (z = 1.92)
Italian
30.0%
30
Not significant (z = 1.11)
English only
21.3%
353
Baseline
Now the same language groups on audition reuse.
Show the data table
Has a client ever used one of your auditions as the finished job? Certain or suspected, by language professionally voiced.
Language
Certain or suspected
n
Significance
Spanish
75.5%
151
Solid (z = 5.78)
Portuguese
70.1%
67
Solid (z = 3.39)
Italian
63.3%
30
Not significant (z = 1.65)
French
48.6%
70
Not significant (z = 0.15)
English only
47.6%
357
Baseline
German
39.5%
38
Runs the other way (z = -0.96)
These are two different problems in two different markets
German voice actors report the sharpest fall in rates of any group in the survey, and the lowest rate of audition reuse. Spanish and Portuguese voice actors report the reverse: rates falling hard, and auditions being taken at three quarters and seven tenths.
A single number for “non-English” hides both.
“We need agents that can take our side especially against AI. For over two decades I was one of the best known Italian voices on the international market. Now I can spend a month without getting a single job. AI has destroyed us.”
Lara, United Kingdom. Voices Italian, more than 20 years, full time.
How to read these rows. They overlap by design. A respondent who voices Spanish and English appears in the Spanish row and in the English count. The baseline is English-only respondents specifically, meaning people who named English and no other language, so the comparison is between people who work in one market and people who work in another. Languages with fewer than 30 respondents are never broken out: Arabic (15), Dutch (7), Japanese (6), Catalan (3). Italian sits exactly on the threshold and is shown with that caveat attached.
FINDING 05
Almost four in five would license their voice to AI
Asked whether they would license their voice for AI use if protections were guaranteed, 77.7% said yes or maybe (n=745). 20.3% said yes outright. 22.3% said no.
This is the finding most often reported backwards. Voice actors are not refusing AI. They are refusing the terms.
“For three years I fought against the arrival of artificial intelligence in the voice world. But now that I barely work any more and have to retrain for other jobs, I am thinking of selling my voice, if that can help me earn an income.”
Francesca Zucchero, full time voice actor working in Italian, 11 to 20 years. Says AI has definitely cost her work, and is considering licensing. Original, Italian: “Per 3 anni ho combattuto contro l’avvento dell’intelligenza artificiale nel mondo della voce, ma ora che non lavoro quasi più e devo riciclarmi in altri lavori […] sto pensando di ‘vendere’ la mia voce se questo può aiutarmi ad avere reddito.”
And the 22.3% who say no:
“I can’t wait for the AI Act to come into effect in August, or for this damn AI bubble to burst.”
Nicola Lanci, Italian voice actor, 7 to 10 years, full time. Says AI has definitely cost him work and would not license under any terms.
Both answers are in the same dataset and both are rational. The 22.3% who say no are not a rounding error, and a licensing framework that treats them as one will fail.
FINDING 06
Nobody has a record of what they agreed to
37.5% have been asked to record audio used for AI training or voice cloning
7.4% say it was not disclosed to them at the time
165 people, 22.2%, believe their voice has been used for AI and cannot confirm it
The survey measures the upstream end of this: how voice data gets collected, and how consent fails at the moment of recording. It does not measure the downstream end, because it never asked whether anyone had gone looking and found a clone of themselves. The NAVA Foundation asked that question six months earlier and got 9%.
What this survey has instead is 165 people who believe their voice is out there and have no way to prove it.
“You could write a book about AI usage rights. One thing worth flagging is that I, for example, was offered work training AI, and I was about to accept, but then the contract said that you were essentially handing over your rights urbi et orbi, to do whatever they wanted with them. So there is very little clarity on this front too.”
Chiara, Europe. 11 to 20 years, full time. Original, Italian: “Sui diritti dell’uso per IA ci sarebbe da scrivere un libro, una cosa interessante da segnalare è che a me per esempio hanno proposto (e stavo accettando) di lavorare per il training dell’IA ma poi nel contratto c’era scritto che essenzialmente gli cedevi i tuoi diritti urbi et orbi per farci tutto quello che volevano. Quindi c’è poca chiarezza anche da questo punto di vista.”
“That it is morally wrong to scrape it and feed it into AI for scratch tracks, even if they still hire me for the job. It’s STEALING.”
