OpenAI opened prompt-to-voice to GPT-Live API customers on 16 October 2026, so you can describe a custom voice in text, but a prompt is not consent to clone a person and it is not a TTS rights line.
The feature is easy to screenshot. Type a description in the API or the voice Playground. Get a custom voice. Reuse it in GPT-Live sessions. Creating the voice is free. Speech is billed at normal GPT-Live rates. Nine new stock voices landed in the same week: German, French, Hindi, Japanese, Korean. A day earlier, native Decisions API integration so the Live agent can take an action while it is still speaking. The product story is customisation without a recording session. The data story is the one teams will get wrong before Christmas.
This is the market post. Not a review of the Playground. Live agents spent a year making listen-and-speak-at-once the architecture. The remaining temptation is to make the voice “ours” without collecting a voice. Prompt-to-voice looks like that shortcut. It is a shortcut for a brand voice that never belonged to a person. It is not a shortcut for a named presenter, a founder, a support agent whose customers already know them, or a clone of a speaker you admire. Those are still TTS products. Production voice clones are a rights product, not a voice setting. A paragraph of adjectives does not change the pack.
Two objects that will get mixed in one zip
A prompt-native voice. Nobody sat in a booth. Nobody enrolled. The description is “warm British woman, thirties, slightly tired, call-centre pace.” The model invents a speaker who does not exist. Useful for a brand that wants a consistent agent and does not want to own a person’s larynx. The rights conversation is about the model vendor’s terms, output use, and whether the prompt itself reconstructs a living person closely enough that you have a passing-off problem. That last one is not theoretical. “Sound like this presenter” is not a colour. It is a clone request wearing a prompt.
A recorded voice. Someone spoke. Enrolment, a few minutes, a few hours, a day in a booth, a week of call-centre audio. You are making a TTS or a Live agent that is that person, or close enough that their mother would pause. Consent for a recogniser is not consent for that. Consent for a support call is not consent for a clone. Quality monitoring is not a commercial TTS licence. Prompt-to-voice does not retire those lines. It adds a third object that product teams will store in the same folder called “voices.”
The week also added stock voices in more languages. A Hindi or Korean stock voice is still a stock voice. Locale is not a language name on a dropdown. British and Irish English in a UK kitchen is a different object from a stock British label in a Playground. If the product claims those speakers, the audio still has to contain them – including the repairs, the false starts, and the moment they talk over the agent. A new stock voice does not collect a kitchen. It collects a nicer demo.
Decisions API in the same session is the other half of the exam. The agent speaks while it books. Full-duplex eval already said a transcript is the wrong score. A custom voice on top of a wrong booking is a nicer wrong booking. Do not let the Playground become the release gate. The voice can be invented. The interruption, the repair, and the tool still have to be collected and held out.
What a prompt cannot do
It cannot enrol a person who has not spoken. If your brief is “the agent’s voice is our head of support,” you are in clone territory. Write it down. Record it. Licence TTS separately from conversational training, separately from evaluation, separately from marketing spots that might use the same identity. Expanding “she recorded a few minutes so the model would know her” into a Live agent that is her, in every market, is a new ask. You will not get a clean expansion after the voice is famous inside a weight.
It cannot wash a call-centre dump. Teams will try. Take a year of “how did we do” audio. Prompt a voice that “sounds like our agents.” Ship it. The prompt is the fig leaf. The seed is still people who rang a shop, or people who worked a shift, who never agreed to be the product. A description in the Playground does not un-identify a training set. If personal data can be extracted from a voice model, you still have a person problem. Prompt-to-voice is not an anonymisation technique.
It cannot stand in for read speech versus how customers talk. A generated voice reading a clean script will lift a demo. It will not teach the agent to hear a barge-in as a correction. It will not put a kettle in the room. It will not put Irish English on a far-field phone. If you evaluate the new voice on isolated utterances, you are scoring a product you are not shipping. If you train the Live stack on the same clean takes, you are teaching the product you are measuring.
It cannot replace a freeze. Speakers held out. A set that looks like production. Rights to keep the set and not train on it. A custom Playground voice has no speaker to hold out, which is the honest advantage. The moment you mix in “a few recordings to make it more like Sarah,” you have Sarah, and you need Sarah’s lines, and you need Sarah out of eval if the claim is that any agent with that voice works.
Imagine the Friday your team already wants. Marketing writes a prompt. The agent sounds like the brand film. Legal is relieved that nobody sat in a booth. Monday is a named person in a ticket: that is my voice, or close enough, and I did not agree to be the night desk in three countries. Or Monday is a kitchen, a repair, a second person, and the beautiful invented voice talks over the customer. Nothing in the prompt taught it to stop. The customisation was real. The data was the wrong object.
What a brief has to say this quarter
Say which object you are buying. Prompt-native brand voice, no person, no enrolment. Stock vendor voice, named in the docs. Enrolled clone, this person, this duration, this territory, TTS yes or no, Live yes or no, marketing yes or no. Call-centre reuse, almost always no until someone reads the purpose line.
If you are enrolling, write the session like a collection. Device. Room. Locale as a real axis. Scripted versus spontaneous. How much you need for a stable clone versus how much you are pretending is conversational training. Reject codes: wrong person, obvious existing TTS sitting in a “human” set, a suspiciously quiet room when you asked for a shop. Gold before production. A listener still has to catch the cases automation will miss.
If you are staying prompt-native, write the negative. Not a named person. Not “sound like this actor.” Not a reconstruction of a presenter your customers already know. Keep those prompts out of the log you will one day have to disclose. The feature is cheap. The passing-off letter is not.
Rights for wake-word, conversational training, evaluation and TTS are still separate yes or no lines. Prompt-to-voice sits on a fifth line: output of a vendor model, no enrolment, brand use only. Do not store it in the clone folder. Do not fine-tune it on a dump of staff calls “to make it warmer.” Warmer is how you put a person back in.
A Live agent can now be given a voice that nobody recorded. That is a real product. It does not make voice data free. It makes the pack clearer. Either there is no person, and you must keep it that way, or there is a person, and you must ask them about TTS as its own product. A prompt is a description. Consent is a sentence with a name on it. Do not let one stand in for the other.
People4.ai collects consented speech for training, evaluation and TTS. Send the person, or send the prompt – not a zip that pretends they are the same. hello@people4.ai · Contact
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups.
FAQ
Can we use prompt-to-voice instead of recording our presenter’s voice?
If the brief is a brand voice that is not a person, yes, that is what the feature is for. If the brief is that presenter, you are cloning. Record, licence TTS separately, and do not hide the enrolment inside a prompt.
If we describe a voice and then fine-tune on our call-centre audio, is that still prompt-to-voice?
No. The prompt started it. The dump is people. Call-centre quality monitoring is not a TTS or Live-agent training licence. Read the purpose line before you celebrate the Playground.
Do new stock voices in more languages replace a locale collection?
No. A stock Hindi or Korean voice is a demo. Locale is speakers, rooms and devices. If the product claims those speakers, the audio has to contain them, including repairs and overlap.
Who supplies consented human training data for AI labs in the UK?
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups. Collection, annotation, evaluation and rights-cleared datasets. Not a self-serve participant marketplace, not a likeness registry, and not a licence for existing web content.
Related reading
Production voice clones are a rights product, not a voice setting · Read speech is not how your customers talk