What AI job language looks like

You may see terms such as the following used in an audition breakdown or job contract. It’s important to be aware of them, as sometimes actors initially agree to record for something like “voice modeling” or “dataset creation” without realizing that it means they are lending their voice for an AI job!

• Text to speech
• Artificial intelligence training
• Dataset creation
• Machine learning
• Voice cloning
• Voice modeling
• Digital double
• Virtual assistant / chatbot
• Real-time generation
• Simulated voice
• Voice synthesis
• Voice skinning

Remember that it is always okay to ask questions—an above-board client should be happy to clarify usage.

Why should voice actors be careful with AI training?

On the surface, AI voice training can seem tempting, especially to a less experienced actor or one who is having trouble finding work. Clients may dangle four or even five-figure buyouts in the job breakdowns, perhaps assuring you that you don’t even need to be a trained actor, you just need to record yourself speaking for a number of hours! If someone is in a dire financial situation, they may think “what’s really the harm?”

However, there are a number of very legitimate reasons that so many actors balk at the idea of selling their voice in this manner. While the decision is up to you in the end, it’s important to be aware of the concerns and potential drawbacks.

You may receive a one-time buyout in exchange for allowing your voice to be generated over and over, years into the future, with no additional compensation.

Most text-to-speech voices are done for an “in-perpetuity buyout” at a flat rate. Not only does this mean the client could theoretically re-generate your voice until the end of time if they wanted to, the company who purchases the rights to your voice data can then turn around and sell that data to a giant third-party corporation, without any additional compensation to you and perhaps without even your knowledge.

For an example of how this can – and does – happen, imagine you agree to create a voice model for a relatively small audiobook company with the purpose of allowing indie authors without a big budget to have their books narrated. You agree on a price you feel is fair for this type of usage and sign a contract allowing this company the rights to use your voice model in exchange for that one-time buyout. All seems well until one day this small publisher starts to go under and assets get bought out by a big tech corporation, who acquires the rights to all their synthesized speech models. You then start to hear your voice used in all sorts of commercials, social apps and other mainstream media – even though that was never the intended usage – with zero additional compensation to you.

Think this type of thing sounds farfetched? See the TikTok voice lawsuit.

This pay—even if it seems good up front—often isn’t worth it in the long run.

Predatory AI voice training gigs are often posted on pay-to-play casting sites, where clients may hope to sucker in desperate and/or less experienced talent who don’t understand what a job of this nature is worth. The icing on this terrible cake is that the site itself may take a commission fee – perhaps as high as 20-25%! – of your payment in exchange for the use of their platform, making your take-home pay much less than expected.

And of course, when a client is able to artifically generate a recording using your voice any time they wish, it means they don’t have to hire and pay you to actually voice that job for them—meaning significant income from ongoing sessions could theoretically be left on the table.

So the next time you see a broad in-perpetuity AI buyout for a few thousand dollars and think “wow, that sounds great, I could pay my rent with that!”, know that in context, that rate is insulting and predatory for what they are asking you to do, and it’s not worth giving up the rights to your voice for. It is understandable to want to avoid having a day job, but it is not worth jeopardizing your desired career for a month or two’s worth of rent. What seems like a large sum of money for a single job actually often turns out to be an insulting offer when you factor in the workload, usage, and potential future ramifications of the job.

Being the AI voice of a major brand could potentially lock you out of doing work for competing brands.

While being the voice of a major brand (think Siri, Alexa, Google Assistant, etc) may sound like a very exciting opportunity to advance your career, it can also backfire on you. If the voice becomes so widely used that the general public associates it with that one specific brand, it may lock you out of potentially lucrative work for other brands—-especially competing ones. There is also no guarantee that you would ever even be allowed to claim credit for being the voice of that brand, as much work of this nature requires contributors to sign very strict NDAs that could even hold them legally liable for disclosing information about the company they worked for.

While it’s true that even non-AI commercial work can give you a conflict for similar brands (for example, you generally couldn’t be the voice of national commercials for two different fast food companies running at the same time), the standard is that you should be paid a usage fee based on where or how long the company wishes to run the ad, and get paid again if they wish to renew for additional usage. This helps ensure that holding the conflict is financially worth it to you. But if you become an AI voice for a company with in-perpetuity usage, that could give you a potential lifelong conflict (another reason it’s not recommended to accept in-perpetuity buyouts for commercials, whether AI or not.)

In many cases, you give up any control of what your voice is made to say.

When you get cast for a regular voiceover job, you read each line on the script in real-time. If you were to see something in the script that you found morally objectionable or offensive, you could ask to remove or change the material (or walk away from the job if they were not willing to do so.) You would still ultimately have control over whether or not you wish to say the words on the page, even if it meant losing the job.