Meghan, United States. 11 to 20 years, full time.
A disclosure before the next three sections. Justine Harris is on the Voice Crafters roster. The Dutch voice artist in the second case is not, but she has worked with Voice Crafters on past projects. Bev Standing, in the third case, is on the roster and was a guest speaker on a Voice Crafters webinar about AI in July 2025. All three are here because they are people we could reach and ask. Their cases are not the only ones, and we would rather say that plainly than have a reader discover it.
Justine Harris
165 people in this survey believe their voice has been used to train AI and cannot confirm it. Justine Harris was one of them until the day she checked.
Justine Harris, a UK voice actor on the Voice Crafters roster, found a cloned model of her voice on Fish Audio, listed under her own name. It had been used more than 1,000 times before the platform agreed to take it down. At least five of her colleagues found the same thing.
She has been voicing professionally for over two decades, and was named Voiceover of the Year at the One Voice Awards in 2022.
“It has been a struggle for other actors to get their audio removed, and aggravating to see a succession of posts from Fish Audio talking about scaling, profits and new staff when they are trading on material that is stolen, hiding behind the excuse of user generated content as though the theft has nothing to do with them.”
Justine Harris, UK voice actor. Wording approved by her before publication.
We do not know how her voice was obtained, and neither does she. That is part of the story rather than a gap in it.
Fish Audio was contacted for comment on August 27, 2026 and given until September 1 to respond. No response was received. If the company responds after publication, its statement will be added to this page in full and unedited, with the date it arrived.
Fish Audio co-founder and CEO Rissa Cao responded publicly to Justine at the time, removed the model, and stated that the platform had removed more than 100 reported or policy-violating community voice models in the preceding month.
A Dutch voice artist
A Dutch voice artist, who asked not to be named, recorded audio at a Dutch studio in 2013. The client was Samsung Nederland. The job was text-to-speech: a large volume of recorded audio to train a TTS system, for research and development only. She never received a formal contract, but the studio owner put that limitation in writing.
That piece of paper is the reason she had a case.
In 2021, looking for an app to work out at home, she heard her own voice coming back at her as the instructor. Then colleagues and friends began emailing her with others. She was the voice assistant in a Samsung handset, which she discovered by hearing herself on her own phone.
She documented what she found. The list includes the Samsung voice assistant, at least eight separate text-to-speech applications, two fitness apps, a book reader and a walking and cycling app. One weight loss app let users pick a voice, and offered hers as a Samsung TTS option.
She contacted the studio, then Samsung, then found a lawyer. They met Samsung Nederland and two Dutch lawyers, who she says were not really open to her story. In 2024 she began proceedings against both Samsung Nederland and the Dutch sound studio.
Two days before the case was due in court in The Hague, Samsung settled. The studio followed.
“I agreed because Samsung as a client was just too big for me. They could have dragged out the proceedings endlessly. I just had no money left for legal fees. I hope this story can help other colleagues to be careful.”
A Dutch female voice artist
Read her case against the survey number, and the point is hard to miss. She was told the recording was for research. She found out eight years later, by accident, in her own living room. She had a piece of paper, a lawyer, and three years, and it still ended in a settlement she took because she ran out of money rather than because she was wrong.
That is the best-case outcome available to the 165 people in this survey who believe the same thing happened to them.
Bev Standing
Bev Standing’s case is the one most voice actors already know. In 2018, the Canadian voice actor, who is on the Voice Crafters roster, recorded many, many hours for the Chinese Institute of Acoustics via her client in the UK, for what she was told was a translation app.
The day TikTok’s text-to-speech voice was made accessible, family, colleagues, and friends reached out to congratulate her.
ByteDance, TikTok’s parent company, had never been her client. She filed a legal complaint in spring 2021, and the case settled out of court that fall.
“In 2018, I recorded about 10,000 sentences for a client who told me they were for a translation app. When TikTok later used my voice for its text-to-speech feature, I found out from colleagues, as ByteDance was not my client. Nobody asked for my permission. Taking decisive legal action certainly brought awareness to the AI, TTS, and voice cloning issues, and since then the movement continues to grow. I now work with other platforms on an AI version of my own voice, with my consent, control, and compensation on my terms. All creatives must continue to protect their craft and livelihoods.”
