Voice acting for AI training, explained

The real voices behind AI voices

“AI voice acting” gets used for two different things: synthetic voices generated by software, and the real people whose recordings those voices learn from. This page is about the second - paid voice acting work for AI training, recorded by people, under a written licence.

Those recordings come from three places: open datasets, scraped audio and paid sessions. Where AI voice data comes from explains each, and why companies still pay for new voices.

The work falls into a few types: text-to-speech corpus recording, voice cloning with written and informed consent, expressive sets for conversational assistants and game characters, multilingual and accent datasets, and telephony voices for AI phone systems. You can see what is casting right now on the AI voice jobs board.

The session itself

Recording for AI training is a different discipline from commercial voiceover. A text-to-speech corpus session means reading hundreds of scripted lines with consistent pace, tone and mic position - the model learns from consistency, so the craft is in sounding natural take after take. Sessions are typically booked in paid blocks, and larger corpora run across multiple sessions.

Expressive projects - conversational assistants, game characters, emotional range sets - are closer to traditional acting work, with direction, retakes and performance notes.

The licence is the job

In AI work, what you are really selling is not the hour in the booth - it is the licence. Before any session through Voices4AI you will see, in plain English: who the client is, what the recordings train, where the resulting voice can appear, and how long the licence runs. Fees reflect that scope. A narrow internal research licence and a broad commercial deployment are priced very differently, as they should be.

What to watch out for elsewhere

The horror stories in this industry come from buried clauses: perpetual and irrevocable grants hidden in marketplace signup terms, “synthesis rights” bundled into ordinary casting contracts, datasets resold to third parties. Our rule is simple - if a term cannot be explained to you in one sentence, it should not be in the contract. That standard is written up in our ethics and consent commitments.

Getting started

Professional voice actors are matched to corpus, cloning and character work. Natural speakers - no VO experience - are in demand for language and accent datasets. Either way, the route in is the same two-minute application on our AI voice jobs page. If you are weighing up whether to licence your voice identity itself, read selling your voice to AI, ethically first.