But if you sell your voice model data with “no strings attached”, that means anyone who has access to the technology could generate a file of your voice saying whatever they want it to. Some people may have moral objections to their voices being used – even artificially – to promote ideas or brands that don’t align with their personal values. And if the general public is able to interact with your voice model using real-time generated responses, even if the company tries to put in guardrails to prevent misuse, there is always the possibility that some users will find a way around it. For an example of how this can go horribly wrong, see the Darth Vader Fortnite controversy.

There are significant environmental concerns associated with generative AI.

Generative AI already has a bad reputation in many circles due being associated with theft from creative professionals such as artists and photographers, whose work is often used to train the algorithms without consent or compensation. This alone is bad enough, but the impact goes even deeper. The data centers needed to power all the generative AI applications demand a large supply of water and electricity. If you’re curious to know why and how, read more here.

The actual workloads may be more than they seem.

• While some jobs of this nature may be recorded at a studio, a lot of them – especially the lowball ones – expect the actor to record at home. The standards for audio quality tend to be very stringent, so you’ll need to be able to record for many hours in a perfectly quiet space with no outside noise.
  
• Although a rudimentary voice model can be made without a whole lot of data to go off of, more sophisticated models require a very large amount of recorded data to make the synthetic voice sound as natural as possible. This means you will likely be recording hours upon hours of random sentences to feed into the algorithm, leaving you with less time, energy, and vocal stamina for other jobs.
  
• In addition to the recording workload alone, chances are you will also be expected to edit your audio to clean out any mistakes, noises, etc. You may think “that doesn’t sound so bad, I already do that for my jobs and auditions anyway!” But cleaning a one-minute audition file is much different than cleaning, say, ten or twenty hours of finished audio. The time adds up, and it probably isn’t going to be compensated extra, as most jobs like this will either pay per finished hour or a flat rate.
  
• Clients may also ask for revisions and pickups for any reason they see fit, and these are often unpaid. When a client gets free revisions, they may be excruciatingly nitpicky about the way you pronounced certain words, or what they think is a minor mouth noise or breath, etc. These revisions can be very demanding and require a quick turnaround.
  
• Some AI training jobs will not even provide a script, and will expect actors to improvise the entire thing!

Lending your voice to AI is a “Pandora’s box” situation.

Once you sell the rights to your voice data and it’s fed into the algorithm, you cannot exactly take it back—even if you regret it down the line. It can also be difficult to prove your voice was used for other AI purposes—for example, if your voice ends up being mixed with a bunch of other actors to train a new model and then shows up somewhere else, it can be difficult to impossible to identify it as belonging to you. Some actors avoid any type of AI voice modeling job for this reason—once it’s out there, it’s already too late.

Companies are marketing AI voices as a cheaper and easier alternative to hiring a human voice actor (aka, you and your peers.)

Nowadays, more and more companies are popping up that market the idea of synthetic voices as a replacement for actual human talent on commercials, video games, audiobooks and more. Needless to say, many industry professionals find this extremely problematic. Even if all the voice models were trained with actors’ consent (which they often aren’t), and even if the actors were compensated well for the use of their voice data (which again, they often aren’t), the fact remains that you would be contributing your likeness to a service that literally aims to put you out of a job. Taking voice training jobs of this nature can harm the ability of everyone in your field—-including you, your friends, and your peers—-to make a living doing voice work.

If you DO decide to do this type of work

Ultimately it is up to you as an actor whether or not you wish to take a job. There are situations where an actor may decide it is worth it if they feel that sufficient protections and compensation are in place.

If you have been offered a contract involving the use of your voice for AI purposes and decide for one reason or another to take it, the following tips may help you.

Get a clear description of intended use of your voice model, and see if the client will agree to contractually limit this usage to that specific purpose.

In this example, imagine an Internet service provider wants to create a “voice assistant” for their phone lines to help callers diagnose their connection problems. See if you are able to negotiate your voice model to ONLY be used for the purpose of interactive customer responses for that specific company (as opposed to say, also using that voice model in their commercials), and ask for it to be spelled out in the contract that they will not sell or give away your voice model data to a third-party company. While clients won’t always agree to it, “single-application” usage protection can help you feel more comfortable about working this type of job.

Ask if the client would be open to a yearly licensing fee.

While many clients will expect a one-time buyout in the form of a lump sum, it’s worth at least trying to negotiate a term limit on the use. For example, if you agree to a one-year license of your voice model for a certain rate, the company would be able to use your voice model in its generative AI applications for the following year, but would need to renew for each additional year after that (upon which you could attempt to renegotiate the rate if you so wished.) If the client balks at an initial one year limit, a three or even five-year usage term is still much better than “in perpetuity” (meaning, forever.)

Another potential option, depending on the structure of the voice model, is a sort of “royalty system” in which the actor is compensated each time an end client chooses to use their voice model. However, this type of usage can be difficult to track in practicality and would only work in certain types of applications.