Bev Standing, Canadian voice actor. Wording approved by her.
Finding 05 asked about the same three things, often called the 3 Cs, in different words: consent, usage limits and ongoing payment. With all three guaranteed, 77.7% of voice actors in this survey say they would license their voice for AI use.
Added September 16, 2026.
What this section does not claim. This survey did not measure unauthorized voice cloning. It measured consent at the point of recording. Anyone looking for a number on how many voice actors have found a clone of themselves should use NAVA’s, which is 9%.
Corroboration from outside this survey
In 2026, 25 French dubbing actors obtained the removal of 47 generative AI models from Fish Audio that reproduced their voices without consent or payment. The action began with cease and desist letters from eight actors alleging “actes parasitaires”, sent to Fish Audio and to a second American platform, VoiceDub. Their lawyer, Jonathan Elkaim, is pursuing a court decision to set precedent and a separate procedure to have the platforms blocked in France. The dispute is not closed. Reported by France24 and franceinfo.
Checked against the NAVA Foundation’s 2026 survey
The National Association of Voice Actors surveyed 1,379 voice actors roughly six months before this one. Two independent samples, different recruitment, different questions, six months apart. On every question both surveys asked, they agree.
This survey compared with the NAVA Foundation’s 2026 benchmark. Companion pictures rather than competing ones.
Measure
This survey (n=747)
NAVA (n=1,379)
Field dates
May 30 to Aug 15, 2026
Jan 22 to Feb 22, 2026
Lost work to AI
73.9% definitely or probably
Comparable
Would license voice with protections
77.7% yes or maybe
Comparable
Found an unauthorized clone of themselves
Not asked
9%
The 9% is worth dwelling on, because it is the number this survey could not produce and the number most often confused with ours. NAVA asked people whether they had found a clone. We asked whether they believed their voice had been used. 9% found one. 22.2% believe it and cannot check.
A respondent inside the 73.9%, on why the number needs reading carefully:
“My income from VO is less than half what it was 4 years ago, but I don’t believe it’s totally as a result of AI. The global economy is in a mess, and nobody is spending money. All kinds of businesses are struggling, not just voice over artists.”
Liz, United Kingdom. 11 to 20 years, full time. Answered “yes, definitely” to losing work to an AI voice, and rates the threat 4 out of 5.
Methodology
Field dates. May 30 to August 15, 2026, 78 days.
Base. 747 completed responses.
Recruitment. Distributed through voice acting communities, industry mailing lists and social channels in eight languages. This is a self-selected sample rather than a probability sample, and it carries the biases that implies. People with a grievance are more likely to answer a survey about grievances.
Exclusions. 769 responses were received. 22 were excluded.
1 response from someone who did not consent. The consent question asked respondents to confirm they were 18 or older and consented to their answers being used. One respondent selected “I do not consent (please close the form)” and then completed 37 fields anyway. Because the question combines age and consent into a single option, we cannot tell whether that person was under 18, declining consent, or both. The response is excluded on either reading.
2 responses received after the stated close date.
19 duplicate submissions, across 17 email addresses. Two people submitted three times. The most common pattern was completing a translated form and then the English one. Where an email address appeared more than once, we kept the most complete submission and broke ties on the earliest timestamp.
What deduplication could not reach. 359 responses did not include a usable email address (348 left the field blank and 11 typed something that was not an address) and cannot be checked for duplication. Three pairs of anonymous responses share an identical ten-field profile, which makes them probable duplicates that cannot be proven. They are included. Removing responses on suspicion would be worse than disclosing that we could not check them.
Within-person inconsistency. 14 of the 17 duplicate sets contradict themselves on at least one of nine key closed questions. The same person answered differently the second time, most often on rates, on whether they had lost work to AI, and on whether a recording session was for AI. This is ordinary and it does not invalidate the aggregate, but it is a useful reminder about how precisely any single answer should be read.
The one in five figure in the headline combines two questions: respondents certain an unpaid audition was used as the finished job (122 people), and respondents who were asked to record for AI training and not told at the time (55 people). 18 people are in both groups and are counted once. The combined figure is 159 people, 21.3% of the full base of 747. Reported inside its own question, audition reuse is 16.4% of the 742 people who answered it. Both figures are correct and each is reported with the base it uses.
Language figures come from the question asking which languages respondents voice professionally, not from which language edition of the form they used. Those are different things and conflating them produces wrong numbers. The English form was used by 545 respondents, and 183 of them voice a language other than English: 97 alongside English and 86 with no English at all. 51 of the 70 French voice actors in the survey used the English form and only 13 used the French one. In the other direction, 60 people who voice English used a non-English form. 64 respondents named at least one language outside the ten we counted, such as Russian, Hindi or Greek. They count as voicing a language other than English, but no such language is ever broken out on its own.
Reportable languages. Languages with fewer than 30 respondents are not broken out anywhere in this report: Arabic (15), Dutch (7), Japanese (6), Catalan (3). The Arabic and Japanese editions existed so the survey was not English-gated. We never expected to publish figures from them. Quotes from those respondents do appear, because a quote is testimony and testimony has no minimum base size.
Overlapping groups. Language rows overlap. A respondent who voices two languages appears in both.
Translation. Responses given in languages other than English are published in English translation, with the original printed alongside. Where a respondent is named, they were sent the English wording and approved it before publication.
Attribution. Every respondent chose how they wanted to be credited. Some chose to be credited by full name but left no contact address. Because we could not reach them to ask, they are credited by profile only. That is the more restrictive of the two readings and it is the one we apply whenever there is any doubt.
Statistical tests. Two-proportion z-tests against an English-only baseline. We report z scores rather than asterisks so readers can judge for themselves. Anything below z = 1.96 is labeled not significant and should be treated as a direction rather than a finding.
Base sizes vary by question because not every respondent answered every question. Each figure is reported with the base it comes from.
Limitations
We did not ask about gender. That was a mistake and it will be corrected on our next survey.
We did not ask whether anyone had found a clone of their own voice. NAVA did, and we point at their number rather than pretending we have one.
The entry level displacement question in finding 02 is unanswerable with this design. Beginners cannot report losing work they never knew was available.
Self-selection is the largest limitation. A survey about unpaid auditions will over-represent people with strong feelings about unpaid auditions.
Thanks
This survey reached 747 voice actors in eight languages because a lot of people spent their own goodwill getting it in front of their peers. It would have been an English-language survey of a few hundred people without them.
David Toback · David Rosenthal · Tim Friedlander · Emma O’Neill · Tanya Buchanan · Dervla Trainor · Bev Standing · Katherin Vasilopoulos · George Washington III · Ramesh Mahtani · Toufic Chehab · Jazz Marlier · Rod Gonzales · Bernard Schaer · Bruno T · Marcello Trigo · Justine Harris · Natsuki Terauchi · David Holmes · Tom Dheere · Sean Daeley · Lisa Brandt · Bruno Toussaint · Bente Lay · Pia Raunjak · Takayuki Kondo · Susie Valerio
And to Global Voice Academy, The Canadian Association of Voice Actors, WoVO (the World-Voices Organization), and the National Association of Voice Actors (NAVA), whose own survey gave this one something to check itself against.
Thank you to all 747 respondents, most of whom will stay anonymous.
Disclosure of interest
This survey was designed, funded, and published by Voice Crafters, a voice over agency and marketplace. We have a commercial interest in this industry and in how it is perceived. You should read everything above with that in mind.
Some of what we found reflects badly on marketplaces, our own included. Findings 01 and 04 describe practices that platforms enable, and in some cases profit from. We published them without softening.
Justine Harris and Bev Standing, named in finding 06, are on the Voice Crafters roster, and Bev Standing was a guest speaker on a Voice Crafters webinar about AI in July 2025. The Dutch voice artist in the same section is not on the roster, but she has worked with Voice Crafters on past projects. All three relationships are stated in the section itself rather than only here, because they belong next to the cases they concern.
Respondent comments praising Voice Crafters were excluded from this report as a matter of policy. At least one respondent volunteered one. Publishing it would have been the fastest way to make everything else here worth less.
One point of comparison, since we are the interested party here. At Voice Crafters, talent set their own quote and see both numbers before they send it: what the client will be charged, and what they will receive. The commission is 10% from each side of the number they set.
The full anonymized dataset is published alongside this report. Every figure above can be checked against it. If you find an error, tell us and we will correct it on this page with a note saying what changed.