Find out if there are any guardrails in place for what your voice model can be made to say.

Is there a chance that your generated voice files could be used to say things that are offensive or that you are otherwise uncomfortable with? For instance, this is probably not a concern if your voice is being used as a virtual assistant for a specific company’s website or app, but much more of a concern if it’s available for download on a website where users can type in anything they want the voice to say.

Is the work otherwise something that could be done by a human performer?

Some voice actors feel more comfortable with the idea of providing voice data (with consent and compensation) if the use is for something where they could not possibly or practically be engaged as a human performer. This might mean providing data for a text-to-speech service that reads news articles aloud, an app that translates one language to another in real-time, or a tech support bot that responds to customer queries.

However, many actors are rightfully concerned if a company wants to use text-to-speech technology to avoid having to hire them for sessions in the future. Here are some hypothetical instances of how this could play out.

Example 1: Video Games
You might be comfortable with an AI voice model if it was used for situations where your character could address the player verbally by name and respond to them with a dynamic conversation based on whatever they ask you (something you could not realistically do as a human performer.) However, you might be uncomfortable with an AI voice model if the company wanted to create it for the purpose of not having to bring you back for pickups, line changes or added material (something you COULD feasibly do as a human performer.)

Example 2: Phone Messaging
You might be comfortable with an AI voice model if it was used to assist patients at a doctor’s office with confirming or rescheduling their appointments when their staff was busy or out of the office (something you could not realistically do as a human performer.) However, you might be uncomfortable with an AI voice model if the company wanted to create it for the purpose of being able to change their phone greetings for holidays or updated information without having to hire you again to record a new message(something you COULD feasibly do as a human performer.)

If you (or the client) are based in the United States, see if the job can potentially be converted to a union contract.

In 2024, SAG-AFTRA called a strike for video game performers in an effort to secure AI protections in game contracts and ensure that members would be appropriately informed and compensated if a game company wished to use any of their recordings to create a synthetic voice model. The standardized union agreements can help protect actors from exploitative terms.

One reason some actors cite for only considering AI voice work if it’s on a union contract is because on a union production, a client can more easily face legal repercussions for violating the contract. On a nonunion job, you’d be responsible for taking the client to court yourself, which may be too difficult or expensive, especially if the client is a large corporation with a robust legal team and/or is located in a different country.

Remember that you do not ever HAVE to agree to lend your voice for AI training, even if the terms seem good.

Imagine you get an audition that checks all the boxes. They offer fair compensation (possibly even with re-use fees), a single-product usage limit, no possibility of being made to say something offensive, and the nature of the work is such that it could not be performed organically by a human actor. Perhaps it even comes through your agency, so you know it’s legit. But something about it still gives you the ick, and you can’t explain why. After all…aren’t they meeting all of the demands of what actors have been asking for?

The thing is, it’s natural to feel uncomfortable with the idea of someone else puppeting your voice, and if any audition gives you a bad gut feeling – AI or not – you are within your right to pass on reading for it. Yes, even if it’s a requested audition or even a direct offer! Clients (and agents) realize that many actors are not comfortable with generative AI in any regard, and should be understanding. And on the off-chance an agent does try to push you into accepting an AI voice job…re-evaluate if this is someone you want representing you.

Can companies make an “ethical” AI voice model?

There are legitimate uses for text-to-speech technology, such as for accessibility purposes (ex: reading onscreen text in real-time) or applications such as spoken instructions for GPS navigation where it would be virtually impossible to hire a human actor to record every conceivable output. Clients who wish to generate AI voices ethically should abide by the following principles.

An actor must provide express informed consent before any AI voice training is to take place.

AI use permissions should never be “slipped into” a contract in a way that could easily mislead a voice actor – whether intentionally or unintentionally – into signing something without being fully aware of what they’re agreeing to. The client should explicitly disclose to the actor that they are requesting AI usage and confirm that the actor agrees to the terms before the job commences.

The client should provide transparency as to how the actor’s voice data will be used.

If an actor is providing their vocal likeness to train an AI model, the client should disclose to the best of their ability exactly where and how the voice is intended to be used—even if it means the actor must sign a non-disclosure agreement first. For example, is it for internal training purposes only, or will the use be public-facing? Is there any chance that the voice model could be made to say something objectionable? What happens to the data if the company gets bought out in the future?

The actor should be compensated appropriately and fairly for the use of their voice data.

Actors should be fairly paid for the use of their vocal likeness for AI/TTS—for example, via ongoing licensing fees, or usage restricted to a single product/brand. If the nature of the work realistically requires a one-time buyout, this buyout amount should factor in any potential conflicts with other brands and/or loss of potential jobs the actor could otherwise be hired for as a human performer. If a perpetual NDA is required, this may also be factored into the rate. The actor should also be protected from having their voice model sold to third-party companies for no additional compensation.